The Discovery Gap: Why Artificial Intelligence Must Learn to Change Its Mind "Rethinking Human–AI Collaboration from Molecule to Market"

Scientific discovery requires the revision of belief: giving up a claim when the evidence turns against it, and knowing what else must change with it.Artificial intelligence (AI) has transformed research -predicting protein structures, generating molecules, screening billions of candidates -but it does not do this.Autonomous agents traverse the discovery loop end to end without holding the beliefs that loop exists to revise.I use dynamic scientific intelligence for the class of system this requires, one that holds hypotheses as persistent objects with provenance, uncertainty and explicit dependencies.Its distinguishing condition is that a system can identify what supports each commitment and what depends on it, and so revise one without disturbing the rest.Because no architecture can revise beliefs on evidence it never receives, the problem is jointly architectural and infrastructural.Using drug development from molecule to market as the test case, I identify where evidence generated downstream should revise commitments made upstream and does not, and three decision points at which the non-delegable human act is adjudication rather than generation.I propose a revision test -an evaluation distinct in kind from any predictive benchmark -and a minimum reporting standard for claims of AI-assisted discovery.Until AI learns to change its mind well, it will remain a tool rather than an infrastructure in which discovery accumulates.