Years into the generative AI debate, in 2026 we are still mostly arguing about cheating. I think that has been an expensive mistake. Weber-Wulff and colleagues (2023) tested fourteen detection tools and found none reliable enough to use, with a bias towards calling AI text human. A score that errs in both directions accuses the innocent as readily as it catches anyone.
Meanwhile the ground has shifted. Ninety-four per cent of UK undergraduates now use generative AI for assessed work, and one in eight paste its output straight into a submission (Stephenson & Armstrong, 2026). Whatever the policy says, that is the cohort in the room.
What has been damaged is not honesty but validity. Corbin, Dawson and Liu (2025) separate discursive changes, which are rules students remain free to ignore, from structural changes, which alter how a task actually works. Traffic-light policies borrow the language of enforcement without any of its machinery. This is the enforcement illusion.
So the useful question is not whether an AI tool was used. It is whether the task still supports the claim you make on the strength of it. An essay was always a proxy, and some proxies survive contact with a language model while others do not.
Reviews
There are no reviews yet.