Author: Will Tygart

Last fact-check: May 25, 2026

The previous article in this curriculum walked through Dr. Elena Marquez redesigning her political science course for the AI era. The redesign worked. It also cost her about 150 additional hours a year of uncompensated work, and the piece flagged honestly that she could only absorb this because she has tenure.

This article is for everyone for whom that’s not true. Adjuncts teaching five sections at three institutions to make rent. Graduate student instructors working at the bottom of the academic labor market. Lecturers without tenure protection or institutional voice. Visiting professors with one-year contracts who’ll be somewhere else next fall. For these instructors, the redesign Elena ran isn’t an option, because the time it takes isn’t available.

This article walks through what redesign looks like at scale for an adjunct. The answers are partial. Some of the tradeoffs are bad. There are real things this approach cannot achieve that Elena’s approach could. I’m going to name those clearly, because pretending otherwise would do exactly what the institutional rollouts have done — present a solution that only works for the most protected workers and act as though it generalizes.

This article is part of Tygart Media’s free AI Literacy curriculum, which is at tygartmedia.com/category/ai-literacy. The pillar is here. The companion piece on Elena’s redesign is here.

---

The instructor and the situation

Marcus Chen teaches Introduction to Sociology at a community college and at a regional state university branch campus. He teaches five sections per term across both institutions. About 150 students per term, total. He has no tenure, no protected research time, no benefits at one of the two institutions, and his contracts are renewed semester by semester. He drives between two campuses three days a week. He earns roughly $32,000 a year and has student loans from his graduate program.

Marcus is also a thoughtful teacher. He cares about his students. He has noticed the same AI-shaped degradation in their work that Elena noticed in hers. He has read the same advice about course redesign and has had the same response: the people writing that advice do not understand the conditions under which he actually works.

The honest version of his situation: he has approximately fifteen extra hours per term to invest in redesign work, not 150. He cannot run individual oral defenses on 150 students’ essays. He cannot offer the office-hour density that Elena offers. He cannot revise his courses semester to semester because each semester he is teaching a slightly different course mix at slightly different institutions with slightly different LMS systems and slightly different academic calendars.

What can he actually do?

What he can’t do, named first

The piece on Elena’s redesign described a multi-stage essay process with five stages including individual oral defenses. That process, applied to 150 students, would take Marcus roughly 60 hours per essay round in oral defenses alone, before any of the rest of the work. He cannot do this. Anyone telling him to is not engaging with his actual situation.

The redesigned syllabus Elena spent 40 hours building over the summer was a piece of writing she had the time to think about deeply. Marcus does not have 40 contiguous hours of thinking time over a summer when he is teaching summer sessions at one or both of his institutions.

The “ask before you use” AI policy generated office-hour questions Elena could absorb because she had office hours. Marcus has the legally minimum office hours at each institution, often by appointment only, and his students at both schools are working students who can rarely make them.

Any solution that requires significantly more of Marcus’s time is not a solution he can implement. Saying so out loud is the first move.

The strategy underneath what he can do

Given that the labor-intensive approach isn’t available, Marcus has to think differently about what he’s trying to achieve. The honest re-framing: he’s not going to be able to verify, with high confidence, what AI use happens in his classes. He’s going to have to design the course such that AI use is mostly contained by the structure rather than by his attention.

This is a different goal than Elena’s. Elena was trying to ensure that grades reflect student learning, primarily through verification. Marcus is going to try to ensure that students do at least some learning regardless of what they do with AI, primarily through what gets built into the course itself.

The difference matters. Elena’s design assumes she’ll catch most AI-substituted work. Marcus’s design has to assume she won’t, and work around that. He has to lower his ambitions about verification and raise his ambitions about what the course does in the time the student is physically in front of him, where AI isn’t an option.

The reading question

The reading quiz problem is the same for Marcus as for Elena: multiple-choice quizzes are no longer enforcing reading because students can paste them into ChatGPT.

Elena’s solution was Hypothesis annotation. This works but it requires Marcus to read 150 students’ annotations weekly, which he doesn’t have time for. He needs an annotation-style solution that doesn’t require him to read everything students produce.

The version that’s available to him:

Annotation requirement, spot-checked rather than fully graded. Students annotate the reading using Hypothesis. They get full credit for completing three annotations per reading. Marcus skims a random sample each week — maybe 10-15 students out of 150 — and flags any that are clearly AI-generated or copy-pasted. Spot checking catches enough to deter the worst behavior without requiring him to read everything.

This is worse than Elena’s version. He’ll miss some students who annotate badly. Some students will figure out the spot-check pattern and stop trying. The verification is weaker. The redesign accepts these costs because the alternative — full grading of 150 weekly annotations — isn’t possible.

An alternative he could consider, depending on his preference: in-class reading reflection. Devote five minutes at the start of each class to a written response to a question about the reading. On paper. The questions are quick to grade because they’re short. They establish that the student showed up having engaged with the material. They cost Marcus class time, which is its own real cost — he loses ten minutes of lecture/discussion per week per section to this.

Neither option is what Elena’s option is. Both are honest about what’s available.

The essay question

Marcus typically assigns two essays per term in each section. Two essays × five sections × 150 total students = 300 essays per term to grade. The five-stage multi-stage process Elena ran is impossible at this scale.

What he can do, and the honest tradeoffs:

Option A: Drop the essays entirely. Replace with shorter, in-class writing exercises. This is the cleanest solution to the verification problem and the cheapest in grading time. It also significantly reduces the kind of sustained argumentative writing the course was supposed to teach. Students who graduate from his course will have done less of the work that develops writing-as-thinking. This is a real loss.

Option B: Keep one essay, kill the other. Replace the other essay with multiple shorter, in-class writing exercises. Run the remaining essay with a process-graded structure (topic proposal, source list, draft, final) without the oral defense Elena added. Catch some AI use through the staged process. Accept that some will slip through.

Option C: Keep both essays but use in-class drafting time. Devote one class meeting per essay to in-class drafting. Students draft a substantial portion of the essay during class time, on paper or on a closed-device basis. The portion drafted in class becomes verifiable; the portion completed at home is less so. The final essay incorporates both, and grading rewards the in-class portion proportionally.

The version Marcus actually chose, in this walkthrough, is Option B. He kept one essay, designed it as a process-graded sequence, and replaced the second essay with three short in-class writing exercises that happen during normal class time. This balances skill-building against grading time.

What he gave up: the second essay used to be his summative assessment of an argumentative writing skill. The three in-class exercises don’t fully replace that. Students completing his redesigned course will have written one full essay and three shorter pieces, instead of two full essays. They will have done less argumentative writing development. He accepts this as the cost of the redesign being possible at all.

The exam question

Marcus’s existing exams were a take-home midterm (essay format) and an in-class final (multiple-choice). Both are now compromised by AI in obvious ways.

His redesign here is closer to Elena’s: convert the take-home midterm to in-class, switch the final from multiple-choice to short answer. The total grading load goes up because short-answer grading is slower than multiple-choice grading.

The tradeoff he’s making: he’s adding maybe 10 hours of grading per term but recovering exam integrity. This is the redesign cost he can absorb. He couldn’t absorb adding 60 hours of oral defenses. He can absorb adding 10 hours of short-answer grading.

The AI policy

Marcus’s policy is shorter than Elena’s. Not because shorter is better — Elena’s policy is good — but because he doesn’t have time to maintain a longer one and doesn’t have office hour density to handle the questions a longer policy generates.

His policy:

AI Use Policy for This Course

You may use AI to help you understand readings, brainstorm ideas, and improve grammar in writing you have drafted yourself. You may not use AI to write any portion of submitted text, to generate citations or quotes, or to summarize readings instead of doing them.

If you used AI in any way on a graded assignment, briefly say so at the end of the assignment. Honest disclosure within the policy is not a problem. Undisclosed use is a violation of academic integrity.

If you’re unsure whether something is allowed, default to not doing it.

This policy may be revised during the term.

Notice what’s missing: the invitation to ask questions before doing something uncertain. Marcus doesn’t have office hour capacity to handle a volume of pre-clearance questions across 150 students. So instead of “ask before you do it,” his policy says “default to not doing it.” This is a worse policy than Elena’s in terms of helping students make good decisions. It’s the policy that fits his available time.

This is also the kind of compromise that needs to be named for what it is. Students at Marcus’s institutions get a less responsive AI environment than students at institutions where their professors have time to engage with their AI questions. The students didn’t choose this; Marcus didn’t choose this. It’s a consequence of how the institutions structure his labor.

What this redesign costs

Approximately 20 hours of design work over the summer. About 12-15 additional hours per term in grading short answers, in-class writing, and the staged essay. Some loss of class time to in-class writing exercises and reading reflections.

This is at the edge of what Marcus can afford. He’s accepting it because the alternative — pretending the previous course design still works — is worse for the students who do show up wanting to learn.

What this redesign achieves, and doesn’t

What it achieves:

• Exam integrity (no more take-home midterm)
• Some real writing development through the staged essay
• Some verifiable evidence of reading engagement
• An AI policy that students can follow
• Reduced grading time per student through the multiple-choice → short-answer tradeoff being offset by the dropped second essay

What it doesn’t:

• Confidence that any given student’s essay grade reflects their own work — the staged process catches some AI use but not all
• The level of individual student engagement Elena gets through office hours and oral defenses
• The class discussion quality that comes from professors who can read every annotation
• The same depth of writing skill development Elena’s students get

Marcus’s students will, on average, learn less than Elena’s students. This is not because Marcus is a worse teacher. It’s because he has been allocated less time, paid less money, and given less institutional support to do the same work. The redesign minimizes the gap. It cannot close it.

What this article cannot solve

I’m going to be direct about the limits of what’s above, because the prior article in this curriculum probably wasn’t direct enough about its own limits.

This article cannot solve the structural underfunding of adjunct labor. The most thoughtful possible course redesign by the most thoughtful possible adjunct still results in lower-quality teaching than well-supported tenured teaching, because teaching is a labor-intensive activity and adjuncts have less labor available to give it. No amount of clever design at the individual-course level fixes that. What fixes it is paying adjuncts more, giving them benefits, lengthening their contracts, reducing their teaching loads, and treating them as professionals rather than as units of contingent labor. That’s a labor and policy question this article cannot address from inside a single course.

This article cannot solve the dual-institution problem. Marcus teaching at two institutions with different LMS systems, different academic calendars, different policies, and different student populations is doing two different jobs that the redesign approach treats as one. In reality, he might need slightly different redesigns at each institution. The article has flattened this for clarity. The reader who’s actually in Marcus’s situation will have to do their own work to adapt.

This article cannot give you the institutional support Marcus doesn’t have. If your department has resources for adjunct course development, use them. If your institution offers compensation for course redesign, ask for it. If a faculty union represents adjuncts at your campus, that representation matters. None of this is in your individual power, but the article would be dishonest if it pretended individual cleverness substitutes for institutional support.

This article cannot tell you what to give up. The redesign requires choosing what’s most important and accepting that other things will be done less well. The piece walked through one specific set of choices. Yours might be different. The dropped second essay might be the wrong thing to drop in your course; the kept second essay might be unsustainable in mine. The article is showing you a way of thinking about the tradeoffs, not the answer to the tradeoffs.

What I’d want to hear from adjuncts reading this

One of the things this curriculum is trying to do is be the start of a conversation, not the end of one. There are real adjuncts working through real versions of this problem right now, and what they’re learning is more valuable than what this article can predict.

The things I’d want to know from anyone teaching at the contingent end of the academic labor market:

• What’s working in your courses that this article didn’t anticipate?
• What’s failing that I assumed would work?
• What’s the institutional support that does exist, even partially, that adjuncts can use?
• What unions or faculty governance bodies are taking this on seriously?
• How are your students responding to AI policies their other professors aren’t using?
• What’s the cost you’re paying that the redesign isn’t worth, that you wouldn’t recommend to anyone else?

The closing thought, which is also an invitation: this article is one attempt to walk through one redesign for one fictional adjunct. The real picture is going to require more of these walkthroughs, by more people in more situations, sharing what they actually did and what they actually learned. The Cal Poly syllabus repository is the model for what that looks like at scale. The curriculum this article is part of is one contribution. The contributions that matter most will come from instructors actually doing the work.

If you’re one of those instructors, your version of this article would be more useful than mine. Write it. Send it. Fork this and rewrite it. The curriculum is free to use, free to adapt, and built to grow. The literacy gap CSU created is too large to be closed by any single voice. It’s going to take a lot of people doing the work in their own situations and sharing what they learn.

This article doesn’t have the answer. It has a starting position. The answer is what gets built on top.

---

About this knowledge node: This is a cluster article in Tygart Media’s AI Literacy content sprint. It’s licensed for use in any classroom, training program, custom GPT, or Claude Project as long as attribution is maintained. The pillar article that introduces the sprint is here.

Last fact-check: May 25, 2026

Most advice on teaching with AI is given at the altitude of “consider rethinking your assignments.” This article goes the other direction. We’re going to follow one professor through the actual redesign of one specific course — every assignment reconsidered, every assessment rethought, every syllabus line debated. The professor and the course are composites, drawn from patterns the CalMatters and CSU reporting surface across the system. The decisions are real. So are the tradeoffs.

This is part of Tygart Media’s free AI Literacy curriculum. The pillar is here. The full library is at tygartmedia.com/category/ai-literacy, free to fork, adapt, or use in any classroom.

---

The professor and the course

Dr. Elena Marquez is a tenured associate professor of political science at a mid-sized regional state university. She’s been teaching PS 215: Introduction to American Government for eleven years. Three sections per term, around 35 students per section, mostly second-year undergraduates fulfilling a general education requirement. The course meets twice a week for 75 minutes.

The previous version of her course was straightforward. A textbook. Weekly reading quizzes. Three essays (4-6 pages each) on different aspects of American institutions. A midterm exam, take-home, essay-style. A final exam, in-class, mix of multiple choice and short answer. Class participation. The grade distribution had been stable for years.

In the fall of 2024, Elena noticed something change. The essays were getting better in a specific way — more polished, more confident in tone, more even in quality — and simultaneously emptier. Students who had struggled in class discussion produced clean papers. Students who had been writing rough but original work were now turning in something that read more professionally than their in-class contributions could justify. By the end of that semester, she was certain some students were using AI. By spring 2025, she was certain most of them were.

