Skip to content

Did they write it, or did the AI?

This piece, and the series it opens, is about business schools and executive education, and about the applied work inside their programmes: the MBA capstone, the master’s thesis, the in-company project a custom programme runs for a client, the assignment a participant takes back to the job. Every one of them ends in a document. For a century the document was the proof: if it was good, the person had learned. That proof broke in about two years, and the ways schools are trying to repair it tell you what the real problem is.

Two failures, one cause

In February 2025 an executive MBA student sued Yale School of Management over a one-year suspension for alleged AI use on an exam. A teaching assistant had flagged answers as long and elaborate with near-perfect grammar, and a professor ran them through GPTZero. The student’s response, reported by Poets&Quants, was to feed papers by Yale’s former president and the school’s own dean through the same detector, which rated them 100% AI-written.

A year earlier, King’s Business School had asked students to declare any AI use in assessed work, with no penalty attached. 74% of the students who had used it did not declare it, according to a May 2026 account in The European Business Review, which also observes that in most assessment only the final document matters, and that the classic MBA case analysis maps almost perfectly onto what a language model does well.

One is a false positive that ended in court. The other is silent use at scale. They look like opposite problems and they have the same cause: nobody could see how the work was made. The detector was asked to reconstruct that from the finished text, and it cannot.

Detectors are being switched off

A tracker of US universities lists 38 that have disabled, discontinued or never enabled Turnitin’s AI detection. The recorded reasons are consistent: false positives that could lead to integrity cases against students (Washington State, February 2026); false positives and little evidence that it works (Boston University); the vendor not explaining how it detects AI text (Vanderbilt). Use, meanwhile, is near universal. In a 2026 survey of 1,054 UK undergraduates, 94% use generative AI to help with assessed work, and 65% say assessment has changed significantly in response.

Business-school faculty know where they stand. In AACSB’s survey of 236 deans and 429 faculty, 39% of faculty never use generative AI for student assessment and 65% say it has had no effect on how they track progress and assess. One faculty comment in the report sums up the position: “It’s not always easy to see where AI ideas stop and student ideas begin.”

The repairs, and what each one costs

Schools are not standing still. Each response is a way of getting a record of the work back, and each has a price.

Orals. Frankfurt School is bringing back oral examinations and final presentations, Nova SBE has lengthened thesis defences to test whether the student can argue and defend, and Imperial’s Peter Todd describes the shift to live assessment as a move to the performative arts, all reported in GMAC’s 2026 deans survey. An oral is a real record, taken once, at the end, by the most expensive person in the building. It gives the examiner one student at a time and gives the programme no view of the cohort. The European Business Review piece makes the same point: orals and live cases are harder to outsource, and most schools have not restructured around them because they cost too much.

AI orals. NYU Stern replaced a written final with a 25-minute oral run by a voice agent and graded by three language models, at about 42 cents a student, as The Decoder reported. 70% of the 36 students agreed it tested their real understanding; 83% found it more stressful than a written exam; 13% preferred it. Cheap, and still a test after the fact of work nobody watched.

Process recorders. Turnitin’s Clarity, announced in March 2025, records pasted text, typing patterns and draft history inside a monitored writing space, per Plagiarism Today. That is a record of how the document was typed. It says nothing about what the student understood, what they were taught, or whether the method was followed.

Logbooks. Rotterdam School of Management requires any permitted AI use in a thesis to be documented in a logbook, and students must be able to show their original work and justify AI outputs if the examiner asks, in its thesis manual. William & Mary allows AI for structured tasks if students document how they evaluated, revised or rejected its outputs, according to AACSB. A logbook is a record written by the student, about themselves, afterwards.

Machine off, machine on. HEC Paris, in the same GMAC report, runs pen-and-paper exercises for foundations and separate exercises built around AI use. This is the most honest of the five, because it stops pretending the question is whether AI was used and starts asking what it was used for.

The question is wrong

“Did AI write it?” has no reliable answer, and it has the wrong shape. Every participant will use AI, and the school wants them to, because their employers will expect it. The question that can be answered, every time, is “how was this made?” A detector cannot answer it. A record can.

What a record must have

  • Made in the open. The work arrives as steps, drafts and returns, not as one upload on the deadline.
  • Tied to the method. Each step has criteria written before the draft, so the record shows the work measured against a bar.
  • Shows the teaching and the checking. What was read, what was asked, what was answered, where the participant was pushed back and what they changed.
  • AI in the open, as a tutor that reads, questions and gives feedback, and never writes the work.
  • A person decides at each gate, with the record in front of them, and the decision is recorded too.
  • It travels with the deliverable, so whoever reads the work later can read how it was made.

A record like this does not prove authorship by itself. It makes the question answerable, by showing how the answers developed, and it turns the oral from a detection exercise into a conversation about something both people can see.

Where Ulern stands

We build Ulern, the applied-learning platform for business schools and executive education, so read this section knowing that.

In Ulern the record is the product. For every participant it holds what was read, what was answered, where the AI tutor pushed back, every draft, every return and every signature, and the finished deliverable exports with it. The AI tutor teaches from the school’s own material and gives feedback on each draft; it never decides. A supervisor does, at each gate. We have not run an external study of whether participants try to game it, and a pasted draft can mislead the AI tutor as it would a person, which is why the gate is a person and the record shows how the draft arrived. The capstone walkthrough shows the record on real screens.


This is the first of seven pieces in Assign and hope, a series on what goes wrong while a programme’s project runs, and what a remedy must have. The others: Activities are not outcomes · Supervision does not scale · The 90 days after the programme · What corporate clients now ask of schools · The thesis in the age of AI · What to ask an AI tool before December 2027

← All posts