AI tutoring systems are evaluated on whether they produce correct answers. They should be evaluated on whether they teach correctly — a different standard with different failure modes (arXiv:2603.17373).
Pedagogical safety is not factual safety. A tutor can be factually correct and pedagogically harmful: giving the answer too quickly (preventing the student from discovering it), correcting a productive misconception before the student has explored it, or escalating difficulty before the student has consolidated the current level. These are not errors in the content. They are errors in the delivery — violations of pedagogical principles that experienced human teachers follow instinctively.
In single-turn interactions, AI tutors fail pedagogically 17.7% of the time. In multi-turn dialogue — where the tutor must track the student's understanding, adjust its strategy, and maintain appropriate scaffolding across exchanges — the failure rate rises to 77.8%. The knowledge is correct. The pedagogy is broken.
The jump from 17.7% to 77.8% is structural. Single-turn pedagogical safety is a property of the response. Multi-turn pedagogical safety is a property of the trajectory — the sequence of scaffolding decisions that must be maintained across many exchanges. The model can produce any individual pedagogically sound response. It cannot maintain a pedagogically sound sequence because each response is generated without persistent commitment to the pedagogical strategy established earlier in the conversation.
This is the familiar multi-turn drift problem, but applied to teaching: the tutor's approach drifts from guided discovery to direct instruction to impatient correction as the conversation proceeds. Each turn is locally reasonable. The trajectory is pedagogically incoherent.
The tutor who knows everything and teaches nothing is the failure mode that benchmarks miss.