Should We Use AI Detectors?

Not as evidence in a misconduct case, and not as an institution's first line of defense. A detector score cannot be validated against ground truth, cannot be cross-examined, and does not meet the balance-of-probabilities standard that academic integrity findings require. Worse, its errors are patterned rather than random: the strongest controlled study in the knowledge base found that detectors flag honest, guideline-compliant AI editing far more readily than unmodified student prose, while deliberate evasion passes almost untouched — and the students most likely to be flagged are non-native English writers.(Why AI Detection Fails for Academic Integrity)(Detecting the Undetectable? Reassessing Academic Misconduct Procedures in the Era of Generative AI) Bassett et al. (2026) go further and argue detection should not be used in education at all, because the technology cannot tell "work created with AI" from "work created by AI." This page is for the people who have to decide: instructors, academic integrity officers, and administrators. The design-side playbook lives in How Can I Reduce AI Cheating in My Course?.

1. The accuracy numbers do not support a finding

Three independent lines of evidence converge on the same conclusion.

The false-positive rate on legitimate work is high, and it punishes transparency. A controlled study of 642 published English abstracts across four domains and two time periods found that at a 0.50 threshold, two commercial detectors flagged guideline-compliant light AI editing at 38–80%, flagged unmodified 2023–25 originals at 9–15% (non-STEM far above STEM, p<0.001), and — once text had been run through a humanizing service — caught fewer than 4% of AI-labeled rewrites, a false-negative rate above 96%. The authors describe this as an integrity catch-22: students who disclose and edit lightly are the ones the tool catches, while students who deliberately evade it are not. Their recommendation is explicit — detector scores should never serve as standalone misconduct evidence.(Why AI Detection Fails for Academic Integrity)

The best independent benchmark does not reach usable accuracy. In the most comprehensive early Benchmark, none of fourteen tools reached 80% accuracy, and simple paraphrasing, minor editing, or humanizing services roughly halve even that performance. At realistic base rates, false positives outnumber true positives.(Detecting the Undetectable? Reassessing Academic Misconduct Procedures in the Era of Generative AI)

And it misses wholesale AI use entirely when it counts. In a covert field study, researchers injected wholly AI-generated submissions into a live online examination system across five psychology modules: 94% went undetected, and the AI work on average outscored the real students. Contract cheating already demonstrated the same structural problem — it leaves no reliable trace to find.(Detecting the Undetectable? Reassessing Academic Misconduct Procedures in the Era of Generative AI)

Accuracy also varies by task in ways a policy cannot anticipate. When researchers tested whether generative AI can reliably detect its own output, detection was dependable for programming and longer reflective writing but poor for short answers, where the model often judged its own text as more human-like than authentic student work — and minor prompt variations sharply reduced accuracy.(Distinguishing Artificial from Authentic: Evaluating LLMs for Detecting LLM-Generated Content) Any threshold you set will hold for some assignments and fail for others.

Detection has never been clearly better than a careful human, and the margin is not the point. Leaton Gray, Edsall and Parapadakis (2025) report machine detection of AI or paraphrased text at roughly 80% against 78.4% for human reviewers, and cite evidence that AI-generated text has passed as human-authored in up to 80% of cases — which they read as a margin far too narrow to ground a misconduct finding, since the machine's advantage disappears into the same error band the human brings. Their review also undercuts the assumption that detection deters the capable: Krou et al.'s meta-analysis finds self-efficacy correlates negatively with cheating while actual ability does not correlate inversely with it at all, so students who could do the work may cheat when they judge the assessment unfair. Detection is therefore neither a reliable instrument nor an obvious deterrent.

2. The errors are patterned, and they land on the wrong students

This is the part that should decide the question for anyone responsible for equity.

The features detectors treat as signals of AI — long-token density, academic word frequency, uniform style — are also features of competent second-language writing, so detectors misclassify non-native English speakers systematically rather than randomly. The harm of misclassification therefore lands on students who are already disadvantaged.(Detecting the Undetectable? Reassessing Academic Misconduct Procedures in the Era of Generative AI) The same skew appears by discipline: non-STEM abstracts were flagged far above STEM ones in the abstract study, which is a property of the writing conventions of those fields, not of their authors' conduct.(Why AI Detection Fails for Academic Integrity) In effect, a detector is a style test, and the style it punishes correlates with language background, discipline, and register — not with whether a student used AI.

