Two studies involving 250 participants link the way people interact with an AI assistant to their subsequent ability to solve problems independently.
The help of an AI improves immediate performance, but several studies suggest that it can degrade it once the assistant is removed. Researchers, including James Zou, analyzed two studies based on problem-solving games (139 and 111 participants) by classifying each exchange with the AI according to a taxonomy borrowed from tutor-student dialogue.
What matters is cognitive participation
Participants who verbalized their reasoning then obtained better results without help, while those who directly asked for the answer did less well. The same pattern is found on the assistant side: explaining the state of the problem is associated with better autonomous performance, directly giving the next blow to less good ones.
The authors also observe that seemingly similar messages can elicit very different aids. Their conclusion: preserving the user's mental engagement, rather than providing them with the solution, would be the key for AI to benefit learning beyond the interaction itself.
Read the full article on the original source
Go to the source →