Should Schools Ban AI? The Bans Aren’t Holding, and Here’s What Works Instead

Should schools ban AI? It is the question every district, dean, and department head has been wrestling with since ChatGPT landed. However, it is the wrong question.

Here is the short answer: banning AI does not hold, and it does not prepare students for a world that already runs on it. In fact, the schools spending their energy on enforcement are spending it on something that does not work. By contrast, the schools redesigning how they teach are the ones actually getting students ready. Ultimately, the real divide is not ban versus allow. It is policing versus teaching.

This is not a hunch. Rather, there is hard evidence behind every part of it, and most of it points the same direction.

TL;DR: Key Takeaways

• Bans don’t hold. Even New York City, the largest US school district, reversed its ChatGPT ban within one semester.
• The enforcement tool is broken. OpenAI shut down its own AI detector, and NYU, Vanderbilt, and Waterloo switched theirs off for false-flagging honest students.
• AI lies with confidence. GPT-4 still fabricates citations, and 2026 audits show newer models have not fixed it.
• The real skill is verification, and it is teachable. Teaching students to check AI’s sources, not banning the tool, is what prepares them for the workplace they are heading into.

Should schools ban AI? The bans aren’t holding

Start with the biggest test case we have. In December 2022, New York City’s Department of Education, the largest school system in the country, blocked ChatGPT on its networks and devices¹. The reasons will sound familiar: concerns about accuracy, and a worry that the tool “does not build critical-thinking and problem-solving skills”.

However, it lasted about one semester. On May 18, 2023, Chancellor David Banks reversed the decision in an op-ed, writing that the “knee-jerk fear and risk overlooked the potential of generative AI to support students and teachers,” and that “our students are participating in and will work in a world where understanding generative AI is crucial” ². Moreover, rather than simply lift the block, the district committed to giving educators implementation resources, a shared repository of what worked, and toolkits for classroom conversations about AI ethics ².

In other words, if the nation’s largest district could not make a ban stick past one semester, that tells you something about the strategy itself.

Furthermore, NYC is not alone in walking back enforcement. Across higher education, the specific tools built to catch AI use are being switched off. For example, Washington State University cancelled its Turnitin AI-detection contract in early 2026, explicitly aligning with a group of R1 peers that had already dropped detectors, including UC Berkeley, Colorado State, Indiana, Michigan State, Oregon State, and the University of Washington ⁴[3].

AI detection can’t enforce a ban on AI in schools

A ban is only as good as your ability to enforce it, and enforcement rests on AI detection. That is where the whole strategy quietly collapses.

Even OpenAI couldn’t build a working detector

Consider who gave up first. OpenAI, the company that makes ChatGPT, shut down its own AI-detection tool in July 2023 because of its “low rate of accuracy” [4]. Specifically, at launch it caught only 26% of AI-written text while falsely flagging 9% of human writing as machine-made [4]. When the company that built the engine cannot build a reliable radar for it, that is not a bug you patch with a better detector.

False positives punish honest students

The classroom data is worse, because the false positives are not random. Instead, they land hardest on the students least able to defend themselves.

• A Stanford study found detectors flagged 61% of essays by non-native English speakers as AI-generated, versus about 5% for native speakers [5]. Same tools, same day.
• Moreover, the same study showed how trivially detectors are fooled. Adding one instruction, “elevate the text with literary language,” crashed detector accuracy from nearly 100% down to 13% [5]. As a result, the tools punish honest writers and wave through anyone who knows a single trick.
• Vanderbilt ran the math on its own usage, roughly 75,000 papers in 2022. Even at Turnitin’s advertised 1% error rate, that meant about 750 students falsely accused in a single year, so the university disabled the detector [6].
• Similarly, NYU found the real-world false-positive rate closer to 4%, or 1 in every 25 honest students, and disabled it too [7]. The University of Waterloo dropped it after internal tests flagged 100% human-written text as “100% AI-generated” [3].

Better models won’t fix it

This is not just an engineering problem that better models will fix. In fact, researchers proved mathematically that reliable AI-text detection is “not possible in the limit,” because recursive paraphrasing defeats watermarked, neural, zero-shot, and retrieval-based detectors alike, and as models improve the statistical gap between human and machine text shrinks toward zero [8]. Likewise, a separate evaluation of 14 detection tools found them “neither accurate nor reliable,” biased toward calling text human-written, and easily degraded by simple obfuscation [9].

Therefore, you cannot police your way out of this. The radar does not work, and it never will.

Isn’t AI destroying critical thinking?

This is the fear underneath most bans, and it deserves an honest answer rather than a slogan. Ultimately, the honest answer is that it depends entirely on how the tool is used.

