A lot is being promised for AI in the lab: describe an experiment in plain language and have a protocol appear, have agents plan and run workflows, have systems adapt on the fly. Much of it is real and useful. But there is a hard boundary that no amount of model capability moves. AI can arrange steps and reason about a workflow; it cannot invent how a particular liquid behaves in a particular tip. That knowledge has to come from somewhere, and that somewhere is a validated liquid class.
What AI is good at, and what it is not
Generating and sequencing a protocol is a language and planning problem, which is squarely what modern models do well. Knowing that a given detergent at 5 microliters on a given tip needs a specific flow rate, settling time, and blowout is not a language problem. It is an empirical fact about the physical world, established by running the liquid and measuring the result. A model can guess at it, but a guess dispensed is still a guess, not a validated transfer.
Ground truth has to be structured
For an automated system to use a liquid class rather than invent one, the class has to be machine-readable: explicit parameters, the liquid and family it applies to, the instrument and tips it was validated on, and the results that back it. Prose buried in a lab notebook cannot be referenced reliably; a structured, versioned record can. The better the structured ground truth, the less an automated system has to improvise about the one thing it cannot safely improvise.
Provenance becomes more important, not less
When a protocol might be assembled by a system rather than a person, being able to ask where a class came from, what it was validated against, and when it last changed matters more than ever. Automation removes the human who would have noticed something looked off, so the record has to carry the caution that the human used to.
AI can decide what to do; it cannot decide how a liquid behaves. Give it validated, structured classes as ground truth, or it will confidently fill the gap with a guess.