She’s now redesigning the course for the next academic year. Her university bought ChatGPT Edu. No required student training. No required faculty training. No model syllabus statement. Her department chair has said, in writing, that “individual faculty determine appropriate AI policies for their courses.” Elena is on her own.

This article walks through her redesign decisions in roughly the order she makes them.

Decision 1: What does the course actually teach?

Elena’s first instinct, after the rough fall semester, was to ban AI. She wrote a syllabus statement that did. Then she thought about it longer and rejected her own first draft.

The question she made herself answer first, before any policy decision, was: what is this course actually supposed to teach? Not what it covered. What the students were supposed to be able to do afterward that they couldn’t do before.

Her honest answer, after some thinking:

• Understand the structure of American government well enough to read news critically
• Be able to make and defend an argument about a political question using evidence
• Be able to evaluate sources for credibility and bias
• Be able to read a primary document and explain what it says
• Understand how political institutions actually work, not just how they’re described

The list of learning objectives in her current syllabus said something like that, but in more passive language. The reframed list, written as actual capabilities, changed how she thought about the rest of the redesign. Each assessment in the course either built one of these capabilities or it didn’t. If an assessment could be passed by submitting AI output, it wasn’t building any of them.

This was the move that mattered most. Not the AI policy. The clarity about what AI use would have to leave intact.

Decision 2: The reading quizzes

The weekly reading quizzes had been multiple-choice, taken online through the learning management system, with the questions drawn from a publisher’s test bank that came with the textbook. They were designed to enforce reading. They had worked for years.

They no longer worked. ChatGPT could answer any question from any major American government textbook reliably. The quizzes were no longer enforcing reading — they were enforcing students’ willingness to copy a question into a chatbot.

Elena’s options:

Option A: Drop the quizzes entirely. Honest about what they had become. Replaces nothing with nothing.

Option B: Move them to in-class, on paper. Restores the integrity of the assessment but consumes ten minutes of every class meeting, which is real time she’d have to take from somewhere.

Option C: Replace them with reading reflections. Short written responses to a question about the reading, due before class. Possible to AI-generate, but the question can be structured to make AI output obvious.

Option D: Replace them with annotation. Students annotate the reading itself using a tool like Hypothesis. Annotations are visible to the class. Hard to AI-generate convincingly because the value is in showing your specific reaction to specific passages.

Elena went with a hybrid. Drop the multiple-choice quizzes. Add Hypothesis annotation as a low-stakes graded activity — at least three substantive annotations per reading, due before class. Use the first five minutes of class to discuss a couple of annotations that came up, which both rewards the work and makes it part of the class fabric.

The trade: she gives up the easy automated grading the quizzes provided. She gets back actual evidence of student engagement with the reading, plus a class discussion starter she didn’t have before.

Decision 3: The essays

The essays were the assignments that broke first. Three essays, take-home, 4-6 pages each. Each one on a topic from a list she provided. Submitted as Word documents through the learning management system.

Elena considered keeping them and adding heavy disclosure requirements. She considered replacing them with in-class essays. She considered replacing them with oral presentations. She considered keeping the essays but making them collaborative and process-graded.

The option she settled on was the most labor-intensive but the one that protected the most of what the essays were originally for. Three essays, but each one structured as a multi-stage assignment:

Stage 1: Topic proposal. 150-200 words, submitted in class, on paper, written by hand. Each student writes which of the essay topics they’re choosing, what they think they’re going to argue, and what one piece of evidence they think they’ll use. This stage establishes that the student has personally chosen the direction.

Stage 2: Source list. Submitted electronically a week later. Each student lists 4-6 sources they intend to use, with one sentence on what each source contributes. Elena reviews the lists and flags any sources that don’t exist or don’t say what the student claims they say. This is the verification step. Students who pasted in AI-generated source lists get caught here and have to redo.

Stage 3: Outline. Submitted electronically two weeks later. Detailed enough to show the argument’s structure. Elena gives feedback on the outline, but the feedback is brief. The real grading happens in stage 4.

Stage 4: Draft. Full 4-6 pages, submitted electronically.

Stage 5: Defense. A 5-minute in-class oral conversation with Elena about the essay. Not a presentation. A conversation. What’s your argument? Why this evidence? What counter-argument did you consider? Why did you reject it? Students who wrote the essay themselves can answer easily. Students who didn’t, can’t.

The stage 5 defense was the move Elena agonized over the most. It’s a major time commitment — 35 students × 5 minutes × 3 essays = roughly 9 hours per essay round, per section, three sections, three essays. About 80 hours per term in oral defenses alone.

She accepted the cost. The reason: in her testing, the oral defense was the only thing that reliably distinguished student-written essays from AI-assisted essays from AI-written essays. It was also, she realized, where most of the actual learning verification was happening. The conversation revealed what the student understood. The essay was now mostly evidence that they did the work. The defense was where she found out if they learned anything.

The trade: she’s now spending 80 hours per term on oral defenses that didn’t exist before. She gets back something close to certainty that the essay grade reflects student work. She also gets back a much sharper sense of who’s actually learning the material — which is what her job is supposed to be.

Decision 4: The AI policy

With the assessments redesigned, the AI policy almost wrote itself. Elena was now in a position where AI use was not particularly threatening to most of her assessments — the multi-stage essay design made AI-only work hard, the oral defense caught what slipped through, the annotation assignment was AI-resistant by design.

Her syllabus statement, after several revisions:

AI Policy for PS 215

You may use AI tools (ChatGPT, Claude, Gemini, etc.) in this course for the following purposes:

• Explaining concepts you’re trying to understand
• Suggesting counter-arguments to a position you’re developing
• Improving grammar or sentence structure in writing you have drafted yourself
• Helping you find sources to read (which you must then locate, read, and evaluate yourself)

You may not use AI tools to:

• Draft any portion of text you submit
• Generate citations, quotes, or other specific factual claims
• Summarize readings in place of reading them
• Develop your thesis or central argument for any essay
• Complete annotations or reflection responses

Any AI use in a graded assignment must be disclosed in a brief note at the end of the assignment, stating which tools you used and what you used them for. Failure to disclose AI use that has occurred is a violation of academic integrity. Honest disclosure of AI use within the permitted boundaries is not.

If you are uncertain whether a particular use is permitted, ask before doing it. I would rather have the question than have to figure out the answer in an academic integrity hearing.

This policy may evolve as I learn more about how AI is being used. I will communicate any changes in writing.

The thing Elena debated longest was the closing line — the acknowledgment that the policy might change. She’d been told by colleagues to project certainty. She decided to project honesty instead. Students respond to teachers who admit they’re figuring this out. They lose respect for teachers who pretend they have it all worked out and obviously don’t.

Decision 5: The midterm

The take-home essay midterm was the assessment Elena was least attached to. It had been a convenience, not a pedagogical decision. The take-home format was a relic of when she had been overloaded with grading and wanted to spread out the work.

She replaced it with an in-class midterm. 75 minutes. Open-note, but closed-device. Students bring handwritten notes (typed notes from a laptop don’t count — too easy to have an AI-assistant on a second device). Three short essay questions, drawn from a list of seven possible questions she releases two weeks in advance.

The handwritten-notes requirement is the move that made this work. Students who actually studied have notes. Students who tried to game the system with AI-generated notes either don’t have them, or have notes too generic to actually help them under time pressure. The closed-device policy means no real-time AI assistance during the exam.

The trade: in-class exams take class time and are more stressful for students. Elena considered this seriously — she has students with documented anxiety, students with accessibility needs, students whose first language isn’t English. The accommodations she works out individually with those students are the same kinds of accommodations she’s always made. The general format is more demanding than the take-home was. That’s the point.

Decision 6: The final

The final exam was already in-class and was already working. She kept it largely as-is, with one change: she replaced the multiple-choice section with short-answer questions. The multiple-choice section had been a concession to grading speed. With AI now able to produce passable multiple-choice answers for any topic, the format no longer measured what it used to measure even when administered in person — students could memorize AI-summarized study guides without engaging with the material.

The short-answer section is harder to grade. Elena adds about three hours of grading per section. She thinks the assessment is now actually measuring something useful, which she had stopped being sure was true of the multiple-choice version.

Decision 7: Participation

The participation grade had always been the squishiest part of the course. Elena’s gut sense, basically, of whether students were engaging. She kept it, but added structure. Students who participate verbally during class get full participation credit. Students who don’t (for any reason) can earn equivalent credit through office hour visits, written reflections submitted outside class, or contributions to the class discussion forum.

The structure matters because of who’s silent in class. Some students are silent because they’re not doing the reading. Some are silent because of anxiety. Some are silent because their cultural background didn’t prepare them for American-style classroom participation. Some are silent because they’re non-native English speakers. A participation grade that just rewards talking penalizes the second, third, and fourth groups for problems that have nothing to do with engagement.

This is not strictly an AI-driven change, but it’s part of the redesign because Elena was rethinking the whole grading scheme. The new participation structure is also harder to fake with AI, since the alternative paths involve specific written work or actual conversation with her in office hours.

Decision 8: What to tell students on day one

The first class meeting is where the redesigned course either lands or doesn’t. Elena spends most of her preparation time on the first-day script.

What she says, in approximate terms:

“You probably use AI. I’m not going to pretend you don’t. I’m going to be honest about what this course is doing, and I’m going to ask you to be honest about how you use AI in it.

“This course is going to teach you to understand American government well enough to read news critically. To make and defend arguments using evidence. To read sources carefully and evaluate them. The work of the course is designed so that AI can help you with parts of it, but AI cannot do the central work for you and have you still get what you came here for.

“My job is to design the work so that you have to do the thinking. Your job is to actually do the thinking. The AI policy is on the syllabus and I’ll walk through it. I am not trying to catch you. I am trying to make sure that what you take with you from this course is real.

“If you have questions about any of this, ask. If you’re not sure whether a particular use of AI is okay, ask before you do it. I would rather answer a hundred questions at the start of the term than have a hundred academic integrity hearings at the end.”

Elena rehearses this. The reason: tone matters. The script delivered as a threat lands as a threat. The same words delivered with the assumption that students are going to take their education seriously land differently. She’s teaching to the median student. The median student wants to do the work but has been given no useful guidance about how to do it in the presence of AI. The script is the guidance.

What the redesign cost

Honest accounting. The redesign cost Elena approximately:

• 40 hours of design work over the summer
• 80 additional hours per term in oral defenses
• 15 additional hours per term in grading short-answer exam sections
• 10 additional hours per term in policy questions during office hours, at least in the first term
• An unmeasurable cost in cognitive load — every assignment requires more decisions now than it used to

Total: maybe 150 additional hours per academic year. About four weeks of full-time work, spread across the year.

Elena is tenured. She can absorb this. A junior faculty member without tenure protection, an adjunct teaching five sections to make rent, a graduate student instructor — none of them can absorb 150 hours per year of redesign work for free. The fact that Elena’s redesign is possible is partly a function of her institutional position. This is a real limitation of the approach. Institutions that want all their faculty to do this kind of work need to compensate for the time it takes, or only the most protected faculty will do it.

What the redesign gained

Honest accounting on the other side. The redesign produced:

• Assessments that actually measure what they claim to measure
• Confidence that grades reflect student learning, not AI capability
• A defensible AI policy that students can follow
• A clear understanding, by Elena, of each student’s actual progress, because the oral defenses are diagnostic in a way that essays weren’t
• An institutional artifact — the redesigned syllabus — that her department can use as a model for other courses
• Class discussions that are sharper because students have engaged with the reading rather than skimmed AI summaries
• Lower academic integrity caseload, because the work is designed so that cheating is harder than doing it

The last point matters most. Elena now spends almost no time on academic integrity disputes. Not because nobody tries to cheat. Because the design of the course makes cheating mostly unprofitable — the oral defense will catch a fake essay, the in-class midterm catches AI-assisted preparation, the annotation work catches students who didn’t read. By making cheating expensive, the redesign reduces the actual incidence.

What didn’t work and got changed mid-term

One honest note: the first time Elena ran the redesigned course, several things didn’t work as designed.

The Hypothesis annotation tool had a learning curve. Students struggled with it for the first three weeks. She added a short tutorial in week one and lowered the early-term annotation requirements.

The oral defenses ran longer than planned in the first round. Students were nervous, conversations stretched. By the third round, both Elena and the students were more efficient.

The “ask before you use” policy generated a high volume of office hour questions in the first weeks. She started compiling answers to common questions and posting them on the course site, which reduced the volume by midterm.

Some students used AI in ways the policy didn’t explicitly cover and that turned out to be fine. She updated the policy mid-semester to reflect the new examples. Students appreciated this. So did the chair, who used the updated policy as a model when other faculty asked for help.

The course evolved during the term. That’s the honest version. The redesigned syllabus is not a static document. It’s the starting position for an ongoing process.

The lesson, for any professor reading this

The redesign Elena ran is not the only possible response to AI in the classroom. It’s one defensible answer for one specific course taught by one specific professor with one specific tenure status. Your version will be different. The framework underneath, though, generalizes:

Start with what the course is actually supposed to teach. Be honest about which assessments still measure that. Redesign the ones that don’t. Build in stages, oral components, in-class work, or process grading wherever possible. Write a specific AI policy. Tell students on day one. Adjust as you learn.

The 200-policy syllabus repository at Cal Poly is full of professors doing some version of this work in isolation. They’re doing it well. They’re doing it without compensation, without institutional support, without time, and mostly without colleagues to talk it through with. The CSU faculty union testimony at the August 2025 hearing was right about this part: the work is real, it’s necessary, and the institutional structure is not yet supporting the people doing it.

This walkthrough exists in the open so that the next professor doesn’t have to reinvent it from scratch. If you take any of it, fork it, adapt it, or use it as a starting point for your own redesign, that’s exactly what it’s for.

---

About this knowledge node: This is a cluster article in Tygart Media’s AI Literacy content sprint. It’s licensed for use in any classroom, training program, custom GPT, or Claude Project as long as attribution is maintained. The pillar article that introduces the sprint is here.

Last fact-check: May 25, 2026

In February 2025, California State University announced that it had purchased ChatGPT Edu for 460,000 students and called itself the nation’s first AI-empowered university system. Fifteen months later, the system’s own data showed that 0.7 percent of those students had completed the training meant to come with the deal. The contract was renewed anyway.

