Universities should assess student learning both with AI assistance and without it, according to a commentary published in Nature on 3 November 2026. The author argues that performance during an AI-supported task does not necessarily show what a student has learned independently.
The commentary responds to Marc Schmitt’s argument that universities should teach students to challenge AI models and judge the quality of their responses. It says assessments should also establish how well students perform after the AI tool is no longer available.
The author points to a randomized trial involving secondary-school mathematics students. Unrestricted access to a chatbot improved scores during practice, but students later performed worse on an exam taken without the tool. Measuring only the assisted work, the commentary warns, could therefore make an AI tutor appear more effective than it is.
Sources: Nature.