When can AI claim a scientific discovery?

Anthropic's decision to launch a molecular biology lab where Claude agents read papers, generate hypotheses, and conjecture about difficult biological problems has renewed a thorny question: when can we say an AI system made a scientific discovery? The lab is designed to explore how AI can contribute to research, from sorting through vast literature to proposing experiments. But the project also highlights how difficult it is to attribute a discovery to a machine. Science is not just an answer; it is a process of framing questions, choosing methods, interpreting noisy results, and convincing a community that a claim is valid. AI can already assist with parts of that process. It can summarize literature, suggest candidate molecules, identify patterns, or draft protocols. In some cases, it may generate a hypothesis that human researchers had not considered. Yet a hypothesis is not a discovery. It becomes knowledge only after testing, replication, and peer scrutiny. Human oversight remains essential for recognizing when an AI output is novel, plausible, or simply wrong. Models can hallucinate, miss context, or reproduce existing ideas without knowing it. There is also a philosophical issue: if a human asks the question, builds the tools, curates the data, and interprets the result, is the discovery really the AI's? The answer may depend on how much independent reasoning and creative contribution the system provides. MIT Technology Review has explored these implications as AI moves deeper into fields such as drug discovery, materials science, and genetics. The practical stakes are high. Research institutions need standards for credit, authorship, safety, and reproducibility when AI agents participate in experiments. Funding agencies may want to know whether AI is accelerating science or just producing more plausible text. Anthropic's lab is less a final answer than a test case. It shows that the debate is no longer hypothetical. As AI agents become more capable, the scientific community will need to decide what counts as machine discovery and how to verify it without exaggerating what the technology has done.