One Class Bans the Internet. The System Bought Everyone ChatGPT.

In one of my classes this semester, the rules are strict in an old-fashioned way. No internet. No outside sources. You work from the material the professor gives you, and that is it.

In another class, the assignment is this blog. Publish in public, use the tools, learn the technology.

Same student, same semester, same campus. Both of those are defensible positions, and the gap between them is where every student I know is currently living.

The contradiction is institutional

Start with what my own university actually says. In May 2024, the SDSU University Senate expanded its definition of plagiarism to cover “representing work produced by generative Artificial Intelligence as one’s own.” Using generative AI that an instructor has not assigned counts as academic dishonesty. Beyond that baseline, instructors are, in the university’s words, “empowered to determine whether students may use genAI in their classes,” and the Center for Teaching and Learning’s recommended syllabus line is simply: “Students should not use generative AI applications in this course except as approved by the instructor.”¹

Now the other half. In February 2025, the California State University system, which SDSU belongs to, announced ChatGPT Edu for more than 460,000 students and over 63,000 faculty and staff across 23 campuses, described at the time as the first systemwide deployment of its kind in the country.² Students received access that July. EdSource reports the first 18-month contract cost $17 million and that CSU renewed it at roughly $13 million a year, over faculty objections that included a petition to cancel it.³

So the system spent tens of millions putting an AI tool in every student’s account, and whether opening it counts as cheating depends on which classroom you walk into that hour.

I am not claiming either side is wrong. The professor banning the internet is testing something real: what you can do with your own head and a fixed set of materials. That is a legitimate thing to measure, and it is one of the few formats that is actually verifiable right now. The system buying AI access is also responding to something real: students graduate into jobs where these tools are standard.

The failure is in the middle. Nobody translated either position into instructions a student can follow across five courses at once.

What students are actually doing

The behavior is not in dispute. In the Lumina Foundation and Gallup 2026 State of Higher Education study, which surveyed 3,801 students in October 2025, 57% reported using AI in their coursework at least weekly and about one in five reported using it daily.⁴

In the UK, the Higher Education Policy Institute’s 2026 survey of 1,054 full-time undergraduates found 95% using generative AI in some way and 94% using it on assessed work. The same survey found only 12% put AI-generated text directly into submitted work.⁵ That gap is the part policies keep missing. Near-universal use is not near-universal ghostwriting.

Those numbers describe a norm, not an epidemic of cheating. The interesting question stopped being whether students use AI. It is which uses a given assignment should count as learning.

Why detection cannot settle it

The obvious reach is for a detector. It does not work, and the reason is technical.

Most detectors score how predictable text is. Language models produce text with low surprise, word after word falling near the most probable option. Human writing tends to be lumpier. The detector measures that smoothness and calls it a probability of machine authorship.

The problem is that plenty of human writing is smooth, especially writing by people still acquiring the language. When Stanford researchers tested seven detectors, they found an average 61.3% false positive rate on TOEFL essays written by non-native English speakers, compared with 5.1% on essays by U.S. eighth graders. Prompting a model to make text “sound more like a native English speaker” made it read as more human to the same tools.⁶

Read those two numbers together. The tool is bad at detecting machines, and the humans it wrongly accuses are disproportionately the ones with the least standing to argue. Vanderbilt disabled Turnitin’s AI detector in August 2023 for exactly this reason. Its math was simple: Turnitin claimed a 1% false positive rate, Vanderbilt had submitted about 75,000 papers the previous year, so roughly 750 student papers would have been wrongly flagged.⁷

Later work has argued the picture is more complicated, including that some of the gap may track subject matter and discipline rather than language background.⁸ Fine. The policy conclusion survives either way: a score from a detector is not evidence of what a student did, and no honest process should treat it as such.

This is the same lesson from the first thing I wrote on this blog, inverted. There, an AI answer was only worth what you could trace to a source. Here, an accusation is only worth what you can trace to evidence. In both cases, the output is confident, the confidence is unearned, and the person on the receiving end is the one who pays.

