Uncross-Examined: When Detection Software Becomes the Star Witness

As of November 2026, the most consequential witness in many American academic-integrity proceedings is not a professor, a proctor, or a peer. It is a dashboard. A student submits an essay; a detection tool returns a percentage; an administrator opens a case. What follows is supposed to be a disciplinary hearing, but it often resembles a reverse burden of proof: the institution presents a score, and the student must explain why the machine is wrong.

That inversion is now routine enough to deserve scrutiny. Universities have spent the past several years buying AI-detection software, writing it into course policies, and routing accusations through offices that were built for plagiarism, not probabilistic inference. The vendors publish confident marketing and cautious footnotes. The campuses publish honor codes. Almost no one publishes the audit that would tell a student, a faculty member, or a regulator how often the tool is wrong in a given discipline, at a given length, for a given population of writers.

The Vendor's Confidence, the Student's Burden

The appeal of detection software is administrative. It promises scale. A writing center can hire more tutors; a dean of students cannot hire a machine-learning expert for every suspected violation. Vendors sell dashboards that flag passages, assign confidence levels, and generate reports that look like evidence. Turnitin, GPTZero, Copyleaks, and Originality.ai have all marketed AI-writing detection to colleges, sometimes with accuracy claims and sometimes with caveats tucked into documentation. The caveats matter: detection is not proof, and a probability is not a finding.

Yet campus procedures often treat the report as the starting point of a case rather than one data point within it. Faculty members receive little training in base rates, false positives, or the way detection tools can penalize writers who use formulaic prose, translation software, or the kind of cautious, templated syntax common in first-year composition. International students and students with disabilities are not hypothetical edge cases. They are the population most likely to be misread by a tool calibrated on an unrepresentative corpus.

Three Hearings, One Missing Record

Three publicly reported episodes show the pattern. In 2023, a Texas A&M University professor told a class he suspected widespread AI use and threatened to fail students en masse; the university later walked back the penalties after the episode drew national attention. The case became a lesson in what happens when suspicion outruns process. But it also exposed a quieter problem: the institution had no shared protocol for distinguishing a professor's hunch from a vendor's score from an actual violation.

A second case, reported in 2024 at a large public university in the Midwest, followed a familiar arc. A student received a detection report showing a high probability of AI generation. The student requested the underlying report and the model's version information. The campus provided a screenshot. The student appealed, arguing that the assignment had been drafted in a campus writing center and revised across three drafts. The appeal succeeded, but only after the student's family paid for a consultation with an academic-integrity attorney. The record did not establish that the tool was wrong. It established that the student could afford to make the institution prove its case.

A third episode, at a private Northeastern college in 2024, ended with a dean vacating a suspension after the student's adviser noticed that the detection report cited a passage the student had written before the assignment began. The vendor's confidence score had not changed. The context had. The college did not revise its contract, its training, or its appeals form. It revised the individual outcome.

Contracts Without Contours

The procurement records are where the due-process gap becomes concrete. Many detection contracts are shielded by confidentiality clauses, or they are folded into larger learning-management subscriptions that few faculty ever see. Campuses rarely disclose which version of a model they use, how often it is recalibrated, whether it has been tested on multilingual writers, or what the vendor's false-positive rate is for the document lengths most common in their courses. A university that would not buy a fire alarm without inspection standards will buy a disciplinary tool with a sales deck.

This is not an argument that every vendor is reckless or every administrator is careless. It is an argument that the ordinary safeguards of institutional evidence have not caught up with the technology. In a courtroom, an expert witness can be cross-examined. On campus, the algorithm is often treated as a neutral clerk. The student sees a number. The panel sees a report. The vendor sees a license renewal.

An Appeal Is Not a Defense

Due process in higher education has never been as robust as a criminal trial, and it should not be. But the minimum obligations are not mysterious: notice, an opportunity to respond, an impartial decision-maker, and a record sufficient for review. AI-detection cases strain each of those commitments. Notice may consist of a generic honor-code citation. The opportunity to respond may arrive after a grade has been withheld. The decision-maker may be the same office that purchased the software. The record may omit the model version, the prompt, the rubric, and the comparison texts needed to evaluate the claim.

Appeals make the asymmetry worse. A student who can hire a lawyer, a consultant, or a sympathetic professor can force the institution to explain itself. A student who cannot may plead with a dean who has forty other cases and no technical training. The result is not justice; it is triage.

What Auditable Integrity Would Look Like

A defensible system would begin with disclosure. Campuses should publish which detection tools they use, how those tools are validated, and what error rates have been observed in their own populations. They should require human review before any accusation leaves a classroom. They should give accused students the full report, the model version, and the assignment context, and they should place the burden of persuasion on the institution, not on the accused. They should audit outcomes by discipline, language background, and disability status. And they should treat an unverifiable score as a reason to investigate, not a reason to convict.

None of this requires abandoning AI detection. It requires refusing to let a vendor's probability become a campus verdict. The question before universities in 2026 is not whether machines can help identify misconduct. It is whether institutions will build disciplinary systems that can survive being audited. Right now, too many cannot.