That is an equity problem before it is a technical one, and it is also a Trust problem. Because detectors are unreliable and formal processes demand detection-grade proof that is functionally unavailable, faculty end up with what one practitioner account calls "suspicion without recourse," while students rationalize their own use and case files stall.(The Best Response to Student AI Use Is Not Detection, It Is Dialog)

3. Why a score cannot carry an integrity case

If you sit in a hearing, this is the section that matters.

• The evidential standard is not met. Academic misconduct findings require evidence meeting the balance of probabilities. Detector scores — alone or in combination with linguistic markers, style comparisons, Large Language Models (LLMs) judgments, or a student's silence — do not satisfy that standard.(Heads We Win, Tails You Lose: AI Detectors in Education)
• Silence and speech are both being misused. Students retain the right to silence; refusing to respond does not tip the scales against them. The legitimate question is not "did you use AI?" but whether the student can demonstrate the learning the Assessment claims to measure, which an oral response can answer.(Heads We Win, Tails You Lose: AI Detectors in Education)
• The probability is unverifiable. Unlike a spam filter or a medical test, a detector's output cannot be independently checked: in real submissions there is no ground truth about how the text was produced, so validation becomes circular, and false-positive/false-negative metrics apply only in controlled tests.(Heads We Win, Tails You Lose: AI Detectors in Education)
• The reference data is outdated by construction. Detectors are trained and tested on pre-generative-AI human writing — Turnitin's own validation used 700,000 pre-2019 papers — which assumes that corpus reflects current student prose. That assumption is unverified, and it shifts with every model release.(Heads We Win, Tails You Lose: AI Detectors in Education)
• The tool cannot be interrogated. Detectors publish no thresholds or training data and do not permit independent replication, so a flagged student cannot answer or cross-examine the accusation. You cannot defend a finding you cannot explain.(Detecting the Undetectable? Reassessing Academic Misconduct Procedures in the Era of Generative AI)
• The binary is the wrong question. Students' work is frequently created with, not by, AI, across a hybrid continuum, and policies that say "in assessment" rarely define when an assessment begins — leaving enforcement to subjective judgment rather than principled criteria.(Heads We Win, Tails You Lose: AI Detectors in Education)
• You take on data risk. Detectors store student work on third-party servers, sometimes overseas under weaker Privacy protections, creating breach, retention, and commercial-exploitation exposure your institution owns.(Heads We Win, Tails You Lose: AI Detectors in Education)
• More monitoring can backfire. Deterrence runs through a detector's discrimination between hidden use and legitimate work, not its raw catch rate. When extra sensitivity produces more new false positives than new true positives, stronger monitoring makes concealment relatively more attractive — you spend credibility on students who did nothing wrong, faster than you identify the ones who did.(Assessment Design Under Imperfect Information: Generative AI, Disclosure, and Student Response in Higher Education)

4. What to do instead

Ask for verification, not provenance. Grand Canyon University moved from "Did this student use AI?" to "Can this student demonstrate understanding of what they submitted?", using short conversations, early drafts, and recorded explanations, implemented institution-wide in fall 2025. The argument is practical: faculty are already qualified to judge understanding, and the shift restores their authority instead of leaving them waiting on proof that will never arrive.(The Best Response to Student AI Use Is Not Detection, It Is Dialog)

Make at least one high-stakes task unaided. Asynchronous oral assessments — just-in-time prompts with brief, time-limited recorded responses graded against embedded rubrics — performed comparably to in-person multiple-choice exams in one study and significantly better in another, with students reporting more active preparation and higher perceived professional relevance. For an Administrators, the relevant property is that this is scalable without proctoring.(Asynchronous Oral Assessments: Enhancing Integrity, Engagement, and Communication in the AI Era)