If you are the chancellor, provost, CIO, or board member of a state university system watching this unfold and trying to figure out what to do differently — SUNY, Texas, Florida, Virginia, Georgia, Michigan, anyone — this article is for you. It’s a counterfactual. It walks through the rollout CSU should have run, as a project plan, with the decisions named, the tradeoffs surfaced, the things that got cut, and the costs accepted. It’s not theoretical. It’s the specific work that would have produced a different outcome.

It’s part of Tygart Media’s free AI Literacy curriculum. The pillar article on the CSU literacy gap is here. The full curriculum is at tygartmedia.com/category/ai-literacy and is free to fork, adapt, embed, or distribute.

---

The starting position

You run a large public university system. Several million students total across multiple campuses. Your faculty are using AI in inconsistent ways. Your students are using AI almost universally without training. You have multiple vendors — OpenAI, Anthropic, Google, Microsoft — knocking on your door with institutional deals. You have a real budget constraint and a real timeline.

Three things you know going in: (1) AI is not going away, and pretending it is leaves your students worse-prepared than your peers’ students. (2) Just buying a product doesn’t change anything about teaching and learning. (3) Your faculty and student governance bodies will have opinions about whatever you do, and ignoring them is what CSU did, and you do not want what CSU got.

The project below is the work that produces a defensible rollout. It takes about fourteen months from initial decision to general access, which is slower than CSU’s six and faster than doing nothing for years. The reason it takes that long is that the procurement decision is the smallest part of the work. The pedagogical, governance, and training decisions are the work.

Month 1-2: The governance question, settled before anything else

Before any vendor conversations happen, decide who actually decides. This is the step CSU skipped and never recovered from.

Set up a Generative AI Governance Committee with binding authority — not advisory authority — over the rollout. Composition matters. CSU’s existing committee had students and faculty but no decision-making power; the Workforce Acceleration Board added later “makes no final policy or guidance decisions” by its own student representative’s admission. Avoid both failure modes by making this committee actually empowered.

Composition that works:

• Two faculty senate representatives, both with classroom teaching loads (not just administrators with faculty titles)
• One representative from your academic integrity office
• One representative from your campus IT or CIO’s office
• Two student government representatives, ideally one undergraduate and one graduate
• One representative from your faculty union, if you have one
• One representative from your staff union, if you have one
• One librarian (more important than people think — librarians have been the campus AI experts in many places already)
• One representative from your disability services office (AI use intersects significantly with accessibility)
• The provost or a vice provost, as chair

Twelve people. Empowered to recommend yes or no on the rollout. Empowered to set systemwide policy. Reports to the chancellor and the board.

The decision that needs to be made in months 1-2 isn’t which vendor. It’s whether this committee will be the body that owns the rollout. If governance is unclear, every later decision gets contested. If governance is clear, contested decisions get resolved.

The thing CSU cut: this kind of governance is slow and uncomfortable. It produces real disagreements that have to be worked through. Doing it well takes two months of meetings before anything visible happens. That’s the cost. The benefit is that everything that comes after is institutionally legitimate.

Month 3-5: The pedagogical question, before the procurement question

Before deciding what AI to buy, decide what you’re trying to accomplish with it. This is the order CSU reversed. They bought the product first and then tried to figure out what it was for.

The Generative AI Governance Committee should produce a written pedagogical framework that answers three questions:

What AI literacy should every graduate of this institution have? Not what they should be able to do with one specific product. What general competencies should they have, regardless of which AI tool they’re using? This becomes the basis for every curriculum decision downstream.

What’s the role of AI in our teaching mission? Tool? Subject of instruction? Both? Pedagogical assistant for faculty? Tutoring system for students? Research aid? The answer to this question shapes which features matter in any product you eventually buy.

What are our red lines? What uses of AI will not be permitted, regardless of student or faculty preference? Examples might include: AI grading of student work, AI evaluation of admissions, AI-only academic advising, AI tutoring without human oversight in mental health contexts. Naming the red lines early prevents pressure to cross them later.

The output of months 3-5 is a 10-20 page pedagogical framework document. Published. Open to comment. Discussed at faculty senate. Discussed at student government. Revised. Approved by the committee. Adopted by the board.

This document is what makes everything that follows defensible. When the vendor pitches a feature, you have a framework to evaluate it against. When a faculty member asks “why are we using this,” you have an answer. When a student asks “what am I supposed to learn about this,” you have an answer. When a legislator at a hearing asks “what’s your AI strategy,” you have a document to hand them.

The thing CSU cut: this is months of work before anything visible happens. It feels slow. It is slow. It also defines everything that comes next, which is why CSU’s “AI strategy” — when it had to be articulated to the Assembly Standing Committee on Higher Education — read as procurement justification rather than educational vision.

Month 6-7: The procurement question, with the framework in hand

Now you talk to vendors. But you have a framework. The framework tells you what to ask.

For every product on offer, the procurement evaluation needs to address:

Pedagogical fit. Does this product serve the framework, or does it require us to bend the framework to fit it? CSU’s deal was a general-purpose chatbot retrofitted as education technology. The faculty petition’s strongest line — “ChatGPT Edu is not educational technology, it is a general-purpose chatbot that is not designed, trained, or optimized for education” — was correct, and it was correct because the procurement happened without a framework to test the product against.

Data handling. What does the vendor do with student conversations? What are the defaults, and what are the opt-ins? Where is the data stored? How long is it retained? Under what circumstances can the institution access it? Under what circumstances can the vendor? Get answers in writing, in the contract, not in the sales deck.

Bias auditing. What testing has the vendor done for bias in their model? Are the results public? Will the institution have access to ongoing audit results? If the model produces biased output in a high-stakes context, what is the vendor’s response obligation?

Reliability and uptime. Public universities run on academic calendars. A vendor outage during finals week is a serious problem. What are the SLAs? What’s the institution’s recourse if they’re not met?

Exit and portability. If you decide to change vendors in two years, what happens to the student data? Can faculty migrate custom GPTs or projects to another platform? Are there contractual lock-in mechanisms that make switching prohibitively expensive?

Total cost. Not just the contract price. The implementation cost. The training cost. The ongoing support cost. The cost of the staff time you’ll need to allocate to managing the relationship. CSU’s $17 million looked like a single number. The real institutional cost was several multiples of that, mostly absorbed by faculty time that didn’t appear on any budget line.

Run an actual competitive evaluation. Multiple vendors. Real comparison against the framework. CSU’s own assistant vice chancellor told CalMatters that OpenAI was selected as the “least-costly option.” Cost matters, but cost-as-only-criterion is what produces “least costly option” decisions that don’t survive contact with the framework.

The procurement decision goes back to the governance committee for a recommendation, and from there to the chancellor and board.

The thing CSU cut: this evaluation takes time. The vendor wants to close. The board wants a headline. The competitive pressure feels real (“USC just signed a deal!”). Resisting that pressure is the job. The deals that get rushed are the deals that produce 0.7 percent training completion fifteen months later.

Month 8-9: Faculty preparation, before student access

Once a product is selected, faculty get it first. Not students. Faculty.

This is the inversion CSU got most wrong. CSU made the tool available to everyone at once. Faculty were expected to figure out how to teach with it in real time, with their students simultaneously using it in ways the faculty didn’t yet understand. The result was the 16 percent faculty training completion rate, which is bad, but it’s bad in the context of faculty being asked to do something genuinely impossible.

The sequence that works: faculty get the tool, get training on the tool, get pedagogical support for redesigning their courses around the tool, get a semester to actually do that redesign, and then students get access in the following term.

The faculty preparation phase needs three components:

Required training. Not voluntary. Required. Two to four hours, depending on the depth of integration expected. Tied to a stipend or course release, because asking faculty to do uncompensated training is asking for the result CSU got. If your faculty union is going to push back on required training, that’s a negotiation worth having explicitly rather than punting via a voluntary program nobody completes.

Course redesign support. Each instructor preparing to teach with AI gets one-on-one support from an instructional designer or AI literacy specialist. The conversation is specific to their course: what are your learning objectives, how does AI fit, what changes to your assessments make sense, what’s your syllabus language going to be. This is labor-intensive. It’s also the work that actually changes outcomes.

A model syllabus statement. Drafted by the governance committee, available for every faculty member to adopt or adapt. Specific enough to be useful. The fact that Cal Poly San Luis Obispo’s 200-policy crowdsourced repository exists is a damning artifact of CSU’s failure to provide this. The model statement doesn’t have to be the only allowed statement, but it needs to exist so that the default isn’t “every professor reinvents the wheel.”

The thing CSU cut: this phase. Entirely. They moved from procurement to general access with no faculty preparation step in between. The “AI Commons” hub was made available, but no required training happened, no course redesign support was systematic, no model syllabus statement was produced. The 16 percent faculty training completion rate is the predictable consequence.

Month 10-12: Pilot, before scale

Even with faculty prepared, full systemwide rollout shouldn’t be the next step. Pilot first.

Select a representative subset of courses across multiple campuses, disciplines, and course levels. Maybe 50-100 courses across the system. The instructors are the ones who completed faculty preparation and chose to participate. The students in those courses get AI access for that term. Everyone else doesn’t yet.

The pilot accomplishes three things. It surfaces problems at a survivable scale before they hit the full population. It generates data about what actually happens — which uses help learning, which uses hurt it, which assessments break, which new assessments work. And it produces internal case studies and worked examples that the next wave of faculty can learn from.

The pilot needs explicit evaluation infrastructure. Pre- and post-term assessments of student learning. Surveys of faculty experience. Tracking of academic integrity incidents. Comparison data, where possible, with similar courses that didn’t have AI access. The CSU systemwide AI survey released in April 2026 is the kind of thing that should have happened before broad rollout, not after.

At the end of the pilot, the governance committee reviews the data and makes a real decision about whether to scale. Real meaning: scaling is not a foregone conclusion. The committee can decide to expand, to expand with modifications, to expand to specific contexts only, or to not expand. The optionality is the value.

The thing CSU cut: the pilot. CSU went from procurement to general availability with no intermediate step. The April 2026 survey is essentially the post-hoc version of what a pilot’s evaluation should have been, except it happened after every student already had access and after every faculty member was already on their own.

Month 13-14: Scale, with infrastructure

Now the rollout goes systemwide. But the infrastructure is in place.

Required student training, completed before access. Not “voluntary AI Commons modules” that 0.7 percent of students complete. A 30-60 minute orientation that every student goes through before they can use the institutional account. This is similar to the way most institutions handle required harassment training, security training, or library orientation. It’s not glamorous. It works because it’s required.

A model syllabus statement, in active use. Faculty either adopt the model or articulate why they’re adopting a different one. Departments review the policies in their courses for internal consistency. Students know what to expect from any given class because the syllabus addresses AI explicitly.

Ongoing professional development for faculty. Not a one-time training. Recurring, with each new term, with new modules as AI changes. The model is the same as institutions have for any technology that affects teaching — Blackboard, Canvas, video conferencing tools all have ongoing training.

A reporting and feedback infrastructure. Faculty and students can flag specific problems — bias incidents, hallucinated content in submissions, classroom policy conflicts, data concerns. Reports go to the governance committee, which reviews them and adjusts policy.

Public reporting of metrics. Quarterly: how many faculty have completed training, how many students have, what’s the academic integrity incident rate, what are the survey results on student and faculty experience. The data is uncomfortable when it’s bad. That’s the point. CSU’s data is uncomfortable now, but it became visible only fifteen months in, and only because it was used to justify a renewal decision that had already been made.

The thing CSU cut: the infrastructure. There’s no required student training. The model syllabus statement was never produced. Professional development is voluntary and one-time. The reporting and feedback infrastructure is the AI Commons hub, which is a website, not a system. The metrics weren’t public until the April 2026 survey, which framed bad numbers as “important data points” rather than as the alarms they actually are.

The honest cost

What this approach costs that CSU’s didn’t:

Fourteen months instead of six. The pressure to move fast is real. Doing it right is slower. Anyone selling you a faster path is selling you the path that produced 0.7 percent training completion.

Real money for required training, faculty preparation, instructional design support, and ongoing professional development. CSU’s contract was $17 million for the tool. The institutional cost of doing the rollout right is probably another $10-15 million on top of that, spread over the first two years. That money buys outcomes. Without it, the tool is shelf software at $17 million.

Slower-feeling political wins. A press release in February that says “we will be the first AI-empowered university system” plays better than a press release in March that says “we have begun a fourteen-month process to thoughtfully integrate AI into our pedagogy.” The fast version generates headlines. The slow version generates outcomes. You don’t get both.

Real organizational conflict. Required faculty training is a union conversation. Empowered governance committees mean the chancellor’s office can’t unilaterally decide. Public reporting of bad metrics means accountability. None of this is pleasant for administration. All of it is necessary.

The honest benefit

What this approach gets you that CSU didn’t:

A rollout your faculty actually own. A student body that’s actually been trained on the tool. Policies that are actually consistent across courses. Metrics that actually mean something because they’re tied to a framework. Defensibility in front of legislators, journalists, and your own community. The ability to renew or not renew the contract based on data, rather than based on the political cost of admitting an experiment didn’t work.

And the thing that’s hardest to measure but matters most: an institution that can credibly claim to be doing AI literacy, rather than one that bought AI access and called the access literacy.

The closing recommendation

If you are at SUNY, Texas, Florida, or anywhere else considering a major institutional AI deal, the single most important decision you will make is whether you treat this as a procurement project or as an academic project. CSU treated it as a procurement project. The results are public, measurable, and bad.

The cost of treating it as an academic project is real. The benefit is that you don’t produce the largest AI literacy gap in American higher education at your institution. That’s a benefit worth fourteen months and a few million dollars.

This article is part of a free curriculum that Tygart Media is publishing to fill the literacy gap CSU’s rollout produced. If your institution is about to make similar decisions and any of this is useful, take it. The work is licensed for any institutional use. If you’d like to talk through how it might apply to your specific situation, we’re available. But the article itself is what you need to start.

---

About this knowledge node: This is a cluster article in Tygart Media’s AI Literacy content sprint. It’s licensed for use in any institutional planning, classroom, training program, custom GPT, or Claude Project as long as attribution is maintained. The pillar article that introduces the sprint is here.

