Is AI Detection Reliable Enough to Base a Misconduct Case On? (2026)

Short answer: not on its own. A detector returns a probability about a text, not evidence about a person’s conduct. At institutional volume even a low false-positive rate produces a substantial absolute number of wrongly flagged submissions. Detector output can reasonably start an inquiry; it cannot carry the institution’s burden of proof.

That is a narrower claim than either side of this debate usually makes, and it is the one an integrity office can actually operate. What follows is the reasoning, the arithmetic, and what a defensible process looks like when the score is treated as a signal rather than a finding.

What does a detector actually produce?

It is worth being precise, because the difference from similarity matching is the whole argument.

The right-hand column is the difficulty. With a similarity report, a panel can look at the matched source and form its own judgement about whether the overlap is quotation, convention, or copying. With a detection score there is nothing to look at. The panel is asked to accept a classifier’s inference, and the only way to interrogate it is to understand how it works.

What did the vendor claim, and who tested it?

The most useful public record of an institution actually evaluating this comes from Vanderbilt University, which announced on 16 August 2023 that it was disabling Turnitin’s AI detection tool “for the foreseeable future,” after several months of using and testing the tool, meeting with Turnitin and other AI leaders, and speaking to other universities with access.

Vanderbilt’s stated reasons are worth reading as a checklist, because each is a question any institution can ask its own vendor:

• The feature was enabled for customers with less than 24 hours’ advance notice, with no option at the time to disable it.
• The institution had no insight into how it works. Vanderbilt noted that Turnitin had said only that the tool looks for patterns common in AI writing, without explaining or defining those patterns.
• At launch, Turnitin claimed a 1% false positive rate.
• Research indicated detectors are more likely to flag text written by non-native English speakers.
• Sending student work to a third-party detector raises its own privacy questions about data handling by another company.

Vanderbilt’s conclusion was unambiguous: “we do not believe that AI detection software is an effective tool that should be used.”

One caveat we will state rather than bury: that assessment is from August 2023. Products have changed since. We attempted to retrieve Turnitin’s current AI-detection documentation directly and could not — the pages returned an automated challenge rather than content — so we make no claim about the tool’s present accuracy. What we can verify is what the vendor’s own public product pages say today, and that is discussed below.

Why does a 1% false positive rate matter so much?

Because integrity offices operate at volume, and percentages are the wrong unit for a caseload.

Vanderbilt did the arithmetic publicly: the university submitted 75,000 papers to Turnitin in 2022. Applying a 1% false positive rate to that volume gives around 750 student papers that could have been incorrectly labelled as containing AI-written material in a single year.

Sit with the operational consequence rather than the statistic. Seven hundred and fifty inquiries is not a rounding error; it is a workload that would swamp most integrity offices, and every one of those cases involves a student who did nothing wrong being asked to prove it. Now apply your own submission volume:

1. Take your annual submissions through the assessment platform.
2. Multiply by the vendor’s stated false-positive rate.
3. Ask whether your office could absorb that many inquiries — and whether your appeals process could.

The second consideration is subtler and more important. The rate applies to submissions, but the harm lands on students, and it does not land evenly. If a detector systematically over-flags a particular group — non-native English speakers being the documented case — then the aggregate rate conceals a much higher rate for that cohort. An institution with an equality duty cannot treat that as an acceptable error term.

Has the market itself moved on?

This is the part of the story that has changed most since 2023, and it is visible on vendors’ own pages rather than in commentary.

Turnitin’s current product overview leads with Turnitin Clarity, presented with the line “Don’t just detect AI. Understand it,” and described as bringing writing analytics and AI insights together for “the visibility needed to guide responsible AI use and empower authentic learning journeys.” The rest of the line-up is Feedback Studio, Gradescope, ExamSoft, Similarity and iThenticate.

Read that as a market signal. The framing has shifted from adjudicating whether a finished document was machine-written toward giving institutions visibility into how a document was produced. That is a meaningfully different product category, and it is a better fit for what integrity offices actually need — because process evidence is inspectable in a way a probability score is not.

