Universities Give Up on Catching AI-Written Work

Vanderbilt, Yale, Johns Hopkins, Northwestern, the University of Waterloo, the University of Cape Town, and Curtin University have all restricted or disabled AI detection tools like Turnitin, GPTZero, and Copyleaks, according to the Financial Times. The reason isn't that AI use in coursework has slowed down. It's that the tools built to catch it can't be trusted to get the call right.

Key Points

• A New York court ruled in favor of Adelphi University student Orion Newby after the school upheld a misconduct finding based on a Turnitin "100 per cent AI-generated" score, despite other detection tools showing zero percent
• A survey of over 6,600 students across seven UK universities found 32% admitted some level of unpermitted AI use in assessments
• A widely cited 2023 Stanford study found detection tools disproportionately flag non-native English speakers as false positives
• More than 40% of 163 audited UK universities have no publicly accessible AI policy, according to Edinburgh Napier University research
• 42% of students surveyed by the UK's Higher Education Policy Institute said fear of false accusation makes them less likely to use AI at all, even when permitted

The Tools Were Never Built to Be Courtroom Evidence

Turnitin's own chief product officer, Annie Chechitelli, describes the detector as "a starting point" and "a data point," not proof. That distinction got lost in practice. Edward Watson of the American Association of Colleges and Universities put the concern plainly: faculty shouldn't treat AI detection as "hard evidence" or "smoking gun proof," yet that's functionally how it's been used in disciplinary cases until courts started pushing back.

Detection Was Always a Stopgap, Not a Fix

Judy Williams, pro vice-chancellor at Queen's University Belfast, reframes the actual problem: "AI detection tools are not the solution... If we want confidence in academic integrity, the answer is good assessment design." That's a shift from policing outputs to redesigning what's being assessed in the first place, oral components, process documentation, iterative drafts. It's a slower fix, but it doesn't depend on a probability score holding up as fact.

The Same Reliability Problem Is Coming for Marketing Teams

This isn't only a higher-education story. Platforms are rolling out AI-content labeling and detection features of their own, and the underlying technology has the same accuracy ceiling universities just ran into. Any brand or agency using detection tools to screen vendor-submitted content, verify originality, or police AI-use policies internally is relying on the same probabilistic guesswork that just got a New York court's attention. Teams building AI-use policy into their own growth strategy should treat detection scores the way Turnitin's own product chief does, as a data point, not a verdict.

Policy, Not Detection, Is Where This Gets Solved

Urszula Lis of the European Students' Union raises the equity problem directly: two people using AI the same way can get different outcomes depending on which institution, or which detector, is judging them. That inconsistency is the real cost of skipping the harder work of writing clear policy. Illingworth's research put it bluntly: many current policies "promise critical thinking but deliver audit trails." Organizations building their own AI-use guidelines, including how content gets vetted, should take that as a warning before scaling a detection-first approach. If you're working through what AI-use policy should actually look like for a content team, that's the kind of groundwork our AI marketing services work covers.