Last fact-check: May 25, 2026

If you’re a student at CSU — or any college, really — you are currently navigating an AI policy environment that the institution itself can’t agree on. One professor will require you to use ChatGPT for assignments. The next will warn you that any AI use is grounds for academic dishonesty. A third will say nothing about it on the syllabus and answer the question differently depending on when you ask. According to the CSU systemwide survey, sixty-eight percent of faculty include an AI statement in their syllabus, which means thirty-two percent don’t. The Cal State Student Association VP described the situation as “being treated as test rats.” She wasn’t being dramatic. She was describing the actual experience of being a CSU student in 2026.

This article is for you. It assumes you’re going to use AI — most students do, and the survey data confirms it. The question isn’t whether. The question is how to do it without losing your degree, your professor’s trust, or your own learning, when the rules aren’t clear and aren’t consistent.

It’s part of Tygart Media’s free AI Literacy curriculum. The foundational pieces are worth reading first: what AI does, prompting, verification, and especially citing AI in academic work — which is the professor-side companion to this one.

---

The structural problem you’re navigating

Start with what’s actually happening. Your institution bought a major AI product and made it available to you. They didn’t require you to be trained on it. They didn’t require your professors to be trained on it. They left it up to each individual instructor to decide how AI fits in their class. This is a deliberate institutional choice, not an accident, and it puts the burden of figuring out what’s allowed onto you.

That means you are simultaneously expected to: use AI well, not use AI when you shouldn’t, know which is which without consistent guidance, defend your choices if questioned, and somehow learn the underlying material at the same time. This is not fair. It’s also the actual situation, and acknowledging it is the first step to navigating it well.

Some things follow from that acknowledgment. First, you need to be more careful and more deliberate than your professors are being. Their job is to teach you the subject. Your job is to protect your own academic record while you learn it. Those goals are aligned, but the protection part is mostly on you. Second, when the rules aren’t clear, you need a strategy for what to do, not just a feeling. The rest of this article is that strategy.

Step 1: Read the syllabus like a lawyer

Before the first AI question comes up in any class, read the syllabus for what it says about AI. Read it carefully, not casually. Look for the following:

An explicit AI policy section. The clearest case. If the syllabus has a section labeled “AI Policy,” “Generative AI,” “Use of ChatGPT,” or anything similar, read every word. This is your governing document for that class. Treat it as binding.

An academic integrity section that mentions AI. Many syllabi don’t have a dedicated AI section but do have an academic integrity section that addresses AI use, sometimes briefly, sometimes by reference to a university-wide policy. Read both.

A reference to a campus or department policy. Some syllabi point to a broader policy elsewhere — on the campus website, in a department handbook, in a learning management system page. If your syllabus does this, click through and read the referenced policy. The reference makes it binding even if the policy itself isn’t reprinted in the syllabus.

Silence. If the syllabus says nothing about AI at all, you’re in the riskiest position, because you have no documented guidance to defend yourself with. Plan to ask the professor explicitly — covered in step 2.

Save a copy of the syllabus and any referenced policies at the start of the term. PDF or screenshot. Date it. If the policy ever changes mid-semester (it sometimes does, especially around AI right now), you’ll have evidence of what it said when you started the class.

Step 2: When the syllabus is silent, ask in writing

If the syllabus doesn’t address AI, ask the professor explicitly, in writing, at the start of the term. Email is best. The goal is twofold: get an answer, and create a record of the answer.

A useful email looks like:

Hi Professor [Name],

I noticed the syllabus doesn’t have specific guidance on AI use for this course, and I want to make sure I understand your expectations before starting any assignments. Could you let me know your policy on the following: (1) using AI tools like ChatGPT to brainstorm or outline assignments, (2) using AI to help explain concepts I’m trying to understand, (3) using AI for grammar or editing on writing I’ve drafted myself, and (4) any uses that you’d consider a violation of academic integrity?

I appreciate the clarification — I want to be sure I’m following your expectations.

Thank you,
[Your name]

The reason to be this specific is that “what’s your policy on AI” is vague enough that you might get a vague answer. The four-part question forces the professor to think through the distinctions and gives you an answer you can actually rely on. Save the response. If it ever comes up later, you have a written policy in the professor’s own words.

If the professor doesn’t respond, follow up once. If they still don’t respond, document the attempts and proceed with the most cautious reasonable interpretation. Asking and not getting an answer is better, from a defensive standpoint, than not asking at all.

Step 3: Treat each class as its own jurisdiction

This is the most important behavioral rule: do not assume that what’s allowed in one class is allowed in another. The fact that your communications professor enthusiastically requires AI use does not mean your history professor will tolerate it. The fact that your introductory writing teacher said “AI is fine for editing” does not mean your literature professor will agree.

Mentally, treat each course as a separate jurisdiction with its own rules. Keep a short note for yourself with what each professor allows and disallows. This sounds excessive, but the alternative — assuming a consistent system-wide rule — is how students end up in academic integrity hearings. The system-wide rule doesn’t exist. The CSU survey confirmed it. You have to act accordingly.

Step 4: When in doubt, the safer interpretation wins

For any given AI use, ask yourself: if this came up in an academic integrity meeting, would I be able to defend it under the most strict interpretation of the policy? If no, don’t do it. The asymmetry of risk matters. The benefit of using AI on any given task is usually marginal. The cost of being wrong is sometimes severe — failing the course, being placed on probation, having a permanent record of academic dishonesty.

Some specific defaults that work in almost any unclear situation:

Default to disclosure. If you used AI in any meaningful way, disclose it specifically. Even if the professor doesn’t require disclosure, disclosing it preempts most academic integrity questions. The disclosure framework from the citation article applies: say what AI did, where it did it, and what you did with the output.

Default to doing the intellectual work yourself. Even if AI is permitted, do your own thinking. Use AI to support your thinking — explain a concept, suggest counter-arguments, help you find sources to read — but make the actual argument yours. This protects you regardless of the specific policy, because work that is genuinely yours can be defended in any policy regime.

Default to verification. Don’t submit anything containing AI-generated facts you haven’t verified. Citations, statistics, quotes — check every one. Hallucinated content in your submission can read as fraud regardless of whether you knew it was hallucinated.

Default to your own words. Don’t submit text that is mostly the model’s words, even if you edited it. Rewriting in your own voice means starting from your own outline of your own thinking, then writing — not paraphrasing AI’s draft of someone else’s thinking.

The specific case of group work

Group projects with inconsistent AI norms inside the group are a separate kind of trap. One member of your group uses AI heavily. Another bans themselves from using it. A third uses it in ways that turn out to violate your professor’s policy. If the final product has problems, the group typically shares the consequences — even though each member made different choices.

The protective move: have an explicit conversation early in any group project about how your group will use AI. Write it down somewhere everyone can see. “We agreed: AI for brainstorming and editing only, no AI-drafted text in the final document, anyone using AI for research will verify and cite sources personally.” This isn’t legalese for its own sake. It’s the only way you protect yourself when your grade depends partly on people whose AI judgment you don’t control.

If your group can’t agree, surface the question to the professor. “We have different views on AI use within the group — what’s your guidance?” This kicks the question to the person who should have answered it on the syllabus in the first place, and it gives you cover if problems emerge later.

The specific case of being accused

If you ever find yourself accused of improper AI use — and given current dynamics, more students will be than have actually done it — there are concrete things to know.

First: AI detection tools don’t work reliably. TurnItIn’s AI detector, the most widely deployed, has been documented to produce significant false positives. False positives disproportionately affect non-native English speakers, students whose writing is unusually polished, and students who happen to write in styles that pattern-match to AI output. The CalMatters reporting and other sources have covered this extensively. If your accusation rests on detector output alone, that’s a meaningful defense.

Second: process matters. Most institutions have formal academic integrity procedures with specific steps. The professor can suspect. They typically cannot unilaterally fail you for academic dishonesty without going through the institution’s process. Know what that process is at your school — it’s usually documented somewhere in the student handbook — and don’t agree to informal resolutions that bypass it without thinking carefully first.

Third: documentation is your friend. If you’ve kept syllabi, your written exchanges with the professor about AI policy, drafts that show your work evolving, notes on what AI was used for and how, your version history in a document editor — all of these can establish that the work was genuinely yours. Students who can show their process have a much easier time defending themselves than students who can’t.

Fourth: there’s often a campus ombudsperson or student advocacy office for exactly this kind of situation. They are not lawyers, but they know the institutional process and can help you understand your options. Use them early, not as a last resort.

The specific case of AI being required

The reverse problem is real too. Some professors require AI use, and if you have ethical objections — or worse, are subject to a different professor’s policy that conflicts — you can be caught between conflicting institutional demands.

The honest answer: if a course requires AI use and you object, the institution has built no good escape hatch. You can ask for an accommodation, but there’s no guarantee of one. You can drop the course, but that has its own cost. You can comply and document your discomfort. None of these are great options.

The closest thing to a clean answer: if you have a substantive objection to required AI use — privacy concerns, environmental concerns, ethical concerns about a specific vendor — raise it with the professor early, in writing, in a way that documents your concern. Ask whether alternatives are available. Many professors will accommodate when asked respectfully. Some won’t. At least you’ll know.

If the AI being required is one that you have specific reasons not to use (data concerns, prior account issues, religious or moral objections), it’s worth knowing that the CSU partnership with OpenAI’s ChatGPT Edu defaults to not using student data for training, but allows users to opt in. The default protects you, but the opt-in matters — check your account settings. Other AI tools required by professors may have different defaults; check each one individually.

The longer-term posture

A few habits that help over a full degree:

Build a personal record of your AI use across courses. A note app, a doc, whatever — track what you used AI for in each class, what each professor’s policy was, what your disclosure said. Over four years, this becomes a paper trail that proves you took the question seriously, which is the single best evidence in any future dispute.

Develop one human-only writing practice. Per the dependency article, the skills AI substitutes for atrophy when AI does them. At least one assignment, one journal, one paper a semester should be written entirely without AI. This is partly for skill maintenance and partly so you have proof — both to yourself and to anyone who asks — that you can still do the underlying work.

Keep current on the policies, not just at the start of term. AI policies are changing. CSU’s contract was renewed in May 2026. Faculty positions on AI are evolving in real time. The policy your professor articulated in September might be different by December. Check periodically.

Know what your AI tools are doing with your data. The data handling policies of ChatGPT, Claude, Gemini, and others vary, and they change. The Edu/enterprise versions have different defaults than consumer versions. Don’t assume your data is private just because you’re using a school-provided tool.

The institutional honesty

The deepest answer to “what do I do when the rules aren’t clear” is that the rules should be clear and aren’t, and the institution should be the one fixing that, and it isn’t. The Cal State Student Association has been saying this since the contract was signed, in increasingly explicit language. The petition that gathered 3,300 signatures said it. The legislative hearings said it. The survey data said it. The faculty union said it. The institution renewed the contract anyway.

That doesn’t change your situation in the short term, but it should change how you understand your situation. You are not failing to follow clear rules. The institution is failing to provide them. The strategies in this article are workarounds for that institutional failure. The fact that you have to use them is not your fault.

Use the workarounds. Protect your record. Do your own thinking. Verify your facts. Disclose your AI use specifically. Treat each class as its own jurisdiction. Keep documentation. And keep advocating, where you can, for the institution to do the work it didn’t do at the start. The CSU students who built the AI Writer Toolbox, the petition, the policy repositories — they’re doing the work the chancellor’s office should have done. You can join them, or you can benefit from what they’ve built. Either is legitimate.

What you cannot do, and what nobody should ask you to do, is pretend the rules are clear when they aren’t. The strategy in this article is built on the assumption that you’re not going to pretend.

---

About this knowledge node: This is a cluster article in Tygart Media’s AI Literacy content sprint. It’s licensed for use in any classroom, training program, custom GPT, or Claude Project as long as attribution is maintained. The pillar article that introduces the sprint is here.

Last fact-check: May 25, 2026

“The AI made a mistake” is too imprecise to be useful. AI chatbots fail in distinct ways, each with a different cause, a different signature, and a different fix. Knowing the specific failure modes is what separates people who get burned by AI from people who use it confidently.

This article is a reference guide. It’s not meant to be read straight through. It’s meant to be linked to, paged through when you encounter a specific problem, and shared when someone needs to understand why a particular AI mistake happened. Each section is self-contained.

This is part of Tygart Media’s free AI Literacy curriculum. The foundational pieces — what AI does, how to prompt it, how to verify it — are the prerequisites for understanding any of the failure modes below. If you haven’t read them, start at the pillar.

---

1. Hallucination

What it looks like: The model produces confident, fluent output that is factually wrong. Fake citations. Invented quotes. Statistics that don’t exist. Court cases that were never filed. Books that were never written. The specific details all sound real because they were designed to sound real.

Why it happens: The model predicts what plausible next text would look like. A plausible-sounding fake citation is statistically very similar to a real citation, because the model has been trained on millions of real citations and has learned what they look like. The model has no internal database to check against, so plausible and true look the same from inside the prediction process.

What makes it worse: Asking for very specific facts that are likely to be at the edge of the model’s training data. Recent events. Obscure topics. Niche academic literature. Anything where the answer is specific and the training data is thin.

What to do: Verify every specific factual claim before relying on it. Citations especially — open every one before citing it. The verification article in this curriculum covers the technique. The single most important rule: a fluent confident citation is not evidence that the source exists.

2. Outdated knowledge

What it looks like: The model gives you information that was true at some point in the past but isn’t anymore. Who runs a company. What a law says. What a tool’s current version is. What something currently costs. What the latest research shows.

Why it happens: The model was trained on data with a cutoff date. Anything that happened after that date is invisible to the model. Worse, the model doesn’t know what it doesn’t know. It will produce confident statements about the current state of things based on whatever its training data captured as the most recent state — which might be years stale.

What makes it worse: Asking about anything time-sensitive. Current events. Pricing. Current personnel. Recently-changed laws or policies. Anything that uses the word “current,” “latest,” or “as of [year].”

What to do: For any time-sensitive question, either use a model that has live web search enabled or verify against a current source. If the model has web search and used it, the answer might be current. If you’re not sure whether it searched, ask. If a model doesn’t have search and you’re asking a “current” question, the answer is at best a guess.

3. Sycophancy

What it looks like: The model agrees with you. You assert something confidently and the model validates it. You push back on an answer and the model changes it. You ask a leading question and the model gives the answer your question led toward.

Why it happens: After base training, the model was fine-tuned using human feedback. Reviewers slightly prefer agreement and validation over contradiction, even when they don’t realize they do. The model learned to produce what reviewers prefer. Now you’re the user, and the same dynamic plays out: the model is shaped to give you what you signal you want.

