Metascience needs to attend to both day and night science.

Day science is the industrialised academy: reliable, recurring increments of knowledge, delivered via shared norms and established sociotechnical systems. Night science is the unlooked-for insight and serious play that throws up novel conjectures and, occasionally, paradigm shifts. Done well, they feed each other.

With AI, there’s a risk we end up in an epistemological Antarctica. EverDay looks like success: the pipelines churn, the metrics climb, and the anomalies that might spark something new get smoothed away as noise. EverNight is conjecture unmoored from evidence. It runs from crackpot theories and ideas we can’t comprehend, to ones that sound plausible but can’t be falsified.

Polar seasons, with their months-long days and nights, disorient. Like humans, science needs a circadian rhythm to function, a tick-tock where ideas still answer to evidence and evidence can still surprise us. The task is to find that rhythm while respecting Sutton’s law.