Redesign tasks so AI shortcuts are less attractive. A case study of undergraduate economics students found only about a third reported any AI use, with disclosure rarer still — and non-disclosure read as rational caution under ambiguous policy rather than dishonesty. Those same students favored real-world, data-based tasks as the fix.(Beyond Detection: How Students Use—and Hide—AI in Online Assessments and What Authentic Tasks Can Do About It) Expecting, declaring, and scrutinizing AI use beats policing it, because authenticity has to be designed rather than enforced.(Beyond Detection: Redesigning Authentic Assessment in an AI-Mediated World)

Make honest disclosure the safe option. Vague or punitive policies drive concealment; specific declaration frameworks tied to cognitive stages, paired with an assurance that truthful disclosure is not penalized, get more information out of students than surveillance does.(Addressing student non-compliance in AI use declarations: implications for academic integrity and assessment in higher)("Should I Tell My Teacher?" Student AI Disclosure Practices, Stigma, and Self-Regulated Learning in Higher Education)

5. The objections you will hear

• "Our vendor advertises a 1% false-positive rate." Vanderbilt's licensed detector claimed exactly that, and when the university sought to validate the figure it could not — 1% of 75,000 annual submissions implied roughly 750 mislabelled students — so it disabled the tool. Independent benchmarks put no tool above 80% accuracy.(Detecting the Undetectable? Reassessing Academic Misconduct Procedures in the Era of Generative AI)
• "We only use it as one piece of evidence." The standard is not "some evidence" but evidence meeting the balance of probabilities, and detector scores combined with markers, style comparisons, LLM judgments, or silence still do not satisfy it. If you would not say the reasoning out loud in a hearing, it is not evidence.(Heads We Win, Tails You Lose: AI Detectors in Education)
• "If we stop detecting, cheating wins." Detection is not what stops the determined: 94% of injected AI submissions passed a live online exam unnoticed, humanizing defeats detection more than 96% of the time, and contract cheating already left no reliable trace.(Detecting the Undetectable? Reassessing Academic Misconduct Procedures in the Era of Generative AI) What detectors do catch, with measured reliability, is honest light editing and non-native prose.(Why AI Detection Fails for Academic Integrity)
• "We only use it privately, to start a conversation." That is the least harmful use, and it still costs something: the students most likely to be flagged are the least likely to be misusing AI, so suspicion lands disproportionately on the wrong people.(Detecting the Undetectable? Reassessing Academic Misconduct Procedures in the Era of Generative AI) Grand Canyon's answer was to stop starting from suspicion at all.(The Best Response to Student AI Use Is Not Detection, It Is Dialog)

6. If you write or revise policy, put these in it

• Detector output is never standalone evidence, never triggers an automatic consequence, and never grounds a finding. State this in the policy itself, not in a memo.(Why AI Detection Fails for Academic Integrity)
• Students must be able to see the accusation and respond to it, with the right to silence preserved and an oral verification route available for any allegation that rests on style alone.(Heads We Win, Tails You Lose: AI Detectors in Education)
• Procurement terms should cover data retention, use for model training, storage location, breach notification, and appeal rights — the risk sits with your institution, not the vendor.(Heads We Win, Tails You Lose: AI Detectors in Education)
• Fund the alternative. The documented faculty complaint is "suspicion without recourse," so the budget line that matters is verification capacity and guidance, not a detection license.(The Best Response to Student AI Use Is Not Detection, It Is Dialog)
• Ask for error rates in writing, then try to validate them locally. If the advertised rate cannot be reproduced on your own submissions — as Vanderbilt found — that is your answer.(Detecting the Undetectable? Reassessing Academic Misconduct Procedures in the Era of Generative AI)

7. The legal risk when a student is wrongly accused

A detector score that becomes an accusation is where this stops being a teaching question. Institutions are not exposed because someone was accused, but because of how the accusation was built and handled, and the exposure usually surfaces first as an internal appeal or a regulator complaint rather than a lawsuit filed in court.

