AI in Education AI Minute Newsroom 2026-08-12

He hid a ridiculous instruction in white text. Forty students handed in the answer it produced.

Roberto Serrano, an economics professor at Brown University, described in Nature on 10 August how he found evidence of widespread AI use on a midterm in his course on welfare economics and social choice theory. He had set the test as a take-home for the first time in nearly two decades of teaching it, because some students were reluctant to return to the classroom after a shooting on campus earlier in the year. Inside one essay question he buried an absurd instruction in white font — invisible on the page, but copied along with the question if a student pasted it into a chatbot, which duly obeyed it. The grades came back with an average of 96, against averages in the 60s to 80s in previous years, and 40 students scoring full marks. Serrano's argument in Nature is that detection software is not the answer and that universities have to change what they ask for: assessment that watches a student think, rather than assessment that collects a finished document.

Why it mattersThe white-text trap is a good story, but it only works once and it only catches copy-and-paste. The number that matters is the average: a class did not get better at welfare economics overnight, and the take-home format that had been safe for eighteen years stopped being safe in one semester. That is the real change teachers are dealing with, and it is not a discipline problem — it is an assessment problem. Anyone marking work this year is effectively grading a document whose authorship they cannot establish, and the institutions moving fastest are the ones giving up on detection and rebuilding the tasks instead: oral defences, supervised writing, and marks for the process rather than the product.

✓ Verified · 3 sources

Read in the app — free, in 9 languages

He hid a ridiculous instruction in white text. Forty students handed in the answer it produced.

Related stories