What should an AI research assistant show before its answer becomes a lab instruction?
Faraday’s published test is narrower than the headlines: Inherent says it recreates figures from 100 machine-learning and AI-for-science papers under fixed time and compute limits. That is software research, not a system at a lab bench. Good. The moment an assistant’s answer becomes “change this setting” or “mix this,” I want its working record on the screen: source paper, data or sample version, instrument, calibration date, assumptions, failed attempts, and the point where it stopped knowing. A recreated chart can be a useful checkpoint. It should not quietly become permission to touch an experiment. What would you need to see before trusting an AI research assistant with a real lab step?
Comments
Before a junior researcher reaches the protocol, put one line at the top: Figure recreation in software — not tested with your materials. A recreated plot can look so finished that the whole page reads like a ready-to-run experiment. When the assistant moves from paper to proposed lab plan, make that a visible switch and leave the local blanks empty: instrument, materials, approval. Do not let the polished chart quietly fill those in.
Keep a blank result visible too. If the assistant cannot support a proposed change, “nothing verified” should be allowed to end the task. Otherwise every uncertain answer gets polished into a Friday experiment because the system is rewarded for producing motion.
Give ‘nothing verified’ its own outcome in the pilot. On comparable requests, count correct holds, false holds that send someone hunting for an answer, unsafe releases, and review minutes. A tool that refuses every uncertain task can look careful while handing the whole queue back to the lab. The point is fewer bad instructions without making a scientist re-check every blank response.
Exactly. And the local blanks should not be decorative. A paper can show how a figure was made without telling an assistant what changed on this instrument last week, how this sample was prepared, or whether the room conditions matter. Leave those fields empty and ask. Guessing fills a notebook; it does not make an instruction safe to follow.
The handoff needs a person’s name too. “Try this change” is not a useful instruction if the junior person who sees it at 6:40 has no idea whether their supervisor signed off or whether it can wait until morning. Put “reviewed by ___” or “hold for ___” beside it. Otherwise “AI-assisted” becomes one more thing a tired person has to interpret alone.