AI Detector False Positives: Cases and Student Appeals

👁 113 reads

An AI detector flags an essay. A student says they wrote it. Before the dispute becomes a misconduct finding, someone needs to establish what the software actually measured and what evidence exists beyond its score. A detector result alone cannot establish authorship or prove cheating.

The best-known cases are not all the same kind of failure. Some involve demonstrably old human writing. Others involve a chatbot used as a detector, a university changing its policy, or a court overturning a disciplinary decision. Keeping those categories separate makes the examples more useful to students and faculty.

Sources checked October 14, 2026, by the Quantslant editorial team. This article reports published evidence, without new detector tests. Our coverage of the French prizewinning novel controversy is available separately.

What counts as an AI detector false positive?

A false positive occurs when human writing is classified as AI-generated. To establish one, you need reliable information about the text’s origin. A writer’s objection deserves consideration, but a disputed allegation and a demonstrated classification error are not identical.

There is another possibility: software identifies language that was revised with an allowed tool. The dispute then concerns the meaning of the label and the assignment’s rules. That is why an instructor should ask what help was used, which part of the writing changed, and whether that activity was permitted.

For the language dimension, see our guide to AI detectors and non-native English writers. A concern about a writing pattern must not become a shortcut for judging a person’s honesty.

Texas A&M-Commerce: asking ChatGPT who wrote an essay

In May 2023, an instructor at Texas A&M University-Commerce asked ChatGPT whether it had written students’ assignments. NBC News reported the dispute on May 18, 2023. This was a chatbot being used to make authorship judgments, rather than a dedicated detector assessment.

The university disputed viral claims that the entire class had failed or been prevented from graduating. It described temporary incomplete grades, individual reviews, several students being cleared, and different outcomes within the class. Its account did not support treating every student as one undifferentiated case.

The lesson is methodological. A confident response to an authorship question is not a record of a student’s writing activity. Before relying on any answer, identify the system’s actual function and the evidence it can access. Repeating the question does not create an independent source of proof.

It is also a lesson in reporting. A dramatic social media headline can omit the institution’s explanation and later developments. When discussing an accusation, distinguish the original claim from what the available record ultimately supports.

The US Constitution example: real, but historical

Ars Technica documented on July 14, 2023 that excerpts of the US Constitution were flagged by GPTZero and ZeroGPT. The example demonstrates a false classification in those historical tests, not the performance of every current detector.

A vivid counterexample can disprove an assertion that a tool never makes mistakes. It cannot tell you how often that tool misclassifies contemporary undergraduate essays. That requires a suitable collection of documents with known provenance and a defined evaluation method.

Do not assume the original screenshot can be reproduced with today’s service. We are reporting an earlier demonstration, not claiming to have rerun it. Software versions, inputs and scoring rules need to be documented before two results can be meaningfully compared.

Why Vanderbilt disabled Turnitin’s AI detection feature

Vanderbilt’s August 16, 2023 announcement described disabling Turnitin’s AI detection feature after reviewing reliability and other concerns. Its illustration of 750 potentially misidentified papers applied a then-advertised 1% false-positive rate to approximately 75,000 submissions. It was not a finding that 750 students had been falsely accused.

That distinction matters for campus policy. A projected number of errors is a risk estimate under assumptions. A count of misconduct findings is an institutional outcome. A count of overturned findings is something else again.

Before purchasing or deploying a screening system, an institution should ask what happens after an alert. Who reviews the underlying submission? What explanation reaches the student? Can the student challenge the interpretation? Those practical questions should be answered before a percentage is allowed to influence a grade.

Newby v. Adelphi: what the court actually decided

In Matter of Newby v. Adelphi University, decided January 28, 2026, a student challenged an academic integrity finding involving a reported 100% Turnitin score. The student disputed the allegation; Adelphi defended its decision as rationally supported. The court annulled the finding and appeal denial, ordered the record expunged, and rescinded the sanction.