What I learned watching a hundred people write

For a year I worked as a writing mentor, sitting with undergraduates and their drafts. That job changed what I think this whole debate is about.

Most students who came to me were not stuck on grammar. They were stuck because they did not yet know what they thought. The draft was bad because the thinking was unfinished, and the useful hour was the one where they talked it out and discovered the argument they actually wanted to make. The writing was downstream of that.

That is the part a generated draft skips. Not the honesty, the discovery. You get a competent paragraph without ever hitting the moment where you realize your point does not hold and you have to change your mind. Submit it and you might get the grade. You just did not get the thing the assignment was for.

Which is also why blanket bans feel beside the point to me. A student who asks a model to explain a confusing source, then argues with it, then writes their own paragraph, is doing more thinking than one who silently paraphrases a textbook. Those two are indistinguishable under most syllabus language, and that is the failure.

What the evidence actually says

Most AI policy I have read is written as though the research does not exist. It does, and it is more interesting than either side of the campus argument.

Cheating did not spike. Stanford researchers surveying students at 40 U.S. high schools found that 60% to 70% reported at least one cheating behavior in the previous month, a rate that stayed flat or dipped slightly after ChatGPT arrived.⁹ That is a high number, and it was a high number before. The tool changed the method, not the prevalence. It is high school data from 2023 and the authors called it preliminary, so I would not lean on it alone, but it does undercut the premise that policy is responding to a sudden collapse in integrity.

Unstructured chatbot access can genuinely damage learning. In a randomized trial with roughly 1,000 students across 50 classrooms, researchers gave some students a plain GPT-4 chat interface, some a version with tutoring guardrails, and some nothing beyond their textbook. During practice, the AI groups did far better: 48% better for the plain interface and 127% better for the tutor version. Then came a closed-book exam with no AI. The students who had used the plain chatbot scored 17% worse than the students who never had AI at all. The guardrailed tutor group landed level with the control.¹⁰

That is the clearest statement of the problem I have found. The chatbot was a crutch during practice, the practice scores looked great, and the learning did not happen.

Well-designed AI tutoring can beat the classroom. In a Harvard randomized trial, 194 introductory physics students alternated between an active-learning classroom session and an at-home session with a custom AI tutor built with pedagogical constraints. Median post-test scores were 4.5 for the AI condition against 3.5 for the classroom, from the same baseline, which the authors describe as more than double the learning gain. Students also spent less time and reported higher engagement.¹¹

Same technology, three very different outcomes. What separates them is not whether AI was allowed. It is whether the design made the student do the thinking. A tutor that asks questions and withholds answers produced the best result in the literature. A chatbot that hands over finished answers produced the worst.

That is the distinction a syllabus has to encode, and “no AI” does not encode it.

What is being tried elsewhere

Four approaches, roughly:

Teach it on purpose. Ohio State committed that “every Ohio State student, beginning with the class of 2029, will graduate being AI fluent.” Generative AI basics are taught in the required General Education Launch Seminar, workshops run through the First Year Success Series, and the university published six learning outcomes, including evaluating AI outputs and weighing the ethical and environmental implications of the technology.¹² Whether it works is unknown, but it takes a position and resources it.

Redesign the assessment. Some institutions are shifting weight to in-class writing, supervised exams, oral examinations, and assessing process rather than final products. Reporting describes rising demand for blue books on several large campuses.¹³ This works for verification. It also comes at a cost that falls unevenly, since timed handwritten work disadvantages multilingual students and students with accommodations, which is the same population the detectors already treat worst.

Agree on principles across institutions. In July 2023 the 24 Russell Group universities in the UK published five shared principles: support students and staff to become AI-literate; equip staff to support students using the tools; adapt teaching and assessment to incorporate ethical use and support equal access; uphold academic rigour and integrity; and work collaboratively to share best practice as the technology evolves.¹⁴ It is a statement of direction rather than a rulebook, but every one of those sentences is a commitment a student could hold a university to.

