A study of nearly 1,000 high school students shows that access to a standard chatbot degrades performance once the tool is removed, unlike a restricted version.
Among the highlights of the summer on AI and learning is an experiment conducted by the Wharton School and published in PNAS among nearly a thousand high school math students, divided into three groups: unassisted, with a standard chatbot, and with a restricted version that provides cues and encouragement without ever delivering the complete solution.
The result is clear and counterintuitive for those who measure effectiveness by the quality of the homework submitted. Students who used the standard chatbot obtained worse results once the tool was removed: the audience had produced correct exercises without producing learning. Those who had the restricted version achieved results comparable to traditional methods.
The design lesson is straightforward: it is not the presence of AI that determines the pedagogical effect, but the way in which it rations aid. This conclusion converges with another shift observed in educational research, where we are moving from randomized controlled trials to implementation work using short measurement cycles in real conditions.
Read the full article on the original source
Go to the source →