• The procedure is the first thing tested. Principles of natural justice require that a student be informed of the allegation and given an opportunity to respond before any determination is made, obligations codified in regulatory standards as well as in sound academic integrity policy; the response opportunity is typically an investigative meeting or panel interview, and whatever the student says becomes part of the evidentiary record. A finding reached without that step is vulnerable regardless of whether the underlying suspicion was reasonable.(How strong is the evidence in generative AI-related academic misconduct allegations? A mixed-methods analysis)
• The evidence standard is the second. Misconduct findings require evidence meeting the balance of probabilities, and detector output does not get there on its own: the scores cannot be validated against ground truth in real submissions, the tools cannot be interrogated about how a given verdict was reached, and a student cannot cross-examine a number. If your case cannot be stated without the detector score, the case is weak on its face.(Heads We Win, Tails You Lose: AI Detectors in Education)
• Patterned error turns a technical defect into a fairness and equality problem. Detector error is not random: the strongest evidence in the knowledge base shows flagging concentrated on non-native English writing and on one discipline's prose conventions over another's, with measured accuracy of 0.69 and 0.61 for two widely used commercial tools and both of them failing on hybrid human-AI text. A finding built on an instrument that misclassifies by language background is a finding that invites a discrimination argument.(Evaluating the accuracy and reliability of AI content detectors in academic contexts)(Who wrote this? Evaluating the reliability of AI detection tools in higher education)
• Blanket "AI use" bans can remove an accommodation. Rules that do not separate transcription and OCR from generative drafting may criminalise the assistive tools students with conditions affecting motor control, handwriting legibility or typing accuracy rely on, several of which have been discontinued with AI transcription filling the gap. Over-inclusive policy is a legal exposure, not just an imprecise one.(Transcription is not generation: Distinguishing non-generative AI tool use from academic misconduct in higher education assessment)
• The data is your liability. Detectors store student work on third-party servers, sometimes overseas under weaker Privacy protections, which puts breach, retention and onward commercial use inside your institution's risk register rather than the vendor's marketing.(Heads We Win, Tails You Lose: AI Detectors in Education)
• Vendor claims will not protect you. A licensed detector advertising a 1% false-positive rate failed validation when the university tried to reproduce it, implying roughly 750 mislabelled students among 75,000 annual submissions; that institution disabled the tool. The claim in the contract does not transfer the risk in the hearing.(Detecting the Undetectable? Reassessing Academic Misconduct Procedures in the Era of Generative AI)

What lowers the risk is procedural rather than technical: never treat detector output as standalone evidence or an automatic trigger; document the evidence standard your process applies; give notice and a genuine opportunity to respond on the record; offer an oral verification route when the case rests on style alone; put retention, training-use and breach terms in procurement; write policy scope so that assistive and transcription tools are explicitly addressed; and keep the audit trail that shows all of it happened. The knowledge base's account of the legal exposure itself is in Legal Issues and Risks, and it is honest about its limits: it documents procedures, evidence categories and instrument reliability, not litigated outcomes.

How this page differs from the neighboring FAQs

• How Can I Reduce AI Cheating in My Course? is the instructor's design playbook: guardrailed tools, assessment redesign, verification, declarations, AI literacy. This page answers the narrower prior question of whether detector output can be used at all.
• How Do I Redesign Assessment So That a Grade Still Tells Me Something Defensible About What the Student Knows or Can Do? covers assessment redesign in depth; this page only points to the redesign moves that specifically replace detection.
• How Do I Write a Course AI Policy and Communicate It to Students? covers writing and communicating a course policy; section 6 here covers the detection-specific clauses an institutional policy needs.
• Legal Issues and Risks is the concept page behind section 7, covering wrongful accusation, surveillance, accessibility and data-protection exposure together.

The bottom line

Do not build a finding on a detector score, and do not buy one expecting it to secure your assessments. The measured behavior of these tools is the opposite of their marketing: they catch honest, disclosed, lightly edited work and competent second-language writing, while deliberate evasion passes 96% of the time and wholly generated submissions passed a live exam system 94% of the time.(Why AI Detection Fails for Academic Integrity)(Detecting the Undetectable? Reassessing Academic Misconduct Procedures in the Era of Generative AI) The integrity question you can actually answer is whether a student can demonstrate understanding of what they submitted — and that is a teaching capacity worth funding.