Science Locked in Local Minima by Path Dependence as AI Scales Discovery
Scientific knowledge exhibits path dependence that traps progress in local optima, a dynamic AI automation risks amplifying at scale according to synthesized primary sources.
The arXiv preprint arxiv.org/abs/2604.11828 by Mabrok frames scientific knowledge as a local rather than global optimum, analogizing its trajectory to gradient descent that follows gradients of tractability, empirical accessibility, and institutional reward. It presents case studies across mathematics, physics, chemistry, biology, neuroscience, and statistics to illustrate cognitive, formal, and institutional lock-in. Kuhn's The Structure of Scientific Revolutions (University of Chicago Press, 1962) similarly identifies paradigm entrenchment; Merchant et al., Nature 624, 80–85 (2023) show AI models accelerating materials discovery but strictly within existing data distributions and frameworks.
AXIOM: Scientific knowledge follows local gradients of convenience and reward, not global truth; AI systems trained on these corpora will automate and scale the same suboptimal traps unless explicit exploration mechanisms are engineered.
Sources (3)
- [1]The Non-Optimality of Scientific Knowledge: Path Dependence, Lock-In, and The Local Minimum Trap(https://arxiv.org/abs/2604.11828)
- [2]The Structure of Scientific Revolutions(https://press.uchicago.edu/ucp/books/book/chicago/S/bo13179781.html)
- [3]Scaling deep learning for materials discovery(https://www.nature.com/articles/s41586-023-06735-9)