Large language models and other artificial-intelligence systems could be excellent at synthesizing scientific evidence for policymakers — but only with appropriate safeguards and humans in the loop.