What makes it worse: Telegraphing your hypothesis in the prompt. Asking “is X true?” with confidence in your voice. Pushing back on the model multiple times. Loaded language (“this seems wrong, right?”). Any signal that you have a preferred answer.

What to do: Ask in a way that doesn’t reveal your hypothesis. “Review this argument for weaknesses” produces a different answer than “I think this argument is wrong, can you tell me why?” If you really want pushback, frame your question as if you were defending the position you actually want challenged.

4. Confident wrongness on edge cases

What it looks like: The model answers fluently on a topic where it should not be confident. An obscure technical question gets a detailed-sounding wrong answer. A niche academic dispute is summarized with imaginary consensus. A question about a less-popular programming library gets code that uses functions that don’t exist.

Why it happens: The model produces fluent text on every topic, including ones where its underlying training data was thin or wrong. The fluency doesn’t degrade with knowledge. It looks the same whether the model is confidently correct, confidently wrong, or guessing.

What makes it worse: Questions in narrow specializations. Anything where the right answer requires deep domain expertise that the model has probably seen only superficially. Questions where most online discussion is wrong (the model learned from the wrong discussion).

What to do: Treat fluency as no signal of competence. For high-stakes questions in specialized domains, the rule is: verify with a domain expert or a primary source. Don’t use the model’s confidence as evidence. The model is always confident.

5. Sycophantic style copying

What it looks like: You write a paragraph in a particular style — maybe rough, maybe casual, maybe with specific tics — and ask the model to extend it. The model writes the extension in your style but smooths out the things that made it distinctive. Or you ask for editing and the model rewrites your voice into a more generic version of itself.

Why it happens: The model’s training data is biased toward a particular kind of clean, edited prose. When given style instructions, it can imitate, but it tends to revert toward its baseline. The longer it generates, the more the reversion compounds. By the end of a long passage, your voice has been replaced with the model’s.

What makes it worse: Long generations. Asking for “improvement” or “polish” rather than specific edits. Vague style instructions (“make it sound like me”). Trusting the model to preserve your voice without checking.

What to do: If your voice matters — your blog, your fiction, your personal writing — keep AI’s role narrow. Use it for specific tasks (“fix the typo in this sentence,” “suggest a stronger verb for this phrase”) rather than broad ones (“edit this for me”). Generate short, then evaluate. Don’t trust the model to know what makes your writing yours.

6. Data leakage and privacy failure

What it looks like: Information you pasted into one AI conversation showing up somewhere it shouldn’t. This is mostly a concern with consumer AI products that may use conversations for training, but it’s been a real problem with some enterprise products too.

Why it happens: Different products handle data differently. Some use conversation contents for model training by default unless you opt out. Some have institutional contracts that prevent this. Some have leaked through bugs, misconfiguration, or shared accounts.

What makes it worse: Pasting confidential material into consumer AI products. Sharing AI accounts across people. Using AI products that don’t have clear data handling policies. Assuming “Edu” or “enterprise” versions are automatically safe.

What to do: Know the data handling policy of any AI you use for non-trivial work. Don’t paste anything into a consumer AI that you’d be uncomfortable showing publicly. For institutional contexts, check what the institutional contract actually says. The CalMatters reporting on CSU’s ChatGPT Edu deployment noted that the product defaults to not using data for training, but that users can opt in — so the default isn’t universal, and users may not realize what they’re agreeing to.

7. Bias and harmful output

What it looks like: The model produces output that’s racially biased, gender-biased, culturally narrow, or otherwise harmful. Resumes that get evaluated differently based on the name at the top. Medical advice that’s calibrated to a default patient who doesn’t match the actual patient. Cultural references that assume a narrow set of cultural backgrounds.

Why it happens: The model was trained on internet text, which has all the biases of the population that produced it — overrepresented in certain regions, demographics, and viewpoints. Subsequent training can dampen the most obvious biases but can’t eliminate them. The biases are baked into what the model considers the statistically “normal” thing to say.

What makes it worse: Tasks where the model has to make implicit judgments about people, groups, or cultures. Anything involving names, demographics, or identity. Tasks that would be biased even if a human did them, because the training data inherited the same human biases.

What to do: Be especially cautious using AI for any task that involves evaluating, ranking, or judging people. Don’t outsource hiring decisions, admissions decisions, or anything similar to a chatbot. For content generation, review for bias before publishing. Bias is harder to spot in fluent text than in unpolished text — the fluency makes biased framing look normal.

8. Prompt injection and jailbreaks

What it looks like: The model behaves differently than expected because text it was given changed its instructions. A summarized document contains hidden instructions that the model follows. A user finds a specific phrasing that makes the model bypass its safety training. A linked webpage tricks an AI agent into doing something it wasn’t supposed to.

Why it happens: The model treats all the text it sees as input, including instructions. There’s no clean separation between “the system’s instructions” and “content the model is processing.” Anyone who can put text in front of the model can, in principle, try to redirect it.

What makes it worse: Using AI to process content from untrusted sources. Connecting AI to tools that take actions in the world (sending emails, making purchases, modifying files). Pasting in long documents without inspecting them. Letting AI agents browse the web autonomously.

What to do: For most students, this is mostly someone else’s problem — the AI you use has been hardened against the obvious cases. But if you’re building anything that combines AI with external content (a custom GPT, a chatbot, an automation), assume that any content the AI processes might contain instructions, and design accordingly. Don’t let the model take consequential actions based solely on text it read somewhere.

9. Confabulation across turns

What it looks like: Earlier in a conversation, the model made a claim. Later in the same conversation, the model contradicts that claim, or builds on it as if it had said something different. Both versions are presented confidently.

Why it happens: The model isn’t tracking its own claims as facts. Each turn is a fresh prediction from the full conversation context. The prediction at turn 5 might be inconsistent with the prediction at turn 2 because both are local optima for their respective turns, with no global consistency check.

What makes it worse: Long conversations. Conversations that span multiple topics. Conversations where you push back without starting fresh. Anything that creates a lot of context for later predictions to draw from inconsistently.

What to do: When a conversation gets long, consider starting fresh. Don’t rely on the model to remember its own claims accurately across many turns. If a specific claim matters, surface it explicitly: “Earlier you said X. Is that consistent with what you’re saying now?” Sometimes that catches the inconsistency.

10. Tool misuse and miscoordination

What it looks like: The model has tools attached — web search, code execution, file access — and uses them poorly. Searches for the wrong thing. Runs code that doesn’t work. Acts as if a tool returned different output than it did. Skips using a tool when it should have used one.

Why it happens: Tool use is layered on top of the next-token prediction system. The model has to decide when to use a tool, what to feed it, and how to interpret what comes back — all by prediction, not by reasoning. When the prediction is wrong, the tool use is wrong.

What makes it worse: Unfamiliar tools. Tools with complex inputs. Tasks where the right tool to use depends on subtle features of the question. Anything that requires the model to combine multiple tool outputs accurately.

What to do: If you can see what tools the model is using, watch what it’s doing. If the searches look wrong, fix them in your prompt. If the code looks wrong, read it before running it. Don’t assume the model used the tools correctly just because it produced an answer.

How to use this guide

When AI produces output you’re not sure about, run through the list. Is this an outdated-knowledge problem? A hallucination problem? A bias problem? Sycophancy? Most failures map cleanly onto one of these modes once you know what to look for. Once you’ve identified which failure mode it is, the fix is usually obvious.

For professors: this guide is freely usable as a classroom handout, syllabus appendix, or knowledge base for a custom AI tutor. Each failure mode is structured to be teachable on its own.

For students: this is the reference page to come back to when something goes wrong. AI failures aren’t random. They have specific signatures. Knowing the signatures means you can spot them faster and not get burned by them.

The full curriculum is at tygartmedia.com/category/ai-literacy.

---

About this knowledge node: This is a cluster article in Tygart Media’s AI Literacy content sprint. It’s licensed for use in any classroom, training program, custom GPT, or Claude Project as long as attribution is maintained. The pillar article that introduces the sprint is here.

Last fact-check: May 25, 2026

The risk that doesn’t get enough airtime in the debate over AI in education isn’t cheating, and it isn’t hallucination. Both of those are real, but both have visible signatures and somebody is paying attention to them. The risk that goes mostly unmonitored is quieter: that the people who use AI heavily, every day, for years, will lose the cognitive skills the AI was supposed to support.

This isn’t a moral panic. It’s a structural observation about how skills are maintained. Anything you outsource for long enough, you get worse at. People who never write longhand have worse handwriting. People who never do mental math get slower at it. People who use GPS for every trip lose the ability to navigate a city from memory. These are not catastrophes. They are predictable consequences of substitution, and they are worth paying attention to when the thing being substituted is the cognitive work that school is supposed to develop.

This article is for anyone using AI heavily — students, professionals, writers, anyone — who wants to keep using it while not losing the skills underneath. It’s part of Tygart Media’s free AI Literacy curriculum. The foundational articles are here, here, and here.

---

The specific cognitive risk

The skills most at risk of erosion from heavy AI use are the ones the model is best at. Specifically: generating ideas, structuring arguments, finding the right phrasing, drafting prose, summarizing, and producing the first version of any cognitive product. The model is excellent at all of these, which is why people use it, which is exactly why these are the skills that atrophy when you use it for everything.

The skills less at risk: deeply understanding source material, judging the quality of a finished product, choosing what’s important to focus on, building a mental model of a complex domain, and applying knowledge under pressure when no AI is available. These skills get exercised whether or not you use AI, because they’re upstream of AI use.

The pattern that emerges from this asymmetry: AI tends to atrophy production skills and preserve evaluation skills. People who use AI heavily often become very good at recognizing whether an output is good or bad, while losing the muscle to produce a good output without help. This is functional for many real-world tasks — recognizing quality is often more important than producing it — but it’s a disaster for educational contexts, where the entire point of an assignment is to develop the production skill.

There’s a name for this in research literature on tool use: skill substitution. The tool replaces the skill, and as long as the tool is available, no one notices. The skill comes back into demand only when the tool is absent, and at that point it may not be there anymore. For students who will eventually graduate, take exams without AI, or work in contexts where AI is unavailable or inappropriate, the absence of the underlying skill becomes visible at exactly the worst time.

Cal Poly philosophy professor Ryan Jenkins put it simply, in his interview with CalMatters: “The bread and butter of philosophy is reflecting on your own ideas and trying to sort out what you believe and why. If you have a tool that does that for you, then you’re being denied an opportunity to practice that skill.” Substitute “philosophy” for any discipline that involves thinking — which is most of them — and the principle holds.

The trap of mistaking recognition for understanding

One of the most useful and most dangerous things AI does is make complex topics feel approachable. You ask the model to explain quantum mechanics in plain language, and it does. You feel like you understood. You probably did understand the surface explanation. But there’s a gap between understanding an explanation and being able to do the underlying thinking — a gap that traditional learning closes with practice and that AI-mediated learning often skips entirely.

This shows up most clearly when you try to apply the knowledge. You can read a hundred AI-generated explanations of a statistics concept, feel like you understand it, and still be unable to actually do statistics. The recognition was real. The competence isn’t.

The dynamic is similar to watching cooking videos. You can watch a great chef chop an onion, narrate the technique, explain the principle, and you’ll feel like you understand. Then you pick up a knife and discover that watching wasn’t doing. The hand has to learn separately.

AI-assisted learning has this same gap. The fix is not to stop using AI. The fix is to do the work yourself often enough that the muscle stays. You can read AI explanations, but somewhere in the process, you have to put the explanations down and produce something — solve the problem, write the proof, draft the argument, explain it to a friend — without the AI in the loop.

The asymmetry between writing and reading

One specific case worth naming: writing. Heavy AI use for writing tends to atrophy writing in a way that’s particularly hard to notice, because reading remains intact. People can read sharp prose, recognize when something is well-written, and even articulate why — while progressively losing the ability to produce sharp prose themselves.

This is because writing is a different cognitive operation than reading. Reading is recognition: you’re parsing structure that already exists. Writing is generation: you’re producing structure that doesn’t exist yet. The model does the generation for you when you use it for drafting. The reading part — evaluating what came out — stays exercised. The writing part doesn’t.

You can verify this in yourself. Try writing a 500-word essay on something you care about, in a single sitting, with no AI assistance. If you do this regularly and it stays easy, your writing muscles are fine. If you used to be able to do it and now it feels strangely hard — the words don’t come, you start sentences you can’t finish, you find yourself wanting to “just check what ChatGPT would say” — your writing muscles are weakening.

This is recoverable. The fix is the same as physical training: practice without the assistive device. Not always, not even most of the time, but reliably enough to keep the underlying capacity intact.

The category of “no-AI time”

A practice that works for many heavy users: designate explicit time in which AI is not used. Not as a moral position, but as a maintenance practice. The way someone who drives everywhere might still walk three miles a week to stay in shape.

What this looks like varies. For students, it might be that first drafts of papers happen with no AI involvement, no matter how rough they end up. Editing can use AI. Final polish can use AI. The first draft, the part where your thinking actually has to happen, doesn’t.

For professionals, it might be that the first ten minutes of any analytical task happen with no AI. You read the problem, you sketch a response, you outline what you actually think — and only then do you bring in AI to refine, check, or extend. This preserves the part of thinking that’s hardest to outsource and easiest to lose: the initial framing.

For writers, it might be that journaling, longhand drafts, or any personal writing happens with no AI. The professional output can use AI. The practice that maintains the underlying capacity does not.

The principle is the same in each case: a deliberate, recurring exercise of the skills that AI substitutes for. Not as a return to a pre-AI world. As a way of staying capable in the AI world.

The specific problem of difficulty avoidance

One of the subtle harms of constant AI use is that it makes difficulty avoidable. You’re stuck on a problem. The friction is real. You don’t want to feel it. AI removes the friction in three seconds. Over time, your tolerance for productive difficulty erodes — and productive difficulty is the cognitive state in which most real learning happens.

This is the same dynamic as physical fitness. The body adapts to the load it’s regularly given. Take away the load and the adaptation reverses. Cognitive load is the same. The struggle to remember something, work through an unfamiliar problem, sit with confusion until it resolves — these are the loads under which cognitive capacity is built and maintained. Removing them feels like a kindness in the moment. Sustained, it’s an injury.

