Stanford researchers debate whether AI can guide humans through wet-lab experiments

On September 25, Stanford computational biologist Anshul Kundaje and researcher suragnair debated on X the idea of AI guiding humans through biological wet-lab experiments. The core disagreement: whether the dexterous manipulation and hard-to-articulate "feel" that many biological experiments depend on can be transmitted to human novices via AI's verbal instructions.

Confirmed

• suragnair's original idea: rather than executing experimental steps itself, the AI learns how to optimally guide humans; after a human performs an operation, the AI suggests corrections based on fully observable information, forming a collaboration loop of "AI proposes, human executes, iterate to a verifiable result"—seen as an incremental path for AI agents to participate in real scientific experiments
• Kundaje objected: most biological experiments rely on dexterous hand movements and tacit "feel"; human novices learn by first watching experts and then trial and error, whereas vision-based AI trained purely on observation lacks explicit motor control, so novices may not be able to perform operations from AI instructions alone
• suragnair responded: in this scenario the AI isn't learning to run experiments itself but to optimally guide humans; he also conceded that AI trained in an "observe everything" manner would be more passive
• When Kundaje pressed why not use supervised learning on "expert vs. novice operations" (seemingly faster), suragnair explained the rationale for RL, centering on the difficulty of attributing blame when experiments fail

Not yet confirmed

• The idea remains at the discussion stage, with no actual system or experimental data supporting its feasibility

Why it matters

• The debate touches a key bottleneck in AI for Science: whether tacit knowledge in scientific experiments (feel, dexterous manipulation) can be verbalized and transmitted by AI. If the "AI plans, humans act" collaboration loop works, AI agents could enter real experimental loops at much lower cost without first solving robotic motor control

2026-09-25 ~ 2026-09-25 · 7 related posts

• [source] Researcher proposes agent-suggests-human-executes loop for real-world science experiments — suragnair · 2026-09-25
• [source] Why RL beats imitation for AI-run experiments: failure attribution is the bottleneck — suragnair · 2026-09-25
• Suragnair responds: AI's role in wet labs is to instruct humans, not execute — suragnair · 2026-09-25
• [source] Bioinformatician Anshul Kundaje doubts AI can talk novices through wet-lab experiments — anshulkundaje · 2026-09-25
• Researchers debate whether observation-trained AI can guide hands-on biology experiments — suragnair · 2026-09-25

2 near-duplicate retellings: anshulkundaje · suragnair