Science in the Age of AI: Golden Age or Long Night

Foundation models are here to stay, and their imprint on the daily practice of scientific research is deepening---from literature synthesis to hypothesis generation to an expanding role in peer review itself. A clear grasp of the epistemological implications of adopting AI in science is therefore timely, if not urgent. In this essay we revisit the most established accounts of how scientific knowledge is generated---Kuhn's paradigms and revolutions, Lakatos's progressive and degenerating research programmes, Popper's falsificationism---and use them as a lens through which to examine the dangers that an uncritical adoption of AI-based methods for discovery may pose to the process. We argue that a science built on statistical reinforcement risks entrenching the prevailing paradigm while smoothing away precisely the outliers from which revolutions grow, and that it lacks a mechanism the human enterprise has always relied upon: the selective death of superseded ideas. We finally sketch an alternative vision---one that keeps the machinery of scientific progress open to the world it is meant to describe---and outline what it would take to ensure that machinery keeps turning.