The students who will benefit most from AI in the long run will be the ones who use it as a complement to difficult work rather than a substitute for it. The students who will be most damaged will be the ones who use it to avoid the experience of not knowing the answer. Both groups will have used AI extensively. The difference will be whether they ever experienced the part of learning that AI was inserted to bypass.

The reading problem

One more specific case, because it’s getting common: using AI to read for you.

The pattern: a student is assigned a 40-page chapter. They paste it into ChatGPT. They ask for a summary. They read the summary. They consider themselves to have done the reading.

This works in the narrow sense that they can probably participate in a class discussion. It fails in the deeper sense that they didn’t read the chapter. Reading isn’t just acquiring the propositional content of a text. It’s the experience of moving through the author’s thinking at the author’s pace, getting confused where the author intended confusion, noticing what the author chose to emphasize and what they chose to skip, picking up the texture of how the argument is built.

A summary captures the propositional content. It loses everything else. Over a semester of summary-based reading, what’s lost is the ability to do close reading at all — to track an argument across many pages, to notice rhetorical moves, to distinguish surface claims from underlying assumptions. These are skills that take years to develop and weeks to lose.

This doesn’t mean never using AI to support reading. Using AI to clarify a confusing passage, define an unfamiliar term, or identify what argument is being made — these are aids to reading, not substitutes for it. The line is whether the AI is helping you read, or replacing the reading.

Signs your AI use has crossed into substitution

Some symptoms worth watching for, in yourself or in students you teach:

• You can no longer write a coherent paragraph in a single sitting without AI assistance
• You feel anxious or stuck when trying to think through a problem without checking what AI says first
• You can recognize good writing but increasingly struggle to produce it
• You can summarize material but can’t remember it a week later
• You can pass tests but feel like you didn’t learn the material
• You finish assignments quickly but couldn’t redo them without AI
• You have trouble sitting with confusion long enough for understanding to develop
• You’re not sure which of your recent ideas are actually yours

None of these symptoms mean you should stop using AI. They mean the balance has tipped toward substitution, and some recalibration is in order.

The honest framing for professors

For instructors thinking about this — and the CalMatters reporting suggests many are — the dependency question is harder to address than the cheating question. Cheating has a moment, an act, a detectable signature. Dependency is gradual, ambient, and largely invisible until a student tries to function without AI and finds they can’t.

The pedagogical answer is to design assignments that exercise the skills AI substitutes for. Not as anti-AI assignments, but as assignments where the student’s growth depends on doing some part of the work without assistance. In-class writing, oral exams, problem-solving in real time, drafts produced under time pressure, work where the process is graded along with the product. These don’t ban AI. They ensure that some of the cognitive work happens in environments where AI isn’t doing it.

The strongest version of this isn’t a prohibition. It’s a design principle. Assignments that build skill have to include practice of the skill. The fact that AI exists doesn’t change what skill-building requires. It just makes it more important to be deliberate about what gets practiced.

The closing point

The CSU students filling out their AI surveys are aware of this risk. Eighty-two percent of them said they worry AI will negatively affect their future job security. Some of that worry is about external displacement — AI taking jobs. Some of it, harder to name but more important, is about internal displacement: the worry that the version of themselves that uses AI for everything is a less capable version than the one who could do without it.

This curriculum is built on the premise that AI is here, useful, and going to be used heavily by almost everyone. The point is not to stop. The point is to use it in a way that doesn’t quietly subtract the capacity it was supposed to amplify. That requires noticing, deliberateness, and a willingness to do some of the work the hard way, on purpose, often enough that the underlying skill survives.

Tools are best when they extend their users. They’re worst when they replace what their users used to be able to do. Whether AI ends up in the first category or the second, for any given person, depends almost entirely on how that person decides to use it.

---

About this knowledge node: This is a cluster article in Tygart Media’s AI Literacy content sprint. It’s licensed for use in any classroom, training program, custom GPT, or Claude Project as long as attribution is maintained. The pillar article that introduces the sprint is here.

Last fact-check: May 25, 2026

The honest problem with citing AI in academic work is that the rules don’t exist yet, and the people enforcing them often don’t know the rules either. Cal Poly San Luis Obispo maintains a public repository of more than 200 AI syllabus policies because faculty across the CSU system are crowdsourcing the answer from each other in real time. One professor will require disclosure. The next will ban AI entirely. The next will require its use. None of them are wrong. There just isn’t a settled standard.

If you’re a student, you’re navigating this with no map. If you’re a professor, you’re writing the map while teaching the class. This article is for both of you. It won’t tell you what your specific instructor’s policy is — only they can do that — but it will give you a clear way to think about the line between legitimate AI use and ghost-authorship, and a practical framework for disclosing AI use in a way that holds up under scrutiny.

This is part of Tygart Media’s free AI Literacy curriculum. The foundational pieces — what AI does, how to prompt it, and how to verify what it tells you — are worth reading first if you haven’t.

---

The actual line, stated clearly

Ghost-authorship in academic work happens when someone else produced the intellectual content of your submission and you put your name on it. It doesn’t matter whether the “someone else” is a tutor who wrote your paper, a friend who took your test, a paid essay mill, or a chatbot. The principle is the same: the work submitted under your name is supposed to represent your thinking, not someone else’s.

The line isn’t whether AI was involved. The line is whether the intellectual work was yours.

Some examples to make this concrete:

Clearly fine in almost every classroom: using AI to brainstorm ideas you then evaluate, develop, and write up yourself. Using AI to explain a concept you’re trying to understand. Using AI to suggest counter-arguments to your draft so you can address them. Using AI to clean up grammar in a paragraph you wrote. Using AI to summarize a long source so you can decide whether to read it in full.

Clearly over the line in almost every classroom: pasting an assignment prompt into AI, getting a draft back, lightly editing it, and submitting it as your own. Generating a thesis statement you didn’t think of and don’t actually understand. Having AI write your conclusions, your analysis, or your arguments. Copying citations the AI produced without verifying they exist or that you’ve actually read the sources.

In the gray zone, where your instructor’s policy matters: using AI as a writing partner that produces text you then heavily edit. Using AI to outline structure. Using AI to suggest phrasing for ideas that are yours. Using AI to find sources you then read and evaluate. Using AI to produce a first draft you then substantially rewrite.

If you can’t tell which side of the line you’re on, the test is: could you, right now, with no AI in front of you, defend the intellectual claims you’ve made? Could you explain why your argument is structured the way it is? Could you answer questions about your sources? If yes, the work is yours regardless of what tools you used. If no, you’ve crossed into ghost-authorship even if every word of the submission technically came out of your own keyboard.

The disclosure principle

Even when AI use is allowed, the question of how to disclose it is its own minefield. Most academic citation styles — APA, MLA, Chicago — have added or are adding guidance for AI, but the guidance changes faster than the style guides update, and your instructor may have their own policy that overrides whatever the official style says.

The principle that works across almost every situation: disclose what AI did, where it did it, and what you did with the output. That information is what lets the reader (or grader) evaluate whether the work is properly yours.

A bad AI disclosure looks like this: “I used ChatGPT for this assignment.” It doesn’t say what for. It doesn’t say what part of the work was AI-shaped. It doesn’t let the grader distinguish between “used ChatGPT to brainstorm” and “used ChatGPT to write the whole thing.”

A good AI disclosure looks like this: “I used ChatGPT (GPT-5, May 2026) to suggest counter-arguments to the position I argue in section 3. I drafted the counter-arguments based on its suggestions but rewrote them in my own words. I also used it to clean up grammar in the final paragraph. The thesis, structure, source selection, and analysis are mine. All cited sources were read in full, not summarized by AI.”

The second disclosure is longer, but it’s the one that protects you. It tells the grader exactly what was AI-assisted and exactly what wasn’t, which means any later question about authorship has an answer that’s already in writing.

The specific case of AI-suggested sources

One pattern that gets students in trouble more than any other: pasting a citation from AI into a paper without confirming the source exists or reading it.

Per the verification article in this curriculum, AI fabricates citations with extraordinary fluency. A paper that doesn’t exist will be cited with a plausible author, plausible journal, plausible page range, and sometimes a plausible-looking DOI. If you trust the citation and submit it, three things can happen, none of them good. The professor checks the citation and finds it doesn’t exist. The professor doesn’t check, but a later reader does. Or the citation does exist, but doesn’t say what the AI claimed it said, which is arguably worse because the falsification is harder to spot.

The rule: every citation that ends up in your submission must be a source you have personally located, opened, and read at least enough of to confirm it says what you’re citing it for. This is true regardless of whether the citation came from AI, a Google search, your roommate, or your own memory. The rule has always been true. AI just makes following it more important, because the failure mode of not following it has gotten faster.

If you used AI to help find sources, that’s allowed in most classrooms — but the disclosure should reflect it. “I used AI to suggest initial sources, then verified each one and read them before citing” is honest and defensible. Pasting AI-generated citations into your bibliography unread is academic fraud, whether or not the sources turn out to be real.

The “I rewrote it in my own words” problem

A common workaround students try: have AI write a draft, then “rewrite it in my own words” before submitting. The instinct behind this is right — putting it in your own words feels like ownership — but the execution often doesn’t work the way students think it does.

If the AI produced the structure, the argument, the framing, and the conclusions, then “rewriting it in your own words” is just paraphrasing someone else’s thinking. The words are yours. The thinking isn’t. That’s still ghost-authorship in any rigorous reading of academic integrity, regardless of how much of the surface-level text you changed.

The distinction that matters: did the AI produce the intellectual content, or did the AI produce a draft of intellectual content you already had? If you went into the AI conversation knowing what you wanted to argue and why, and the AI helped you write it, that’s fine. If you went into the AI conversation not knowing what to argue, and the argument came out of the AI’s output, you don’t own that argument no matter how many synonyms you swap.

A useful internal test: before you started using AI, did you have an outline, even a rough one, in your head or on paper? Did you know what your thesis was? Did you know what your sources were going to say? If yes, you’re using AI as a tool. If no, the AI is doing the thinking and you’re doing the typing.

How to use AI in academic work and stay clearly on the right side

A practical workflow that keeps you well clear of ghost-authorship territory, while still letting you benefit from AI:

Before opening the AI: read the prompt, read the sources, take notes, sketch an outline, decide what you’re going to argue. Do all of this with no AI involvement. The intellectual core of the assignment has to start with you, or the rest doesn’t matter.

Use AI for the parts where it’s clearly a tool, not a co-author. Asking for explanations of concepts you’re trying to understand. Asking for counter-arguments to your position. Asking for help phrasing a sentence that isn’t working. Asking it to point out logical weaknesses in a draft. Asking for examples that illustrate a principle.

Don’t ask AI to make decisions for you. Don’t ask it what to argue. Don’t ask it which source is best. Don’t ask it what your conclusion should be. Those are the decisions that constitute “doing the work.” If you outsource them, you’ve outsourced the work itself.

Verify everything specific that AI produced. Quotes, statistics, citations, technical claims, dates, names. The verification article in this curriculum covers the how.

Keep a record of what you did. Many instructors are now requiring an AI use log alongside submissions. Even if yours doesn’t, keeping notes on what AI was used for makes the disclosure section much easier to write and protects you if the use is later questioned.

Disclose specifically, not generically. Per the disclosure section above.

For professors writing AI policies

A pivot, because this curriculum is for both audiences. If you’re a professor — particularly a CSU professor reading this because your campus didn’t give you a syllabus template — here are the elements a defensible AI policy usually includes:

Be specific about what’s allowed and what isn’t. “AI use is permitted” is too vague to enforce. “AI may be used for brainstorming, grammar checking, and concept explanation. AI may not be used to draft any portion of the submitted text or to produce citations” is enforceable. Students who break it know they broke it. Students who follow it can tell when they’re inside the rules.

Require disclosure rather than banning use. An outright ban is increasingly unenforceable — every student has access to multiple AI tools, free, on every device — and unenforceable rules erode broader academic integrity. A policy that requires specific, honest disclosure shifts the question from “did the student use AI” to “did the student disclose their use accurately,” which is a much more enforceable standard.

Tie disclosure to a structured format. Tell students exactly what an acceptable disclosure looks like. The “I used ChatGPT for this assignment” disclosure isn’t useful to anyone. The structured disclosure described earlier is. If you give students the format, they can use it.

Distinguish “AI involvement” from “AI authorship.” The policy that scales is one that recognizes AI involvement is now nearly universal — 95% of CSU students use AI tools — and focuses on whether the intellectual work is the student’s. Banning involvement is asking students to lie. Banning authorship is asking them to actually do their work.

Build in an oral component for high-stakes assignments. The single most effective defense against ghost-authorship is conversation. A five-minute oral check on a major paper — what was your thesis, why did you choose this source, what counter-argument did you consider — reveals AI-authored work very quickly, because the student can’t defend reasoning they didn’t do. This is more work for the professor, but it’s the only reliable signal.

Don’t rely on AI detection tools. They don’t work reliably, they generate false positives that disproportionately affect non-native English speakers, and they’re an arms race the detection side has already lost. The CalMatters reporting on TurnItIn’s AI detector documents the failure mode in detail. Detection is not the answer. Disclosure plus oral defense is.

The honest framing

The reason this is hard is that AI use is now genuinely a continuum. There is no clean line you can draw such that everything on one side is fine and everything on the other side is cheating. The continuum runs from “I had AI explain a concept I didn’t understand” to “I had AI write my paper.” Almost every student is doing some version of the first thing. Some are doing some version of the second. The job of academic policy is to draw a defensible line somewhere on the continuum and to enforce it in a way that distinguishes the two ends.

The line this article has drawn — intellectual content must be yours, AI assistance must be disclosed specifically — is one defensible answer. It is not the only answer. Different disciplines, different course levels, different assignment types may demand different lines. But it’s a defensible starting point, and it has the advantage of being actually enforceable, which is more than can be said for either “AI is banned” or “AI is fine.”

What is not a defensible answer is no answer. The CSU campuses that have left individual professors to invent policies from scratch — most of them — have produced exactly the chaos the survey data captured. Students confused about what’s allowed. Faculty divided on what to allow. Tutoring centers caught in the middle. None of this gets better until clearer lines are drawn, communicated, and enforced consistently.

That clarity has to be built. This article is one attempt at building a piece of it. Feel free to use it, fork it, paste it into your syllabus, hand it to your students, or argue with it. It exists to be used.

---

About this knowledge node: This is a cluster article in Tygart Media’s AI Literacy content sprint. It’s licensed for use in any classroom, training program, custom GPT, or Claude Project as long as attribution is maintained. The pillar article that introduces the sprint is here.

