Why science matters for AI safety
Scientific answers rest on mechanisms, methods and evidence. A model can produce an explanation that uses the right vocabulary and still describe something that does not happen.
Researchers and students increasingly use models to read papers, plan experiments and interpret results. Training those models well takes scientists who can tell where an argument breaks.
What science experts do
Science experts write worked answers and explanations, rank model responses on accuracy and reasoning, and annotate where a chain of reasoning goes wrong.
They evaluate research and lab-assistant agents, review how models interpret data and figures, red-team models on hazardous procedures and requests, and produce alignment data on uncertainty and the limits of the evidence.
- Expert-written explanations and reasoning
- Ranking and grading model responses
- Evaluation of research agents
- Red-teaming on hazardous procedures
- Alignment data on uncertainty and evidence
Failure modes science experts catch
Scientists catch mechanisms that sound right but are not, results misread from a table or figure, and unit or order-of-magnitude errors. They notice when a single study is presented as consensus.
They also catch practical risk: a procedure missing a safety step, or a response that gives hazardous detail a model should withhold.
How we vet science experts
Science candidates are interviewed by AI voice agents on how they reason through problems in their discipline. They then complete real-world work tests, such as assessing an experimental method or judging a model's scientific explanation.
Credentials and experience are reviewed as part of vetting. Physicists, chemists and biologists are each tested in their own field.