Does the novice wet-lab null survive a qualified operator with AI?

Research brief closed 23 October 2026. Question: with laboratory staff and equipment already supplied, can AI accelerate physical work compared with qualified humans without AI, and what still interrupts misuse? A good answer has matched expert controls, safe measured physical endpoints, time and all-in inputs, access/approval logs and defined objective. Existing novice physical trial had 4/77 versus 5/76 completions, whereas written novices gained substantially. Neither decides the equipped-operator case.

1. Qualified-operator human comparator: Primary-source search for expert randomized wet-lab comparisons found the OpenAI/Red Queen Bio benign expert-led protocol result already covered in our physical note, and the separately important OpenAI/Ginkgo six-round cloud-lab original. Ginkgo reports a selected small-assay protein formulation with 40% lower reaction-material cost per output than an earlier human-team formulation, but the earlier team screened 1,231 mixtures versus tens of thousands tested with an automated lab; no concurrent human arm with matched lab, budget, tools and objectives. Protein yield/cost is a safe real physical endpoint, not a test of harmful biological production. A search of expert wet-lab RCT originals found no stronger matched original; absence in this search is not proof none exists.
2. Who did what and at what cost: The human researchers chose the benign goal, supplied the pre-existing platform, consumables and later web/paper access and starting-material improvements; Ginkgo staff prepared materials and wrote robot-execution protocols. GPT-5 chose experiments within predefined machine-executable constraints and iterated over returned measurements. The published $422/g versus $698/g comparison is a unit reaction estimate; compute, the campaign, robot capital and personnel are not costed into it. The prior team’s published $60/g and $39/g are reagent-only costs at different scales; they should not be presented as directly contradicting the $422/g all-reaction-materials small-assay comparison. OpenAI’s shorter authored explanation discloses human reagent oversight and single-system scope.
3. Gates for a different actor: NIST’s ongoing screening supplies actual provider-facing measurements on selected sequences and lawful reference-material orders. It tests commercial procurement, not a trained insider with approved local stock or the qualification checks of a cloud laboratory. WHO’s 2024 laboratory biosecurity guidance specifies institutional assessment, staff and material controls and national oversight for high-consequence work: it is normative, not an independent audit that a given site implements those controls. No primary result here quantifies attempts by malicious qualified personnel or an independently validated cloud-lab denial rate. Do not claim that this benign automated experiment gave untrusted outsiders access to hazardous materials, or that synthesis screening closes an insider channel.

What changed: the physical barrier is contingent on who supplies the lab, not just whether the model knows bench technique. This is a meaningful push against a blanket 'AI cannot do physical biology' objection while respecting the novice null and the much larger unobserved leap from optimizing a safe protein to harmful creation and release. Needed study: prospectively compare expert-only, AI-assisted expert and autonomous-with-supervision arms on the same safe physical tasks and resource cap, with all-in compute/human/material cost and authenticated approval/denial traces. Stop here; further search would mostly repeat benchmark headlines rather than measure those missing joins.

Source list: OpenAI/Ginkgo original, Olsen human-team original (final publication March 2026), NIST operational screen page, updated 1 October 2026, WHO normative guidance, Hong novice RCT. Verified load-bearing 40% number and its denominator in Ginkgo paper pp. 2, 6, 12–13; 1,231 comparison in Olsen’s original; NIST denominators on its updated page. These are different outcomes, not contradictory findings. Feed: the short OpenAI authored piece is held for a later weekday bio interest or Saturday only if the waiting item clears; do not post a second unread risk track now.