The opinion addressed the basis for the decision and failures in the university’s procedures. It did not establish a nationwide ban on AI detectors or scientifically determine their accuracy. This account describes that decision, without claiming a later appellate outcome.

For students, the useful point is to read the actual decision rather than a headline about a detector being “defeated.” An appeal can turn on ignored evidence, the opportunity to respond, or whether an institution followed its own rules. A legal ruling and a benchmark study answer different questions.

For faculty, the lesson is equally practical: a review should explain why each relevant piece of evidence matters. A conflicting detector score should not be ignored, but it should not automatically be treated as conclusive proof either.

Why a low error rate can still produce disputed flags

Consider a deliberately simplified example, not a measurement of any product. A batch contains 1,000 essays: 990 are human-written and 10 are generated. Suppose a detector flags all 10 generated essays but also incorrectly flags 1% of the human essays. That produces roughly 10 false alerts alongside 10 correct alerts.

In this hypothetical batch, about half the flagged essays would be human-written despite the 1% false-positive assumption. The calculation is approximately 990 multiplied by 0.01, rounded for illustration. Change the prevalence or detection rate and the result changes.

This does not show that half of real detector alerts are wrong. It shows why an error rate is not automatically the probability that a particular accused student cheated. To interpret the result, you need information about the population being screened and the evaluation conditions.

How to respond to a false AI accusation

The following is an evidence-organizing approach, not a promise that an appeal will succeed. Begin by preserving the exact submission. If you rewrite it immediately, you can lose the reference point for the discussion.

• Request the report. Ask for the tool, date, analyzed file and passages at issue.
• Identify the rule. Find the assignment instructions and applicable academic integrity policy.
• Collect existing process evidence. Drafts, source notes, outlines and feedback may help explain the work.
• Describe assistance accurately. Separate tutoring, proofreading, translation and generation.
• Build a brief timeline. Connect each document to an identifiable stage of the assignment.
• Check the response deadline. Follow the institution’s actual procedure rather than advice from an unrelated case.

If you wrote offline or in one file, say so. Missing version history does not supply a complete account of authorship. Explain what records do exist and what you can demonstrate about your reasoning. Never manufacture drafts or alter timestamps to fill a gap.

A useful response focuses on the disputed passages. For example, explain why you used a particular source or changed an argument between drafts. Sending hundreds of files without context can make it harder to understand the material that actually matters.

A fair review needs more than competing percentages

We recommend treating the report as a question to investigate. Keep the observed fact, the inference and the final finding separate. “The tool flagged this passage” is a reportable observation. “This person violated the policy” requires a supported assessment of both conduct and rules.

An interview can help clarify the writing process, but it should have a clear purpose. Ask about research choices and revisions. Allow the writer to supplement an answer with existing records. Do not turn nervousness or difficulty speaking spontaneously into an authorship test.

For book submissions, the relevant standards may differ from classroom rules. Our guide to AI policies for writing contests, literary prizes and publishers explains why an allowed proofreading workflow cannot be assumed acceptable everywhere.

Related: my test of the Orélien novel in three detectors, what to do if you are falsely accused.

FAQ

Can an AI detector prove I used ChatGPT?

A score alone cannot establish your writing process or a policy violation. Ask what the indicator means and what other evidence supports the allegation.

Was the whole Texas A&M-Commerce class failed?

The university disputed that viral account. The NBC report described temporary incomplete grades and individual reviews. Avoid repeating the blanket claim as an established outcome.

Did Newby v. Adelphi outlaw AI detection?

No. The decision addressed a particular academic integrity finding and the university’s handling of it. It was not a general ban or an accuracy study.

Will another detector clear my name?

A different result can demonstrate disagreement between systems. It does not automatically settle authorship. Include process evidence and address the actual allegation.

Should I intentionally make my writing less polished?

No. Preserve clear writing and an honest record of how it developed. Deliberately changing the style to chase a score makes the process harder to explain.