AI in scientific discovery and the bottleneck in testing hypotheses

DeepMind Institute says its study draws on 15 million Gemini interactions, more than 2,600 specialized AI models and a survey of more than 600 scientists. It finds that LLMs and specialized models serve different tasks, while surveyed scientists report backlogs of untested hypotheses and time spent checking AI outputs. The authors argue that AI could open up new kinds of discovery, but doing so will require investment in testing and changes to scientific institutions.