AI needs a new philosophy of science
Artificial intelligence (AI) is transforming every phase of research, from drafting proposals with large language models to identifying patterns in huge datasets. Today, AI tools help with understanding literature, generating hypotheses, designing and conducting experiments, and even writing and reviewing papers. This surge in capability raises important questions: might AI eventually “solve” science?2 Is it a certain possibility or just a desire, an illusion?3 To navigate this new era, we suggest backing AI-guided science with a three-part philosophy of science. First, integrate AI responsibly through collaborative workflows, epistemic pluralism, and open science. Second, maintain rigor, transparency, and ethical accountability in all AI-driven work. Third, address AI’s blind spots by grounding explainability, trust, and causal reasoning in established philosophical principles. Together, these steps ensure AI advances science responsibly and transparently.
AI as a partner in discovery: Responsible integration
Von Glasersfeld’s radical constructivism views science as building models that work rather than uncovering a hidden truth; knowledge is judged by how well a model helps us predict or act, not by its match to an objective reality. Applied to AI, this means treating its outputs—whether a predicted protein fold or a bold hypothesis—as provisional tools: valuable if they achieve our goals but always subject to skepticism, empirical testing, and open scrutiny.4 This “computational empiricism” invites humility and frames AI as a responsible partner and co-creator in science rather than a replacement for human insight.
But AI alone cannot “solve” science: complex systems often require thorough, step-by-step exploration that no shortcut can replace.5 Human insight and laboratory tests remain vital; for example, AlphaFold’s models still require experimental validation, and AI-designed drug candidates must be synthesized and tested before any claims of discovery can be made. True progress arises from co-creativity: machines suggest bold, unconventional ideas or hidden patterns,3 while researchers apply their domain expertise, critical judgment, and rigorous experiments. This pluralistic interaction echoes Feyerabend’s praise of methodological diversity and recalls that many past breakthroughs resulted from serendipity rather than deliberate planning.
To realize this vision, we must adopt open-science practices, sharing AI models, code, and data so the community can reproduce and challenge findings,4 and invest heavily in educating researchers to explore AI’s assumptions, measure uncertainty, and verify outputs against theory and experiments.2 To harness AI’s creativity reliably, we must thus embed it into core scientific practices.
