The first wave of student-facing AI tools — Photomath, Chegg, Course Hero — optimized for a single metric: time-to-answer. The user's question in. The user's answer out. Session over. Retention success.
The second wave — ChatGPT and its imitators — optimized for the same metric, dressed up. The chat interface felt like tutoring. The answer arrived faster than any textbook. The illusion of learning was persuasive because the artifact of learning (an explanation) was present. But the substrate of learning (the student's own generative work) was absent.
The metric these products chase is not learning. It is convenience. Convenience is a fine metric for many product categories. It is a poisonous metric for a learning product.
Brainback's metric is different. We optimize for the session that ends with the student thinking they've done real work — because they have. Attempt count, hint depth, explain-back accuracy, weekly Brain score movement. None of those metrics reward giving the user the answer.
This alignment problem — the product's metric versus the user's outcome — is why the AI tutor category needs a different design brief. Not another chatbot. A tool that reliably makes users capable of thinking without it.
Written by Brainback Research. Published Jun 22, 2026. Filed under Product design.