Here is the answer up front: the AI-detection tools schools rushed to buy do not reliably do the job. They flag real human writing and miss real AI writing, and in January 2026 a New York court ordered a university to erase a plagiarism finding because a single detection score was not proof of anything. The response that actually works is not a better detector. It is assessment where the work is visible, plus a firm rule that no score alone can punish a student.
This is where AI in schools actually stands in 2026. Cheating is up, students know it, detectors are unreliable in both directions, and the teachers getting ahead of the problem have stopped trying to catch AI and started designing work that AI cannot fake.
One essay, one score, and a court order
Orion Newby wrote his paper himself. Turnitin’s AI detector said he did not.
Newby was a first-semester student at Adelphi University, writing a World Civilizations assignment on Christianity and Islam. His professor ran the paper through an AI-detection tool, and Turnitin marked it as fully AI-written. Newby ran the same paper through two other detectors; both said a human wrote it, according to court records. The university gave him a zero, denied his appeal, and a second allegation raised the possibility of expulsion.
His parents hired an attorney and sued. A New York court ruled the plagiarism finding was without merit and ordered the university to remove the charge from his record. Newby told Inside Higher Ed (Feb 11, 2026) that it “feels incredible to finally have my name cleared.” His family said they had spent six figures on legal fees to get there.
Two things make this case matter beyond one student. First, Newby had tutoring support through a university program for students with learning and neurological disabilities, which is exactly the kind of legitimate help an AI detector cannot see and cannot account for. Second, and more important, the ruling did not decide whether AI wrote the paper. It decided that a single detection score was not enough to punish someone, and that a student facing serious consequences is entitled to real due process first.
As the research-administration group CASRAI put it in a July 2026 review, the question has shifted from “how do these tools work” to “how much weight can this score legally bear.” The answer, increasingly, is not much.
Why a detection score was never proof
The accuracy problem was documented years before any of this reached a courtroom. The most cited study, by Liang, Yuksekgonul, Mao, Wu, and Zou at Stanford, was published in the journal Patterns in 2023. It tested seven GPT detectors against 91 TOEFL essays written by non-native English speakers. The average false-positive rate was 61.3%, and more than 91% of those real, human-written essays were flagged as AI by at least one of the tools. The same detectors produced a near-zero false-positive rate on essays written by native-English-speaking U.S. eighth-graders.
The study found something else that should have ended the debate. Light, prompt-based rewriting was enough to push genuinely AI-generated text below the detection threshold. In other words, the same tools were over-flagging real human writing and under-flagging real AI writing at the same time.
That is not a flaw that got patched. OpenAI itself built an AI text classifier in early 2023 and shut it down by July of that year. Its own numbers: it correctly identified 26% of AI-written text, and it wrongly labeled human-written text as AI 9% of the time. A tool that is wrong in both directions is not evidence. It is a hint.
Even the vendors agree, quietly. Turnitin, which powers most school detection, reports a false-positive rate of under 1% for documents it judges to be more than 20% AI-written, based on its own testing. It also tells institutions plainly that the indicator should not be used by itself to penalize a student. That instruction is the whole case in one sentence.
The false positives are not spread evenly
Detection errors do not land on everyone equally.
- Non-native English writers. The Stanford study is the clearest example: the tools read plain, careful, grammatically regular prose as machine-like, and that is exactly how a second-language writer is taught to write.
- Students who get legitimate help. Newby’s tutoring support is not an edge case. Accommodations, editing centers, and structured writing programs all produce cleaner text.
- Students who simply write in a flat register. Formulaic academic writing, the kind schools reward, is the pattern detectors associate with AI.
The cost of a false positive is not a private embarrassment. It is a zero, an expulsion threat, a permanent record, and in Newby’s case a six-figure legal bill.
What the cheating numbers actually say
The panic is not imaginary. Use really is up, and the best data we have comes from RAND’s American Youth Panel, a nationally representative survey released March 17, 2026.
- Between May and December 2025, the share of middle school, high school, and college students using AI for homework rose from 48% to 62%.
- 67% of students said using AI for schoolwork harms critical thinking, up from 54% earlier in the year.
- 71% reported using at least one AI tool for school. ChatGPT led at 53%, and Google Gemini use more than doubled to 28%.
