Harvard's Dean Says Scrap AI Detection. The Evidence Says Scale It Back Instead.

Quick answer

Harvard College Dean David Deming told students and faculty in a September 2026 email that the college should get "out of the AI-detection business" entirely, calling it a losing race that damages trust between students and faculty. A week-old AFP wire report, picked up nationally on October 6, shows what that actually looks like in practice: professors dropping take-home writing assignments altogether rather than finding a middle path. But most of the schools that have already acted on AI-detection skepticism this year didn't scrap it. They narrowed what a score is allowed to do. That distinction is the whole story.

What Deming actually told Harvard

On the first day of fall classes, Deming sent undergraduates a message naming "AI use policies" as a priority for the year. In it, he floated "AI acceptance, or even encouragement" for writing-based courses, paired with a shift toward in-class, proctored assessments for anything AI could easily do at home. The line that traveled furthest: "We would all benefit from getting out of the AI-detection business because it damages trust between faculty and students," as reported by The Harvard Crimson. On his Substack less than ten minutes later, he went further, describing AI writing and AI detection as "trapped like Alice and the Red Queen in a race that will leave us running ever faster just to keep up."

It's worth being precise about what Deming proposed, because "stop detecting AI" and "stop treating a detector score as a verdict" are not the same policy, even though headlines tend to compress them into one. Deming's email argued for abandoning detection as an enforcement tool in writing courses specifically, replacing it with assessment redesign: fewer take-home essays, more supervised exams and project work.

The reaction inside Harvard wasn't unanimous

The Crimson's own follow-up reporting found humanities and social-science faculty split, not aligned. English professor Deirdre Lynch said she'd keep a blanket AI ban in her two fall courses regardless of Deming's suggestion, arguing that "writing is sort of a personality and individual thoughts clashing with something shared called language. You figure out what your style is through that clash." History professor Peter Gordon warned that encouraging AI in writing classes would push students to "outsource their thinking" and lean on "conventionalized experiential reports" instead of forming their own responses to texts. A Crimson analysis of roughly 600 fall-2026 course syllabi found AI use was permitted in just 27% of Arts and Humanities courses and 35% of Social Sciences courses, versus more than half in science and engineering — a department-by-department split Deming's one-size-fits-all framing didn't fully account for.

Context mattered too. Deming's email landed weeks after Harvard folded its standalone Writing Center into a broader support center, a timing critics didn't let pass quietly. Writing in The Atlantic, Tyler Austin Harper argued the sequence raised a real question about how seriously the institution still takes unassisted writing as a skill worth protecting.

This isn't just a Harvard story anymore

An AFP wire report published October 6 shows the same logic spreading well past one dean's email, and it's less tidy in practice than "stop detecting, start trusting" sounds on paper. University of Wisconsin-Madison biology professor Timothy Paustian built a hidden-prompt trap that caught 60 of 350 students submitting AI-written work in a single class — then told AFP he's "pretty much thrown in the towel on classic writing assignments for lower-level classes" because newer models increasingly ignore or strip out exactly that kind of trap. Cornell University's own guidance, quoted in the same piece, states plainly: "It can be very difficult to accurately detect AI-generated content." A national AACU/Elon survey found 95% of faculty fear students are overrelying on AI, and misinformation researcher Timothy Caulfield called the current moment "a bit of a technology arms race... the detectors versus the evaders," getting harder on both sides at once.

Paustian's own case is the clearest illustration of why "detection doesn't work, so drop it" undersells what detection-with-limits actually buys a school. In a 2024 peer-reviewed study he co-authored before this fall's trap-and-catch exercise, detectors correctly separated AI from human writing 88% of the time across 153 student essays, per the published results in Frontiers in Education. The method, not the underlying score, is what broke down two years later — newer models evade the specific trap he built, and a single score was never designed to carry a disciplinary case on its own.

The schools that already acted chose the narrower version

That's the part the "scrap it" framing skips. We've covered the earlier wave of this story: Yale, Vanderbilt, Johns Hopkins and Indiana restricted AI-detection scores from being used as sole evidence of misconduct. Northwestern, Georgetown and NYU went further and disabled Turnitin's AI-detection feature specifically. None of those schools, as of this year, told faculty to stop caring whether a student used AI. They told faculty a percentage alone can't decide a case — which is a narrower, more defensible claim than "the whole category of detection is a waste of time."

That narrower claim also happens to match what a recent Harvard-adjacent research effort found about the cost of getting this wrong in either direction. The Center for Digital Thriving's October report on "two-way suspicion" describes a different failure mode entirely from Deming's "stop detecting" proposal: teachers developing an unprovable hunch ("AI-dar") and students hiding or disguising their process defensively, regardless of whether a formal detector is even in the room. Abandoning detection doesn't make that suspicion disappear. It just removes the one tool that could, used correctly, give both sides something more concrete than a hunch to talk about.

What this actually means if you're deciding your own policy

The honest version of this debate isn't "detectors work" versus "detectors don't." It's "what decision is a score allowed to make on its own." A detection score that flags a passage for a human conversation is doing its job. A detection score that ends a case by itself is being asked to do something no single number, from any vendor, was built to do — that's the same reason Turnitin's and GPTZero's own published guidance tells schools not to use a score as sole evidence. Dropping detection to avoid misuse just trades one failure mode (over-trusting a score) for another (losing the earliest, cheapest signal that a conversation needs to happen at all).

FAQ

Did Harvard officially ban AI detection? No. Deming's September email described his own preference and said he would spend the year developing advisory guidelines through "extensive consultation" with faculty and students — a college spokesperson confirmed the email was a prompt to continue discussion, not a final policy.

Is Harvard the only school moving this direction? No. The AFP report that circulated nationally on October 6, 2026 quotes Cornell's own guidance questioning detection accuracy and describes Yale among schools that have limited detector reliance, alongside Brown and Alcorn State faculty who went viral this year after catching apparent mass AI use with hidden-prompt traps.

What's the difference between "restricting" and "abandoning" AI detection? Restricting means a score can flag a case for human review but can't be the only evidence used to discipline a student — the policy most schools that acted this year actually adopted. Abandoning means dropping detection as an input entirely, which is closer to what Deming proposed for writing courses specifically.

See what a flagged passage actually looks like before a policy debate turns into your case

Whichever way your institution lands on this, the question that actually matters isn't "detector yes or no" — it's "what evidence does a flag need before anyone acts on it." Run a passage through TheChecker.AI's free demo to see the kind of signal a score gives you on its own, and what it doesn't tell you without more context.