Every AI-use policy I have read this year is doing two jobs at once. One is the job the executive who signed it believes they signed for. The other is invisible to them, and it is the one that actually moves the outcome.
The promotion packet that scored high and lost
Last week I was on a coaching call with a Chief People Officer at a professional services firm in Amsterdam. Her firm was three weeks into promotion committee season. She wanted to talk about the AI-disclosure line she had added to the packet template. Then she got quiet and told me what had actually prompted the call.
Her strongest technical candidate, a senior consultant with a clean book of business, had been flagged in the committee read. Two members had noted, in different words, that the packet read as if it had been polished by a model. Neither would say it in the room. Both had scored the packet lower than the candidate's numbers said they should.
She had spent the weekend re-reading it. The writing was good. The argument was tight. The judgement was there.
That was the problem.
What Harvard just measured
A new working paper from Isley, Gaebler and Goel (Harvard and NYU, September 18) covers six admissions cycles at a large U.S. public policy master's program. That is 7,462 applications, 2020 to 2025. Plus a controlled experiment with five admissions officers scoring 100 essays.
Three findings, all in the same paper.
The ban failed. 56% of 2025 applicants submitted at least one primarily AI-generated essay despite an explicit prohibition. 69.3% of international applicants.
The AI improved the writing. Prose fluency, structural clarity, mechanical quality all measurably up after ChatGPT's release.
The reviewers detected AI writing at AUC 0.70 (95% CI [0.65, 0.75]) and marked those essays down. The paper's clean magnitude, quality-adjusted:
a penalty of 2.6 p.p. for each AI-written essay, or about 6.0 p.p. for the typical applicant who used AI.
The essay got better and the applicant got worse.
That is the two-jobs problem. The written policy is doing the compliance job badly, and a majority ignore it. The unwritten policy, the reviewer's detection reflex, is doing its job well enough to move the outcome. If you own an evaluation process in 2026, banning harder will not help. Your options are a disclosure norm that ends the guessing game, or a scoring rubric that stops rewarding the surface features the detection reflex is reading. My Amsterdam client spent this week rewriting the rubric.