Used passively, AI absolutely can be an atrophy machine. For instance, if an assignment only asks a student to find a source, summarize it, and cite it, AI automates that overnight, the student offloads it, and no thinking happens. And right now, most students are using it passively. Among surveyed university students, only 27.5% consistently verify AI-generated content, while 60.8% have submitted AI text with minimal editing [10]. Similarly, among teens aged 13 to 18, 70% have used generative AI, and among those using it for schoolwork, nearly half did so without their teacher’s permission [11].

That last number is the tell. In short, verification is a teachable skill, and teaching it changes behavior.

So the move is to push the work up a level. Instead of “summarize three sources,” the assignment becomes “here is what the AI told you, now find where it is wrong.” And it will be wrong, which we will get to in a moment. Because a student who has to catch a fabricated citation is evaluating credibility and defending a judgment, which is harder and more valuable than any book report. In other words, AI did not kill critical thinking. Lazy assignment design did, and AI just made it impossible to hide.

Can we even trust AI? Not blindly, because it lies with a straight face

The most dangerous thing about AI is not that it fails. Rather, it is that it fails confidently, handing you a polished, plausible answer with a real professor’s name attached to a paper that does not exist.

The evidence on fabricated citations

The research on fabricated citations is sobering:

• For example, when researchers asked ChatGPT for medical references, 69% of the citations were completely fabricated, complete with real authors’ names, plausible titles, and real-looking journals [12].
• In addition, another study found only 7% of ChatGPT’s medical citations were both real and accurate, with a wrong identifier appearing in 93% of them [13].
• Notably, this is not just the old models. A Scientific Reports analysis found GPT-4 still fabricated 18% of citations outright, and of the real sources it did cite, 24% had major errors in the journal, volume, or page [14].

When AI hallucinations reach the courtroom

These are not abstract risks. Instead, they are producing real consequences for people who should know better. For instance, a federal court fined lawyers $5,000 for a brief built on six ChatGPT-invented cases [15]. In a 2026 Oregon case, a court imposed $110,000 in sanctions and dismissed a $12 million lawsuit over 15 hallucinated citations [16]. In 2026 alone, a California appellate court affirmed $6,000 in sanctions against a firm and three attorneys [17], and a Mississippi federal judge sanctioned lawyers on both sides of a case for AI-hallucinated citations [18]. Consequently, if trained attorneys with everything on the line get fooled, a 19-year-old writing at 2am does not stand a chance.

So can we trust AI? Not blindly, and anyone who tells you otherwise is selling something. Therefore, the right answer is not “trust it” or “ban it.” It is “verify it.” Trust the sources you can check.

Banning AI is a preparation problem, not a cheating problem

Here is the part that bans get exactly backwards. The workplace students are heading into already runs on AI. For example, a 2025 survey of 3,000 workers across the US, UK, and Germany found 78% already use AI in their role, and 97% of those use it weekly [19]. Likewise, a separate 2025 survey of more than 1,300 organizations found roughly 95% use AI, with 58% using it daily [19].

As a result, a blanket ban does not protect students from that world. Instead, it sends them into it missing a core skill, while pretending the tool does not exist. It is like teaching someone to drive by hiding the car.

That said, “just let them use it” is equally lazy, because most of them are not checking. The goal is not AI-proof assignments. Rather, it is AI-literate students. In practice, that means guided use, with verification built into the work: show your sources, defend your reasoning, and flag what the AI got wrong.

What works instead of banning AI: teach verification

If detection is dead and bans do not hold, what is left is the thing schools should have been building all along. The skill that matters now is not finding information. Instead, it is interrogating it. Can you trace a claim back to a real source? Can you tell when a citation is a ghost?

Verification is the new literacy. In five years, checking an AI’s sources will be as routine as running spell check is today. As a result, the schools that teach it now will graduate students who can think with these tools instead of being fooled by them.

This is the problem CoChat was built to solve, and it is worth being specific about how, because “makes it easier” is not an answer. CoChat is an AI research workspace where verification is baked in. Specifically, when a student is working on a paper, it checks citations against real academic databases, CrossRef, Semantic Scholar, OpenAlex, and arXiv, so a hallucinated source gets caught before it ever reaches the bibliography [20]. Moreover, it reads the full text of sources, not just the abstract. And it turns research into things students actually use: literature review tables, citation-managed exports, and flashcards pulled from the findings.

To be honest about the scope, this is assistive, not magic. For example, on a deep search, CoChat screens up to 1,000 papers and reports on the most relevant 50, so a human still makes the final call on what belongs. That is the point. The tool does the checking a student cannot do at scale, and the student does the judging that no tool should do for them. In short, it is the difference between an AI that hands you an answer and one that shows you the receipts.

The question was never whether students use AI

It was whether they can tell when it is wrong.