Grade the level of use. The approach I find most sensible is the AI Assessment Scale, which replaces a yes-or-no rule with five defined levels, from no AI at all through full AI use, chosen per assignment and stated up front.¹⁵ Its authors report adoption across hundreds of institutions, and the framework has been picked up by quality agencies in other countries.

The common thread in the ones that work is specificity. The instructor decides what the assignment is for, then says which level of help serves that purpose. “Don’t cheat” is not a policy. It is a mood.

What I would want on a syllabus

Writing this as a student who also builds with these tools, here is the version I would find usable:

1. A level per assignment, not per course. The reflection paper and the take-home analysis do not need the same rule. Say which is which.
2. Specifics about tasks, not tools. “You may use AI to check your grammar and to explain readings. You may not use it to generate your argument or your prose” tells me more than any list of banned apps.
3. Disclosure instead of surveillance. Ask me to write a line at the end saying what I used and how. Cheap, checkable in context, and it makes honesty the easy path.
4. No detector scores as evidence. If authorship is in question, talk to me about my work. I can explain anything I actually wrote.

For my part: I use AI in my own process, including on this blog, and I think saying so plainly should be normal rather than brave. The policy I want is the one I would be happy to follow in public.

Footnotes

1. San Diego State University Library, AI at SDSU, citing the University Senate’s May 2024 revision of the plagiarism definition and the Center for Teaching and Learning’s recommended syllabus statement. ↩

2. California State University, CSU Announces Landmark Initiative to Become the Nation’s First and Largest AI-Empowered University System, February 2025. Deployment figures as reported by Campus Technology, 5 February 2025. ↩

3. EdSource, California State University renews controversial systemwide contract with OpenAI, 2026. ↩

4. Gallup and Lumina Foundation, AI Is Routine for College Students, Despite Campus Limits, published April 2026. Web surveys of 3,801 students, fielded 2 to 31 October 2025. ↩

5. Stephenson and Armstrong, Student Generative AI Survey 2026, Higher Education Policy Institute Report 199, published 12 March 2026. Polling by Savanta, 1,054 full-time UK undergraduates, fielded December 2025, sponsored by Kortext. ↩

6. Liang, Yuksekgonul, Mao, Wu and Zou, GPT Detectors Are Biased Against Non-Native English Writers, Patterns (Cell Press), 2023. Seven detectors tested against TOEFL essays and U.S. eighth-grade essays. ↩

7. Vanderbilt University, Guidance on AI Detection and Why We’re Disabling Turnitin’s AI Detector, 16 August 2023. ↩

8. See, for example, Style as a Confound: False Positives in AI Detection of Non-Native Academic Writing, 2026, which argues the gap may partly reflect subject matter and disciplinary conventions. ↩

9. Stanford Graduate School of Education and Challenge Success, reported in What do AI chatbots really mean for students and cheating?, December 2023. Anonymous surveys at 40 U.S. high schools; the authors describe the findings as preliminary. ↩

10. Bastani, Bastani, Sungu, Ge and Kabakcı, Generative AI without guardrails can harm learning: Evidence from high school mathematics, Proceedings of the National Academy of Sciences, 25 June 2025. Randomized controlled trial, roughly 1,000 students across 50 classrooms. ↩

11. Kestin, Miller, Klales, Milbourne and Ponti, AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting, Scientific Reports, 3 June 2025. 194 students, crossover design, introductory physics. ↩

12. Ohio State University Office of Academic Affairs, AI Fluency, including the six published learning outcomes. ↩

13. Axios, Blue books make an “out of step” campus comeback in the AI era, 2026. ↩

14. Russell Group, Principles on the use of generative AI tools in education, July 2023, adopted by all 24 member universities. ↩

15. Perkins, Furze, Roe and MacVaugh, The Artificial Intelligence Assessment Scale (AIAS): A Framework for Ethical Integration of Generative AI in Educational Assessment, Journal of University Teaching and Learning Practice 21(06), 2024. ↩