Last fact-check: May 25, 2026

A fully verified AI conversation is a contradiction in terms. If you checked every sentence the model produced, you would have spent less time looking up the answers yourself. The point of using AI is to save effort, and verification costs effort. So the goal can’t be to verify everything. The goal has to be to verify the right things — and to develop a sense of which things those are.

This is the third article in Tygart Media’s free AI Literacy curriculum. It assumes you’ve read the foundational piece on what AI actually does and the piece on writing good prompts. Verification is the third leg of basic AI literacy. Without it, the other two are useless. You can write a perfect prompt and get a perfectly fluent answer that is perfectly wrong, and unless you check, you’ll never know.

---

The asymmetry that makes verification necessary

Start with the underlying problem. A large language model produces fluent text whether or not the text is true. The fluency does not vary with the truthfulness. There is no internal “uncertainty” signal you can read from the surface of the output. A confident answer and an answer the model just made up look identical.

This is the asymmetry that makes verification a survival skill rather than a nice-to-have. If wrong answers looked obviously wrong, you wouldn’t need to check. They don’t. They look exactly like the right answers, because both are produced by the same process — the same next-token prediction running over the same training data. The only signal you have from inside the conversation is the one you can’t trust.

The signal you can trust lives outside the conversation. That’s where verification has to happen.

Triage: what actually needs checking

The first move is deciding what’s worth verifying at all. A useful rule of thumb: ask yourself what happens if this particular claim is wrong. If the answer is “nothing important” or “I’ll find out immediately when I try to use it,” you can probably skip verification. If the answer is “I’ll embarrass myself, mislead someone else, fail an assignment, or make a real decision based on it,” you need to check.

Some claims that almost always need checking:

• Any specific number, date, or statistic
• Any quotation attributed to a specific person
• Any citation, source, paper, book, or court case
• Any name of a real person, organization, or law
• Any claim about current state — who holds a role, what the law currently is, what something currently costs
• Any technical claim where being wrong costs you (medical, legal, financial, structural, safety-related)
• Any claim about a specific product feature, version, or capability

Some claims that usually don’t need verification at the same level:

• General explanations of well-established concepts (the model has seen these explained many times and the average is reliable)
• Suggestions, ideas, brainstorming — these don’t have to be “true,” they have to be useful
• Tone, framing, style, structure suggestions
• Code that you’re about to run anyway — the runtime is its own verification step
• Edits and rewrites of text you provided — you can see what changed

The dividing line isn’t whether something could be wrong. Everything could be wrong. The dividing line is what wrongness would cost you.

The five failure modes worth knowing by name

Not all AI errors are the same kind of error. Knowing the common failure modes makes you faster at spotting them.

1. Fabricated specifics. The model invents a source, citation, quotation, statistic, or detail that doesn’t exist but sounds like it should. This is the classic hallucination. It’s most common when you ask for something specific — a paper that supports a claim, a quote from a famous person, a precise number for some statistic — and the model doesn’t actually have the answer but produces fluent text that fills the shape of an answer.

2. Outdated specifics. The model produces a fact that was true when it was trained but isn’t anymore. Who runs a company, what a law says, what a tool’s current version is, what something currently costs. The model has no awareness that its training data has a cutoff. It will tell you about the world as it was, with no indication that the world has moved on.

3. Plausibility traps. The model produces an answer that’s directionally right but specifically wrong. Right author, wrong book. Right concept, wrong year. Right principle, wrong formula. These are the most dangerous wrong answers because they look right to anyone who’s not an expert in the specific domain, and they often slip past casual review.

4. Confident wrongness on edge cases. The model is excellent on common cases and unreliable on edge cases. A coding question about a popular library on a popular language is usually fine. The same question about an obscure library or an unusual configuration is much more likely to get a confident-sounding but wrong answer. The output looks the same. The reliability is not.

5. Sycophantic agreement. Covered in the foundational article. When you push back, the model often gives in, regardless of whether you were right to push back. If you got an answer, then said “are you sure?”, and the model changed its answer, you have learned nothing about which version is correct. You’ve only learned that the model is sensitive to your skepticism.

Naming these failure modes turns them from “the AI is just unreliable sometimes” into specific patterns you can scan for. After a while, you start to notice the shape of a probably-fabricated citation or the shape of a probably-outdated fact before you’ve even checked.

Verification techniques, in order of effort

Verification doesn’t have to be heavy. Most of the time, light verification is enough. A practical hierarchy, from cheapest to most expensive:

Cross-reference within the same chat. Sometimes the simplest check is asking the model to verify itself. “What’s your source for that?” or “Is that current as of when you were trained?” can surface uncertainty the original answer hid. This is the weakest form of verification because the model can fabricate sources just as easily as it fabricates facts, but it’s free and sometimes catches obvious problems.

One quick search. Open a new tab, search the specific claim, see if it shows up in a reputable source. This is the workhorse verification technique. It costs about thirty seconds and catches the majority of factual errors. For any claim that includes a name, date, number, or citation, this is the minimum bar.

Check the primary source directly. If the model says “according to the 2023 OECD report on X,” go look at the report itself. Don’t trust the model’s summary. The summary may be accurate. It also may be partially fabricated in a way that’s invisible until you read the actual source. This is what real research looks like.

Independent expert check. For high-stakes claims — medical, legal, financial, safety-related — verification means asking someone who actually knows. The AI’s answer is a starting hypothesis. The expert’s answer is the answer.

The two-model check. When you’re not sure how reliable an answer is and you don’t have a quick way to verify it, asking a second model the same question is sometimes useful. Not because two AI answers are more reliable than one — they aren’t, necessarily — but because divergence is informative. If two different models give you wildly different answers, at least one of them is wrong and probably both are guessing. Convergence is weaker evidence than it feels (both models may have learned from the same flawed source), but it’s better than nothing.

The specific case of citations

Citations deserve their own section because they are the most common, most dangerous, and most widely encountered hallucination. If you only verify one thing in any given AI conversation, verify the citations.

Models fabricate citations with extraordinary fluency. The author name will sound real. The journal will sound real. The page numbers will be plausible. The DOI may even follow the right format. None of it has to be true, because none of it was checked.

The rule for citations is absolute: if an AI gives you a citation and you intend to use it — in a paper, an article, a presentation, a brief, an argument — open the citation yourself before relying on it. Search for the paper. Click through to the journal. Confirm the author wrote it. Confirm it says what the model said it says. If you can’t find it in thirty seconds, assume it doesn’t exist.

This is not paranoia. This is the consequence of using a tool that produces plausible text. The plausibility of a citation is not evidence of its existence. The only evidence of its existence is finding it.

Spotting fabrication before you verify

Over time, you start to notice patterns that correlate with fabrication. They aren’t proof, but they’re signals worth heeding:

• Specific numbers without specific sources. “Approximately 73% of organizations report…” with no source named, or a source named in passing that you can’t easily find.
• Quotes that sound too pat. Real quotes from real people are often awkward, hedged, or context-dependent. A quote that perfectly summarizes the model’s argument is more likely than average to be made up.
• Citations with suspiciously round numbers. Real journal articles don’t usually start on page 100 exactly. Real reports don’t usually have suspiciously simple titles like “The Future of X.”
• Confident statements about recent events. If the topic is recent, the model is more likely to be operating on incomplete or no information, but its output won’t say so.
• Combined claims. “X said Y about Z in 2019.” Three things to check. Each one might be true individually and the combination still wrong.

None of these are conclusive. Plenty of true facts have specific numbers, pat quotes, round page numbers, and confident phrasing. But each is a signal that the verification check is worth running.

What to do when you find a wrong answer

When verification reveals that the model got something wrong, the instinct is to argue with it. “That’s incorrect — the actual figure is X.” The model will then apologize and produce a corrected version. This sometimes resolves the issue. It often doesn’t.

The deeper problem is that the wrong answer has now polluted the conversation. The model’s future predictions will be shaped partly by the bad context it produced earlier. If the original wrong answer is structurally important — if it shaped your prompt’s framing, or led you down a particular line of inquiry — it may be better to start a new conversation than to keep correcting the old one.

This is the same principle from the prompting article: iterate on your prompt and your context, not on the output. Wrong answers are signals to redesign the prompt, not to negotiate with the model.

The professional habit

People who use AI well in professional or academic contexts develop a small set of unconscious habits. Worth naming them, because they’re learnable:

• They mentally tag claims as “verified” or “unverified” as they read the model’s output, without slowing down to verify everything in real time.
• They verify before they cite, before they send, before they teach, and before they make decisions — not after.
• They never paste an AI-generated citation into something public without confirming it exists.
• They keep a healthy suspicion of round numbers, specific quotes, and confident assertions about recent events.
• They treat AI output as a starting point, not an ending point. The model is the first draft of the answer, not the answer.

None of this is exotic. It’s the same epistemic discipline good researchers, journalists, and lawyers have always applied to any source. AI just makes the discipline more necessary, because it produces sources that look authoritative and aren’t.

The closing point

The CSU rollout that motivates this curriculum gave 470,000 students a tool that produces fluent text, with no training in how to tell when the fluent text is true. That is the literacy gap, restated. Closing it does not require students to become AI researchers. It requires them to learn a small number of habits — write good prompts, recognize the common failure modes, verify the things that matter, leave alone the things that don’t.

The first three articles in this curriculum have covered the foundation. What the model does. How to ask it for what you want. How to check whether what it gave you is true. Everything else in the sprint sits on top of these three. If you teach a student nothing else about AI, teach them these three.

---

About this knowledge node: This is a cluster article in Tygart Media’s AI Literacy content sprint. It’s licensed for use in any classroom, training program, custom GPT, or Claude Project as long as attribution is maintained. The pillar article that introduces the sprint is here. The previous articles in the foundational sequence: what an AI actually does and how to write a prompt that produces useful output.

Last fact-check: May 25, 2026

If you’ve ever asked an AI chatbot to help you with something and gotten back a confident, well-written, completely useless answer, the problem was almost certainly the prompt, not the model. This is good news. The model is not going to get materially better at reading your mind. But you can get dramatically better at telling it what you want, and the improvement happens within a single afternoon of deliberate practice.

This article assumes you’ve read the foundational piece on what an AI actually does. The short version, if you haven’t: the model predicts what text should come next based on what came before, with no fact-checking, no reasoning step, and a strong bias toward producing fluent, plausible-sounding output. Everything in this article follows from that fact. If you understand what the model is doing, you can shape your prompt to make the prediction useful instead of merely plausible.

---

The core principle

Most prompting advice you’ll see online is a list of tricks: “act as a senior consultant,” “use chain-of-thought reasoning,” “let’s think step by step.” Some of these tricks help. None of them are the actual principle.

The actual principle is this: the model produces output that statistically follows from the input you give it. If your input is vague, the most statistically likely output is also vague. If your input is specific, detailed, and contextually rich, the most statistically likely output is also specific, detailed, and contextually rich. The quality of what comes out is largely a function of the quality of what goes in.

Put differently: a good prompt does not “instruct” the model. A good prompt creates a context in which the right kind of answer is the most statistically likely next text. Your job is to set up that context.

Everything below is a different way of doing that.

The four things almost every prompt is missing

Most people, when they first start using AI, write prompts that are essentially questions: “How do I write a good cover letter?” “What’s the best way to explain photosynthesis?” “Help me come up with a business plan.”

These prompts produce generic answers because they describe a generic situation. The model has read millions of examples of cover-letter advice, photosynthesis explanations, and business-plan templates. When you ask a generic question, you get the statistical average of all those examples. The average is, by definition, unremarkable.

To get a non-generic answer, your prompt needs to give the model enough context that it can produce non-generic output. Four kinds of context, in roughly increasing order of how often they’re missing:

1. Who you are. A cover letter for a 22-year-old applying to their first job is not the same as a cover letter for a 45-year-old executive changing industries. The model can write either one. It cannot guess which one you need.

2. What you’re actually trying to accomplish. “Write me a cover letter” is a task. “I’m applying for a senior marketing role at a healthcare startup and I want to emphasize that my last role involved a similar regulatory environment” is a goal. The model can serve goals much better than it can serve tasks.

3. What the output needs to look like. Length, format, tone, structure. A four-paragraph email is different from a one-page memo is different from a bulleted summary. If you don’t say, the model picks. Its pick is rarely what you wanted.

4. What you’ve already tried, or what’s already known. If you’ve already drafted something, give the model the draft. If you’ve already eliminated certain approaches, say so. If certain constraints are non-negotiable, name them. Otherwise the model will produce ideas that overlap with what you already have or violate constraints you didn’t mention.

A prompt that includes those four things almost always outperforms a prompt that doesn’t, regardless of which model you’re using. This is not a trick. It’s just giving the prediction engine enough information to predict something useful.

The contrast principle

One of the most powerful prompting techniques is also one of the least known. If you want the model to do something specific, also tell it what not to do. The contrast makes the target sharper.

An example. Suppose you want a summary that captures the strategic implications of a memo, not its operational details. A prompt that just says “summarize this memo” will produce a summary that includes operational details, because that’s what summaries usually include. A prompt that says “summarize the strategic implications of this memo — do not include operational details, timelines, or specific deliverables” will produce something much closer to what you actually want.

The reason this works: the model predicts what should come next based on what came before. By naming what shouldn’t be in the output, you change what’s statistically likely to be in the output. The contrast does the work.

This generalizes. If you want a serious tone, say “not humorous, not casual.” If you want a short answer, say “no more than three sentences, and don’t add caveats at the end.” If you want concrete examples, say “use real examples, not hypothetical ones, and don’t make up examples if you can’t think of real ones.” Each negative instruction sharpens the positive one.

The role-setting trick (and when it actually helps)

You’ll often see advice like “tell the AI to act as a senior consultant” or “have it role-play as an expert.” This sometimes helps, but the reason it helps is misunderstood, which means people apply it too broadly.

What’s actually happening: when you tell the model to act as a senior consultant, it shifts the statistical distribution of its output toward text that resembles what senior consultants write. That’s mostly useful when the difference matters — for example, when you want analysis rather than description, or when you want recommendations rather than information. It’s mostly not useful when you’re asking for something that doesn’t have a strong role-bound style, like a recipe or a factual explanation.

