We appear to have spent the best part of the last decade teaching machines (and now AI) to catch cheats. This summer, in the same breath, the sector began switching those machines off. That is not a contradiction. It is a confession…

1. The great switch-off

Let me start with the fact that should really make us all think within the test security ‘sector’.

On the 5th August 2026, Inside Higher Ed reported that a growing list of universities now forbid their own faculty from treating AI detectors as proof of anything. Yale. Vanderbilt. Johns Hopkins. Indiana. Each has moved to ban or discourage reliance on detection software as the sole evidence in an integrity case. At least a dozen more, Northwestern, Georgetown and New York University among them, have gone further and simply disabled Turnitin’s AI detection outright. The reason given is the one the vendors spent two years denying. The false positive rate is (allegedly!) higher than advertised.

Read that again, because the timing is the story. These are not institutions going soft on cheating. Nobody thinks the problem has receded. They are switching off the traditional and widely-used ‘alarm’ not because the building is safe, but because the alarm keeps screaming at empty rooms. An innocent student flagged as a cheat is not a rounding error. It is an accusation. And a tool that manufactures accusations at scale is not a safeguard. It is a liability wearing a detective’s lanyard.

So here is the uncomfortable position the higher education institutions now occupy. Cheating tools have never been more accessible. Detection on the other hand, has never been less trusted. The gap between those two sentences is where this week’s edition dives in.

2. The invisible ink

If the detectors are now out, what rushes in to fill the vacuum? Ingenuity. Some of it magnificent, some of it faintly desperate.

Meet the ‘trap professors’. At Brown University, the economist Roberto Serrano set a take-home midterm and buried an instruction inside the paper in white font, invisible to a human eye, plain as day to any large language model asked to read the question. At Alcorn State University, Jason Gibson did much the same, and his ‘trap’ went viral on TikTok just this last week. – Link here, which works for those with and without TikTok: https://www.tiktok.com/@iamjasongibson/video/7664639999479958797

The technique is simple and elegant. Paste the question into a chatbot and the hidden instruction hitches a ride, surfacing in the answer like a name written in lemon juice held to a flame. Very 007…

The results were not subtle either. Serrano’s cohort returned an average of ninety six, with forty students scoring full marks, against a historic range that sat somewhere between sixty five and eighty. You do not need a standard deviation to feel something is wrong. Something broke.

I admire the craft, given my wife thinks I’m a spy (or ex-spy). I also want to be honest about what the craft admits. A hidden trap in an exam is not a security control. It is nothing more than a sting operation. And a sting operation is what you reach for once you have already conceded that the assessment itself cannot tell an honest answer from a dishonest one. This is the Mirror Moment (from previous blogs) in its purest form. The professor looks at the exam, sees his own flawed design staring back, and rather than change the design he lays a snare for the student who exploits its weakness. Clever, yes. But the cleverness is aimed at the symptom. The take-home essay, as an instrument for certifying that this person knows this thing, is the patient on the table. Nobody in these stories is treating the patient.

p.s. As you know, I work for risr/, so I’m allowed my medical analogies… 🙂

3. A clean session proves nothing

Now lift your eyes from the seminar room to the exam hall, because this is where my readers live, and where the stakes stop being a grade and start being a licence to practise.

The university story and the exam/certification story are the same story told twice at different volumes. Talview’s AI Threat Index, which anchors this month’s In-Center Edition of the Exam Security Summit, puts the matter in language I’ve previously used as part of the board of the AI Dome.

A single, clean proctored session is no longer proof of a secure credential. Sit with that. The camera saw nothing. The lockdown browser held. The proctor signed off. And none of it tells you whether the person who now holds the qualification can do the thing the qualification asserts.

The report in question documents exam completions clocked at five hundred milliseconds, and at nine minutes. Those are not fast students. Those are temporal signatures no human hand can produce. They are the fingerprints of an agent that logged in, read the paper, answered it and left, while the proctoring stack watched a still webcam and saw a candidate behaving beautifully. The session was immaculate. The credential it produced was hollow.

Here is the line I would underline and have amplified at various conferences this year already. If a machine can reliably pass your exam, your exam may no longer measure competence. That is not a security failure you can proctor your way out of. It is a validity failure. And validity is the one word in our field that cannot be outsourced to a vendor, because the moment your credential/assessment stops predicting real capability, the credential is worth precisely nothing, however secure the room in which it was earned.

The Measured Response

Let me pull these threads into one.

The detectors are being unplugged because they accuse the innocent. The ‘trap professors’ are trending because they have given up on the instrument and gone hunting instead. The credentialing/awarding bodies are staring at clean sessions that certify nothing. Three stories, one shape. Every one of them is a failure to answer a single question. Who actually produced this work?

That is the Attribution Gap, and it has been the quiet engine under all of this from the beginning. We built a decade of tooling to answer the wrong question. Not who wrote this, but does this look like it was written by a machine. Those are not the same question, and the fast growing false positive scandal is the bill arriving for pretending they were. A detector guesses at provenance from style. Style is exactly what a machine can now counterfeit and a nervous human (or one with English as a Second Language) can accidentally imitate. Build your enforcement on that quicksand and you will punish the anxious and wave through the automated.

The way through is not a better detector. There is no better detector coming anytime soon that will not carry the same flaw baked into its premise. The way through is to move attribution back to where it can actually be established. Design that shows the working. Assessment that captures the process and not only the artefact. And judgement, human judgement, an Expert in the Loop who is asked to weigh evidence rather than rubber stamp a probability score handed down by a black box. The ‘Guardians of Assessment’ do not win this by buying a sharper adversary detector. They win it by refusing to certify what they cannot attribute.

Switching off what might be perceived as a broken tool is not surrender. It is the first honest act in the room. What matters is what you reach for next. Reach for a trap, and you have admitted the design is beyond saving. Reach for better assessment design, or “Re-Innovation” (as described in Measured Response 2) and you might just close the gap.

That’s the Measured Response.

Over to you…

If your detector is off, your proctor sees nothing and your session is clean, what exactly is your evidence that the candidate did the work? I’d love to hear your thoughts!

Until next week…

Paul

Sources:

  1. Inside Higher Ed – 8th July 2026 https://www.insidehighered.com/news/faculty/learning-assessment/2026/07/08/brown-professor-suspects-most-his-class-used-ai-cheat
  2. Inside Higher Ed – 5th August 2026 https://www.insidehighered.com/news/tech-innovation/artificial-intelligence/2026/08/05/ai-detectors-are-out-new-approaches-are
  3. TikTok – Professor Jason Gibson https://www.tiktok.com/@iamjasongibson/video/7664639999479958797

Subscribe for the latest email updates