Handing Labs to AI: Who Signs Off?

I just saw news that a materials scientist dared to hand over his lab to AI. My first reaction was occupational hazard kicking in: Has clinical validation been done? Who takes the blame for failed samples?

Back when I was doing AI imaging at United Imaging, hospitals initially asked if the model was accurate. Later, doctors asked more practical questions: Can it integrate with PACS? Can reports be written back? Who is responsible for missed diagnoses?

No matter how pretty the metrics are, if no one dares to sign off, the product won’t land. Materials labs seem "braver" than hospitals, but the bottlenecks are the same: AI usually outputs a string of experimental plans awaiting verification.

Google’s Co-Scientist has been discussed a lot. It’s said to integrate Gemini into real scientific research workflows, designing experiments, generating execution code, reading experimental feedback, and even linking with lab equipment. The direction is aggressive, but I agree with Yu Peiyuan from SUSTech who works on AI + Materials: the real difficulty lies in verification. Generating a bunch of hypotheses is fast; filtering out the one that is reproducible, explainable, and fits into standard processes is what kills you.

So my judgment is that AI can take over repetitive labor in the lab, but it cannot take over scientific responsibility. It’s suitable as a high-throughput trial-and-error machine, not as the final judge. Robots can enter furnaces, pipette, and read values; models can rank candidate solutions. But once results need to be written into papers, patents, or clinical translations, the person signing off must still understand why they trust it.

You can let AI run the lab, but someone has to sign the failure report. If no one signs, the process stops at AI.