AI was great until it was not

A physicist recounts arriving at Oxford from the U.S., describing a competitive, impersonal academic environment where classmates increasingly delegated problem sets to ChatGPT. Early encounters with the chatbot produced nonsense, reinforcing skepticism, but widespread reliance among peers eroded collaborative problem solving and shifted norms: homework became a perfunctory step rather than exam practice or a space for learning with others. Faced with a compressed month‑and‑a‑half study period for multiple MSc courses in 2025, the physicist used ChatGPT as an explanatory bridge - feeding it small bits of dense notes to get clearer, example‑driven explanations (for things like torsion‑free conditions, Penrose diagrams, Kerr black holes), which accelerated comprehension and allowed passing closed‑book exams with high passes and merits.

The account draws a line between useful pedagogical assistance and inappropriate research dependency. At Harvard, AI was used sparingly - mainly to find background literature, clarify derivations, and once to patch code with Claude Opus - while real research progress still came from human mentors, notebooks, and hands‑on coding. The essay acknowledges striking successes where internal models produced nontrivial proofs but warns that relying on AI for encoding theories, learning through doing, or producing dissertations undermines craftsmanship and collaboration. It’s a cautionary tale: AI can fast‑track learning by making explanations accessible, but overuse risks hollowing out the skills and social practices that produce original physics.