Two stories from this spring tell the same story from opposite ends. A New York court reversed a university's AI-cheating finding and ordered the record expunged - AI-detection software failed here. A Stanford senior published, in the country's paper of record, an obituary for what a degree means. Both point at the same root cause — and at the same answer.
What a Stanford graduate is telling us
In his May 17, 2026 New York Times essay What A.I. Did to My College Class, Theo Baker — a graduating Stanford senior and Polk Award-winning journalist — documents what most people in higher education already know. Cheating is omnipresent. He cannot name a classmate who has not used AI on some assignment. He describes students signing "I did not use ChatGPT" declarations with ChatGPT open in the next window. A campus survey found that 49 percent of computer science majors would rather cheat on an exam than fail. Proctored handwritten exams — banned at Stanford for a century — returned in April 2026 as the institution's only remaining answer.
The essay closes with five words that read like an obituary: "For us, this was college."
What Baker does not offer — and this is not a criticism; he is a journalist, not a vendor — is a path forward. He diagnoses. He does not prescribe. We would like to.
What the Adelphi ruling is telling us
On January 28, 2026, a New York State Supreme Court justice issued a decision in Matter of Newby v. Adelphi University that should be required reading in every general counsel's office. Orion Newby, a first-year student with documented learning differences, was found responsible for academic dishonesty on the strength of a single Turnitin output flagging his paper as 100% AI-generated. Two other detectors classified it as human-written. Adelphi refused to consider that evidence. The same administrator who issued the determination heard the appeal. The court vacated the finding.
The operational headline: detector output alone is no longer a defensible basis for sanction. Similar cases are now pending against Yale, the University of Minnesota, and the University of Michigan. The FTC's August 2025 final order against detector company Workado — marketed as 98% accurate, measured at 53% — is a parallel signal from the consumer-protection side.
The common root - where AI-detection fails
The model universities have used for academic integrity for a quarter-century is structurally simple: a student submits an artifact, a tool produces a probability score, the score becomes the case. This worked when the only way to get cheap text was to copy it. Generative AI removes the referent. The text is original. It is also, often, not the student's.
What replaces detection is not better detection. It is a different question. Instead of asking, after the fact, is this artifact suspicious?, the question becomes: do we have evidence, beyond the artifact itself, that this person actually did this work?
Introducing the corroboration standard
We propose a name for what comes next: the corroboration standard. An academic-integrity finding cannot rest on the output of any single tool. It must be supported by independent evidence from at least one of three categories:
- Process evidence — a verifiable record of how the work was created, generated with the student's knowledge and consent.
- Conversational evidence — an oral defense or structured interview demonstrating substantive familiarity with the work.
- Contextual evidence — sources, notes, drafts, or related coursework corroborating authorship.
Detector output, where used, becomes one input among several. It can flag a paper for examination. It cannot, alone, support a sanction.
Three consequences follow. The legal exposure that Newby exposed evaporates. The credibility problem Baker describes begins to reverse — credentials carry information again because the standard of evidence is higher, not because surveillance is greater. And critically, the corroboration standard does not require universities to take a position on whether AI may be used. A student who uses AI heavily, transparently, and as a documented collaborator passes it. A student who pretends to have written something alone, when in fact they did not, cannot.
What to do Monday morning
- Audit your current policy for Adelphi-vulnerability. Three failure modes to look for: detector output as primary evidence, appeals heard by the original decision-maker, vague procedural rights. Any two together is the Newby fact pattern.
- Stop treating detector output as a verdict. One paragraph of standing guidance: detection scores trigger inquiry, not findings.
- Adopt the corroboration standard formally. Update the integrity policy to require that no sanction rest on detection technology alone, and to define the three evidence categories.
- Add a contestation clause to syllabi. Give students a written, defined path to demonstrate authorship — drafts, version history, oral defense, or process evidence — and commit to honoring it.
- Run a pilot. One course, one term, voluntary process-evidence reporting. Measure what happens. Institutions doing this are reporting something policy documents miss: students like being able to show their work.
- Communicate publicly. An institution that can say clearly — to applicants, alumni, and the press — "we no longer rely on AI detectors as sole evidence, and our students can voluntarily demonstrate authorship using a verifiable record" has stopped losing the narrative.
An invitation, not a sermon
Theo Baker's essay closes with "For us, this was college." It is a true sentence and a sad one, and it deserves a response. Ours is this: the system Baker watched fail is not the only one available. The students walking onto campus this fall do not have to inherit the resignation that defined his class. There is another way. It exists. It works. It is not a future we are waiting for — it is a Monday-morning decision.
We would like to invite you to make it.
See the corroboration standard in practice
Mentafy's Authorship Report is the process-evidence layer that makes the corroboration standard operational — for one course, one department, or your whole institution. We are running pilots now. See how it works → or talk to us about a pilot.
No comment yet, add your voice below!