A student’s chat history cannot by itself tell you whether learning happened

Researchers analyzed roughly a million conversations associated with higher-education email addresses, then filtered for student academic use. They found a mixture of direct answers, collaborative problem solving, and content creation.

The same visible request could represent very different behavior. Asking for a solution might be cheating on an assessment or checking a practice exercise. A multi-turn explanation might support understanding or simply outsource the thinking more politely.

This ambiguity limits conclusions about both harm and benefit. The study also captured one product’s early adopters during a short window, not a representative census of students. Its practical value is in documenting the work students were asking AI to do—and showing why that evidence needs to be connected to assessment rules and measured learning before it can settle arguments about education.