- Only about one third said their school had a schoolwide AI policy. Many said the rules changed from teacher to teacher.
Here is the detail that should reframe the whole debate. Students are not hiding this. Only 45% considered getting direct answers to homework to be cheating, while about 80% said using AI to understand an assignment was not cheating at all. Most students do not think the common uses are dishonest. The real gap is that most schools never wrote the rule down, so everyone is guessing.
What teachers are doing instead
The teachers who are handling this best have stopped policing output and started designing work where the process is visible. That is the pivot that makes detection irrelevant.
- In-class writing. The essay happens in the room, on paper or on a locked-down device, where the only available assistant is the student’s own brain.
- Oral check-ins. A five-minute conversation about your own draft is very hard to fake and very easy to grade.
- Graded drafts with version history. Google Docs and Word both keep a full edit trail. A student who can show the messy middle of a paper has the evidence no detector can see.
- Open-book tasks that reward reasoning. If the question cannot be answered by retrieving a fact, a chatbot is less useful than the student thinking.
There is a second, quieter move happening at the institutional level. Universities are starting to separate text-matching from AI detection, because the two tools have very different reliability records. Text-matching, the classic similarity check for copied sources, is mature and defensible. AI-writing detection is neither. Curtin University announced it would switch off Turnitin’s AI-writing feature from January 1, 2026, while keeping the originality check running. It is not alone in that reasoning.
The rule every school should adopt
If there is one policy lesson from the Newby case, it is this: no AI-detection score, on its own, can support a finding of misconduct. A number can start a conversation. It cannot end one with a punishment.
That means a human review and an appeal step before any consequence, every time. For students, it means keeping the paper trail: drafts, revision history, notes, and any messages with a tutor or advisor. That process record is the one form of evidence a detector cannot generate and cannot dispute.
If you are flagged, do not argue about the percentage. Ask what the tool is being used as, bring your drafts, and ask the institution to show you the written policy it is applying. In most cases, that policy does not actually say a score alone is enough.
The same blind spot, one year earlier
AI systems judging people on thin evidence is not new to classrooms. We broke down how hiring algorithms screen out qualified candidates for reasons no human ever sees.
FAQ
Can AI detectors reliably tell if a student used ChatGPT?
No. They produce both false positives (real human writing flagged as AI) and false negatives (real AI writing missed). The Stanford study found a 61.3% average false-positive rate on essays by non-native English speakers, and simple rewriting defeats the same tools. Treat any single score as a signal to look closer, never as proof.
Do schools still use Turnitin?
Mostly yes, but the pieces are being pulled apart. Similarity checking for copied sources is reliable and staying. AI-writing detection is the part facing pushback, with universities like Curtin disabling it while keeping the originality check. Turnitin’s own guidance says the AI indicator should not be the sole basis for a penalty.
What should a student do if they are flagged for AI?
Do not fixate on the score. Ask whether it is being used as sole evidence, request a human review, and bring your drafts and revision history. If your institution has a written policy, ask them to show it. Keep every version of the work as you go, so the proof exists before you need it.
Is using AI for schoolwork always cheating?
No, and most students do not think it is. RAND found that about 80% of students consider using AI to understand an assignment to be acceptable, while only 45% call getting direct answers cheating. The problem is that most schools never defined the line, so the rule changes from classroom to classroom.
The Short Version
- Detectors do not work. They flag human writing and miss AI writing. A 2023 Stanford study found a 61.3% false-positive rate on essays by non-native English speakers; OpenAI’s own classifier was wrong in both directions and was retired in 2023.
- A court agreed. In early 2026, a New York court reversed Adelphi University’s plagiarism finding against Orion Newby and ordered it removed from his record. The ruling did not say whether AI wrote the paper. It said a single score is not enough to punish a student.
- Use is genuinely up. RAND found AI homework use rose from 48% to 62% between May and December 2025, and 67% of students now worry it harms critical thinking.
- The gap is policy, not honesty. Only about a third of students say their school has a schoolwide AI rule. Most do not consider the common uses cheating.
- The fix is not a better detector. In-class writing, oral check-ins, graded drafts with version history, and a hard rule that no score alone can trigger a penalty. Make the process visible and the detection question stops mattering.
About The Author