Banning AI answers the first question and ignores the second, which is the one that actually determines whether a student is ready for college, for work, and for a world where confident, fluent, fabricated information is everywhere. Detection cannot enforce a ban. Bans do not hold. And neither one teaches the skill that fixes the real problem.

So the honest answer to “should schools ban AI” is no. Stop policing. Instead, start teaching students to verify. That is the assignment now.

Want to see what verification-first research looks like in practice? Explore CoChat or read more about how we fixed citation hallucinations.

---

References

1. Chalkbeat. “NYC bans access to ChatGPT on school computers, networks.” January 3, 2023.
2. Chalkbeat. “ChatGPT caught NYC schools off guard. Now, we’re determined to embrace its potential.” (David Banks op-ed). May 18, 2023.
3. University of Waterloo. “Discontinuing use of AI detection functionality in Turnitin.” September 2025. (Washington State University and R1 peer detail, February 2026.)
4. OpenAI. “New AI classifier for indicating AI-written text.” January 31, 2023 (discontinued July 20, 2023).
5. Liang W, Yuksekgonul M, Mao Y, Wu E, Zou J (2023). “GPT detectors are biased against non-native English writers.” Patterns. DOI: 10.1016/j.patter.2023.100779
6. Vanderbilt University. “Guidance on AI Detection and Why We’re Disabling Turnitin’s AI Detector.” August 16, 2023.
7. New York University, Office of the Provost. “Disabling the AI Tool in Turnitin.” September 7, 2023.
8. Sadasivan VS, Kumar A, Balasubramanian S, Wang W, Feizi S (2023). “Can AI-Generated Text be Reliably Detected?” arXiv:2303.11156.
9. Weber-Wulff D, et al. (2023). “Testing of detection tools for AI-generated text.” International Journal for Educational Integrity. DOI: 10.1007/s40979-023-00146-z
10. Naik GR, Pednekar R, Shaikh M, Khan U (2026). “Artificial Intelligence (AI) Tools in Higher Education: An Empirical Study.” International Journal of Computer Science and Engineering. DOI: 10.5281/zenodo.20683366
11. Common Sense Media (2024). “The Dawn of the AI Era: Teens, Parents, and the Adoption of Generative AI at Home and School.” Survey of 1,045 parent-teen dyads; 70% of teens have used generative AI, and 46% of those using it for schoolwork did so without permission.
12. Gravel J, D’Amours-Gravel M, Osmanlliu E (2023). “Learning to Fake It: Limited Responses and Fabricated References Provided by ChatGPT for Medical Questions.” Mayo Clinic Proceedings: Digital Health. DOI: 10.1016/j.mcpdig.2023.05.004
13. Bhattacharyya M, Miller VM, Bhattacharyya D, Miller L (2023). “High Rates of Fabricated and Inaccurate References in ChatGPT-Generated Medical Content.” Cureus. DOI: 10.7759/cureus.39238
14. Walters WH, Wilder EI (2023). “Fabrication and errors in the bibliographic citations generated by ChatGPT.” Scientific Reports. DOI: 10.1038/s41598-023-41032-5
15. Mata v. Avianca, Inc., S.D.N.Y. (June 2023). $5,000 sanction for six fabricated ChatGPT citations.
16. Couvrette v. Wisnovsky, No. 1:21-cv-00157-CL (D. Oregon, 2025 to 2026). Judge Mark D. Clarke dismissed plaintiff Joanne Couvrette’s claims with prejudice as a terminating sanction (Dec 12, 2025) and approved $110,204.38 in combined sanctions (a $15,500 fine on lead counsel Stephen Brigandi plus $94,704.38 in fees), over 15 nonexistent cases and 8 fabricated quotations; the ~$12 million case was dismissed. (Family dispute over Valley View Winery.)
17. Quinteros v. Harbor Distributing, LLC, No. A174202 (California Court of Appeal, June 11, 2026). Affirmed $6,000 in sanctions ($5,000 payable to defendants, $1,000 to the court) against Lipeles Law Group and three attorneys.
18. Withers v. City of Aberdeen, No. 1:24-cv-00218 (N.D. Miss., June 8, 2026). Senior Judge Sharion Aycock sanctioned lawyers on both sides for AI-hallucinated citations: pro hac vice counsel Kathleen M. Wilson fined $3,500 and Kathryn Y. Williams $2,500, local counsel $1,000 each, with two-year bars from the district.
19. SnapLogic / 3GEM “AI at Work” survey (3,000 workers, US/UK/Germany); American Management Association 2025 AI survey (1,365 respondents).
20. CoChat verifies citations against a database of more than 200 million papers (via CrossRef and Semantic Scholar) before a source reaches your draft.