One further change worth recording for anyone maintaining a vendor list: Ouriginal’s service ended on 30 June 2026, per Turnitin’s own page, which now directs former users toward Turnitin Similarity. Comparison content naming Ouriginal as a current option is out of date.

What does a defensible process look like?

The question an integrity office should answer is not “did AI write this,” which may be unanswerable, but “can this student account for this work.” That reframing is legally and pedagogically sturdier, and it does not depend on any vendor.

1. Treat any detector output as a trigger, never a finding. Write that into the policy in those words, so a panel cannot treat a score as dispositive.
2. Require corroboration before an allegation. Vanderbilt’s own guidance to instructors is a reasonable starting list: compare the work to the student’s previous writing for style, tone and level; look for inaccuracies in sources, arguments and facts, since generative tools may fabricate sources entirely; and talk to the student.
3. Make the conversation the primary instrument. A short viva about the submitted work — why this method, where this source came from, what this paragraph means — distinguishes authorship far more reliably than a classifier, and it produces a record a panel can review.
4. Never disclose a score as the accusation. “Our software says 82%” invites a fight about the software. “Please talk us through how you produced section 3” invites an account.
5. Record the burden explicitly. The institution alleges; the institution proves. A policy that quietly shifts the burden onto a student to disprove a machine output will not survive scrutiny on appeal.

What about designing the problem out?

The durable answer is assessment design, and it is the one every serious source lands on. Vanderbilt recommended reformatting assessment — in-class writing, requiring students to write about specific material discussed in class, focusing on current issues — alongside clear expectations and citation of AI use where it is permitted.

That is not a counsel of despair about technology; it is a recognition that authorship is easiest to evidence when the process is visible. Assessments that generate intermediate artefacts — proposals, annotated bibliographies, drafts, supervision records — give both students and panels something concrete. An assessment whose only artefact is a finished document at a deadline will always be the hardest case.

If you are weighing how to give an institution that process visibility without adding a detection contract, we are happy to talk through what an evaluation would involve at your institution. Request an institutional evaluation.

Frequently asked questions

Can universities detect AI-generated essays?

Tools exist that estimate the likelihood text was machine-generated, but they produce probabilities rather than proof, and their accuracy claims are vendor claims that institutions should test against their own volume and cohort.

Can a detector score alone support a misconduct finding?

It should not. A score is not evidence of intent and provides no inspectable source. Use it to open an inquiry, and corroborate before alleging.

What false positive rate did Turnitin claim?

At launch of its AI detection tool, Turnitin claimed a 1% false positive rate, as recorded by Vanderbilt University in August 2023.

Why did Vanderbilt disable Turnitin’s AI detector?

It cited the false-positive risk at its submission volume, the absence of any explanation of how the tool works, evidence of bias against non-native English speakers, and privacy concerns about third-party detection.

How many false positives would we see?

Multiply your annual submissions by the vendor’s stated rate. At Vanderbilt’s 75,000 papers, a 1% rate implies roughly 750 papers a year.

Are detectors biased against international students?

Vanderbilt cited research finding detectors more likely to label non-native English speakers’ text as AI-written. Any institution with a large international cohort should treat this as a live equality issue rather than a technical footnote.

Is AI detection the same as plagiarism detection?

No. Similarity matching shows you a source document you can inspect; AI detection infers from the text’s own properties and produces no source.

Is Ouriginal still available?

No. Turnitin’s own page states that the Ouriginal service ended on 30 June 2026 and points former users to Turnitin Similarity.

Should we tell students whether we use detection?

Yes. Transparency about what is checked and how findings are used is a basic procedural fairness expectation, and undisclosed screening is difficult to defend on appeal.

What should replace detection in our policy?

A clear permitted-use and disclosure rule, corroboration requirements before any allegation, and assessment design that produces inspectable process evidence — the components we set out in what a university AI policy should include.

How should we pilot an alternative before committing?

Run a bounded departmental pilot with success criteria agreed in advance rather than an institution-wide rollout — the approach described in our guide to running a departmental pilot.