AI tutors struggle to know when to help versus push students, study finds

The tension TutorMoments highlights sits at the heart of educational practice: helpful AI assistants are trained to do the hard work for the user, but in tutoring, doing the work for the student can short-circuit the "productive struggle" that learning research has long tied to deeper understanding. Most existing benchmarks for AI tutors reward single fixed behaviors (never give away the answer, always offer a hint) without checking whether those behaviors fit the student's actual moment. TutorMoments shifts the frame by grounding evaluation in real tutoring data and asking a harder question: can the model read the room and adapt?

The scoring results reveal two layers of limitation. First, models' default behavior under generic prompting is to over-help—a sign that the training objective (be helpful) conflicts with the pedagogical goal (let the student do the work). Second, even when the prompt spells out the trade-off in detail, models improve across the board but remain inconsistent and less reliable than human tutors. The gap is particularly stark on rigor: models rarely push students to explain their reasoning or work independently. The researchers note, importantly, that the evaluation uses a simulated student, so these numbers measure tutor behavior at a decision point, not actual learning outcomes—a limitation the team plans to address as the framework expands.