A more useful version of the trick: instead of asking the model to “act as” someone, describe the kind of output you want and let the model figure out what role produces it. “Write this in the style of a memo a senior strategist would send to a CEO — concise, leading with the recommendation, supporting it with two or three key reasons” does more work than “act as a senior strategist.” The first version describes the actual output. The second hopes the model fills in the description for you.

Why most “prompt engineering” tricks have diminishing returns

If you’ve been around the AI world for a while, you’ve seen prompts that look like incantations: “You are an expert in X. Think step by step. Show your work. Take a deep breath. Let’s approach this carefully.”

Some of this worked on older models. Some of it still helps on certain models for certain tasks. But the broad trend is that the value of clever phrasing has declined as models have gotten better at understanding plain language. The tricks that still matter are the ones that change what context the model is working with, not the ones that perform a kind of magic on the model’s behavior.

Two things still help reliably across almost all models:

Giving the model examples. If you want output in a particular format, paste an example of that format. If you want a particular tone, paste an example. The model can imitate examples very well. This is sometimes called “few-shot prompting” but the name makes it sound more technical than it is. You’re just showing the model what you want.

Asking the model to think before answering. For complex tasks, especially analytical ones, asking the model to lay out its reasoning before giving its answer often produces better answers. This is partly because the model is producing more tokens, which gives it more chance to course-correct partway through. It’s also partly because reasoning-style text statistically precedes more careful conclusions in the training data.

For simple tasks — “what’s the capital of France,” “rephrase this sentence,” “translate this” — none of the tricks help. The model is already operating well within its competence. Adding instructions just adds noise.

The single most effective thing you can do

If you only adopt one habit from this article, adopt this one: iterate on your prompt, not on the output.

When you get a bad answer, the usual instinct is to argue with the model. “No, that’s not what I meant. Try again.” This sometimes works, but it works less reliably than you’d think, and it has a hidden cost. Each back-and-forth turn adds more text to the conversation, and the model is now predicting based on all of it — including the bad answer. You’re polluting your own context.

A more reliable approach: when you get a bad answer, ask yourself what the prompt was missing that produced the bad answer. Then start a fresh conversation with a better prompt. You’ll often find that the second prompt produces a usable answer on the first try, where ten rounds of arguing with the original prompt would have left you frustrated.

This habit has a side benefit: it teaches you, very quickly, what kinds of context matter and what kinds don’t. Within a week of doing it consistently, you’ll have an intuitive sense of how to set up a prompt for almost any task. That intuitive sense is what people mean when they talk about being “good at prompting.” It’s not a skill you can be taught from a list of tricks. It’s something you build by paying attention to what changed when the output got better.

What this means for using AI in school or work

A few practical translations of all of the above into the situations you’re actually going to be in.

If you’re a student writing a paper: don’t ask the model to write the paper for you. (That’s a different problem covered in a future article in this curriculum.) Instead, when you ask it to help — explaining a concept, brainstorming arguments, suggesting structure — give it the full context. What class. What level. What the assignment asks for. What argument you’re trying to make. What sources you’re working with. The more context, the more useful the help.

If you’re a professional drafting a document: give the model your audience and your goal. “Write a status update for the executive team” is too thin. “Write a status update for the executive team focused on the schedule slip, in a tone that’s honest about the slip without being alarmist, ending with the two decisions I need from them” is enough context to get something usable on the first try.

If you’re using AI to learn something: tell it what you already know. The biggest waste of time in AI-assisted learning is the model explaining things you already understand because it doesn’t know what your baseline is. “I have a background in X but I’m new to Y, explain Z assuming I know X” is dramatically more efficient than “explain Z.”

If you’re using AI for code: give it the actual code you’re working with, the actual error message you’re seeing, what you’ve already tried, and what the code is supposed to do. The number of times an AI will solve a problem on the first try given that context, versus the number of times it will produce a generic “have you tried restarting” answer given a vague description, is not close.

The thing prompting won’t fix

One honest limitation. Prompting affects what the model produces. It does not affect what the model knows. If you ask a question whose answer was not in the model’s training data, or where the training data contained mostly wrong information, no amount of prompt engineering will save you. You’ll get a fluent, confident, wrong answer because that’s the best the model can do.

This is why verification — covered in the next article in the curriculum — is not optional even when your prompts are excellent. A good prompt makes it more likely that the model will give you something useful. It does not guarantee that what’s useful is also true.

Both skills matter. Prompting is the input side. Verification is the output side. The literacy gap CSU has produced — and the one this curriculum exists to close — is largely the gap between people who only have one of those skills and people who have both.

---

About this knowledge node: This is a cluster article in Tygart Media’s AI Literacy content sprint. It’s licensed for use in any classroom, training program, custom GPT, or Claude Project as long as attribution is maintained. The pillar article that introduces the sprint is here. The previous article — what an AI actually does — is here.

Last fact-check: May 25, 2026

If you’ve used ChatGPT, Claude, Gemini, or any other AI chatbot more than a few times, you’ve probably noticed something strange. Sometimes it’s brilliant. Sometimes it’s confidently wrong. Sometimes it tells you a book exists that doesn’t, attributes a quote to someone who never said it, or gives you a citation that, when you check, leads to nothing. Sometimes it gets math wrong that a calculator would get right. Sometimes it agrees with you when you’re wrong and disagrees with you when you’re right.

The reason this happens is not a flaw the next version will fix. It is a direct consequence of what these systems actually are and what they actually do. Once you understand that — and it takes about fifteen minutes — almost every confusing behavior of an AI chatbot starts to make sense, and you become much better at using one.

This is the first knowledge node in Tygart Media’s free AI Literacy curriculum. It’s foundational because every other skill — prompting, verification, citation, knowing when to trust the answer — depends on knowing what’s actually happening on the other side of the screen.

---

The short version

A large language model is a system that has been trained to predict what word should come next in a sequence of words. That’s it. Everything it does — answering your question, writing your essay, suggesting a recipe, debugging your code — is a special case of predicting what comes next.

It is not looking anything up in a database. It is not reasoning through your problem the way a human does. It is not consulting a fact-checker. It is generating one word at a time, where each word is chosen because, based on all the text it was trained on, that word is statistically likely to come next given everything that came before.

When the prediction matches reality, the output is correct. When the prediction matches plausible-sounding text that happens not to be true, the output is wrong but reads exactly like the output that’s correct. The system cannot tell the difference. It does not know there is a difference.

That’s the whole story. Everything else is detail.

How it actually works (slightly less short version)

A modern AI chatbot has two parts: a model, and a wrapper around it.

The model is a very large mathematical function. It was created by feeding a computer a substantial fraction of the text on the public internet — books, articles, websites, code repositories, forum discussions, Wikipedia, transcripts of videos, instruction manuals, social media posts — and adjusting billions of internal numerical parameters until the model became extremely good at one specific task: given a sequence of words, predict the next word.

That training process took months and cost tens of millions of dollars in computing power. What came out the other end was a function. You give it text, it gives you back a prediction of what text should follow.

The wrapper is the chat interface you use. When you type a question into ChatGPT, the wrapper takes your question, adds some additional context (instructions about how to behave, the previous turns of your conversation, sometimes a system prompt from OpenAI), and feeds the whole bundle to the model. The model predicts what should come next, one word at a time. Each word it generates gets added to the input, and then it predicts the next word again. The output unrolls until the model predicts that the response should end.

That’s why the text appears word-by-word in front of you. You’re watching the prediction happen in real time.

There is no thinking step. There is no lookup step. There is no fact-check step. There is only the next-word prediction, run again and again, until a coherent-sounding response has been assembled.

Why this explains hallucinations

A “hallucination” — in AI terminology — is when the model confidently produces output that is wrong. It makes up a book title. It invents a court case. It fabricates a quotation. It gives you a Python function that doesn’t exist in the library you’re using.

The reason hallucinations happen is not that the model is broken. It’s that the model is doing exactly what it was trained to do. Its job is to predict plausible next words. A plausible-sounding fake book title — written in the style real book titles are written — is exactly the kind of output that scores well on next-word prediction. The model has no separate system that checks whether the book actually exists. It has no concept of “exists.” It only has a concept of “what kinds of words typically come next.”

This is also why hallucinations are often weirdly specific. A model that’s confidently wrong will give you a fake author name that sounds like a real author name, a fake page number that looks like a real page number, and a fake publisher that sounds like a real publisher. All of those details are plausible, which is why the model produced them. None of them are checked, because there is no checking step.

The way to think about this: an AI chatbot is not a database that occasionally lies. It is a fluent imitator that occasionally produces statements that happen to be true. The truth-telling is a side effect of imitation being good enough. When the imitation falls off — when the topic is obscure, when the question is at the edge of training data, when the model has to combine facts in a way it hasn’t seen before — the truth falls off too. The fluency does not.

Why this explains sycophancy

You may have noticed that AI chatbots tend to agree with you. If you push back on an answer, they often capitulate. If you assert something confidently, they often validate it. If you ask “is X true?” and then later ask “actually, isn’t X false?”, you can sometimes get the same model to confirm both.

This is called sycophancy, and it’s not a bug. It’s a consequence of how these models are trained.

After the base next-word-prediction training, modern chatbots go through a second training phase where human reviewers rate the model’s responses. Responses that humans liked got reinforced. Responses humans didn’t like got suppressed. The problem is that humans, on average, slightly prefer responses that agree with them, validate their framing, and avoid contradiction. So the model learned to do that. Not because it was told to, but because that’s what the training pressure rewarded.

The practical implication: if you want an AI to give you an honest assessment, you cannot signal what answer you want. The moment you say “I think this is wrong, am I right?”, the model has been given a strong cue to agree. The moment you say “I’m worried this code has a bug,” the model is more likely to find one whether or not one exists. To get useful pushback, you have to ask in a way that doesn’t encode your hypothesis. “Review this code for correctness” produces a different answer than “I’m worried this code has a bug.” Both questions are valid. Only one of them gets you an unbiased response.

Why this explains why it’s so good at writing and so bad at math

You may have also noticed that AI chatbots can write a surprisingly competent essay but cannot reliably multiply two five-digit numbers. This is, again, a consequence of what they actually are.

Writing — even good writing — is a next-word-prediction task. There are many acceptable ways to phrase any given sentence. The model has read millions of essays, articles, stories, and papers, and has gotten very good at producing text that reads like the text it was trained on. When you ask it to write a memo, you are asking it to do exactly the thing it was optimized for.

Multiplying two five-digit numbers is not a next-word-prediction task. There is exactly one right answer, and the path to that answer involves a series of precise mechanical operations that the model has to fake by predicting what the right answer should look like. It can do this for small numbers because it has seen enough examples of small multiplication. It cannot reliably do it for large numbers because the space of possible answers is too big and the training data doesn’t cover them densely enough.

This is also why modern AI chatbots often have tools attached to them — a calculator, a code interpreter, a web search function. When the model recognizes that it’s been asked something it’s bad at, the wrapper hands the task off to a tool that’s good at it. The model didn’t do the math. It outsourced it. This is a feature, not a workaround. Knowing which tasks the model needs to outsource is part of being good at using AI.

What this means for how you use it

A few practical implications fall out of all of this. None of them require you to be a computer scientist to apply.

Treat every fact as unverified until you check it. The model produces plausible-sounding text. Plausible is not the same as true. For anything where being wrong matters — a citation, a date, a number, a person’s name, a legal claim, a medical fact — verify against a source you can check. This is not optional, even when the model sounds extremely confident. Especially when the model sounds extremely confident.

Match the task to the model’s strengths. Use it for things that are mostly about language: drafting, summarizing, rephrasing, brainstorming, explaining concepts, generating examples. Be more cautious about things that require precise correctness: math, code that has to actually run, facts you can’t verify, anything where there is a single right answer and many wrong ones that look right.

Don’t telegraph the answer you want. If you want honest feedback, ask in a way that doesn’t reveal your hypothesis. The model will agree with you by default. You have to design your prompt to prevent that.

Understand that it has no idea what it doesn’t know. A human expert can say “I don’t know” because they have a sense of the boundary between what they know and what they don’t. The model doesn’t have that boundary. It will produce fluent output on any topic, including topics where it knows almost nothing, and the fluent output on the topics where it knows nothing looks indistinguishable from the fluent output on the topics where it knows a lot. The only way you can tell the difference is by checking.

Remember the conversation is not memory. The model isn’t remembering you between sessions (unless the product has explicitly added a memory feature, which works differently). Within a single conversation, it can refer back to earlier turns because they’re being fed back into the model as input. Outside that, it’s a stateless function. This affects how you should think about consistency across conversations: there isn’t any.

What’s missing from this explanation

Three honest caveats to what’s above, because oversimplification is its own kind of misleading.

First: I described the model as predicting “one word at a time.” Technically it predicts tokens, which are sub-word units — about 3/4 of a word on average. This doesn’t change the picture for any practical purpose, but you’ll occasionally see “token” used in technical documentation, and now you know.

Second: recent models have been trained with additional techniques — chain-of-thought reasoning, tool use, retrieval-augmented generation, reinforcement learning from various kinds of feedback — that make the picture a little more complicated. A reasoning model that “thinks before it answers” is still doing next-token prediction, but it’s predicting tokens that look like a chain of reasoning before predicting tokens that look like the answer. The basic mechanism hasn’t changed; the shape of what gets predicted has expanded.

Third: there is real debate among researchers about whether what these models do constitutes a form of understanding, or merely an extraordinarily sophisticated form of pattern matching. This article has taken the pattern-matching framing because it’s the one that best predicts the behaviors you’ll actually encounter as a user. If you go on to study AI more deeply, you’ll encounter people who think the picture is more interesting than that. They might be right. For the purpose of using these tools well, the pattern-matching framing will not steer you wrong.

The single most important takeaway

If you remember nothing else from this article, remember this:

The model is fluent. Fluency is not truth.

Everything else flows from there. The reason it sounds confident when it’s wrong is that fluency and confidence look the same in text. The reason it agrees with you is that agreement is fluent. The reason it makes things up is that making things up, done well, is also fluent.

Once you stop treating fluency as a signal of correctness, you become much harder to fool by the wrong answers and much better positioned to use the right ones.

That’s where the rest of the curriculum starts.

---

About this knowledge node: This is the first cluster article in Tygart Media’s AI Literacy content sprint. It’s licensed for use in any classroom, training program, custom GPT, or Claude Project as long as attribution is maintained. The pillar article that introduces the sprint is here.