BE HUMAN LABS Talk to us
← From the lab
Research note

When AI can pass the exam, the exam has to change.

AI now reaches gold-medal level at the world's hardest school mathematics contest. Exams that test recall and procedure tell us less than they used to. What to measure instead, and how, without burying teachers.

Shikhar Anand, FounderOctober 20264 min readOpening only

In July 2025, an AI system from Google DeepMind reached gold-medal standard at the International Mathematical Olympiad, confirmed by the competition's own graders. OpenAI reported the same score for its system, graded by former medallists. Both wrote their proofs in plain language, within the same two four-and-a-half-hour sessions as the human contestants. Only a few years earlier, leading AI systems struggled with school-level word problems.

An exam is a sample. In a few hours, it tries to tell us what a student has learned over a year. For a long time, recall and procedure stood in reasonably well for understanding: a student who could remember a method and apply it had usually understood something. That link is now much weaker. Recall and procedure are exactly what AI does best, and students know it.

If a machine can produce the answer, the answer alone tells us very little about the child.The starting point for rethinking assessment.

Talk to us

About: