Same mark, different work: 150 student-AI conversations show that the assessment no longer recognises who has put thought into it

A study on ‘epistemic ownership’ shows that assignments marked in the same way conceal very different levels of intellectual engagement between the student and the AI.

When a student works with a generative AI, the final note tells about the quality of the rendering, but less and less about the reflection that produced it. This is the observation of an article published on arXiv on October 2, 2026, which analyzed 150 conversations between students and an IA assistant.

Three dimensions of participation

The authors use the notion of "epistemic property" to identify, in the exchanges, the part of the intellectual work that really belongs to the student. They observe it in three directions: that guides the approach (direction), that assembles ideas (integration) and that judges the relevance of the responses (evaluation).

The result is clear: assignments with comparable scores are based on very different ways of spreading cognitive effort with the machine. The note therefore does not allow to reconstruct what happened during the interaction. Researchers see this as an empirical basis for rethinking evaluation, integrating traces of the process rather than judging only the final product.