A Co-Author That Cannot Sign: Stress-Testing Language Models at the Bench

In 2026, McKnight Lab ran a small controlled stress test of large language models in three workflows that touch nearly every project here: protocol drafting, literature triage, and analysis-code review. The question was not whether the models could produce plausible text. They can. The question was whether a careful reader, six months later and one bench over, could reproduce the work after a model had touched the record.

The short answer is sometimes. The longer answer led us to add a brief AI-use disclosure to every methods section. Not as a confession, and not as an endorsement, but as a piece of provenance. A language model can draft a sentence. It cannot sign for the experiment.

Why We Ran the Test

Our lab has no blanket policy against AI tools, and we did not want one. We wanted evidence about where these tools help and where they quietly alter the scientific record. The test was deliberately modest: twelve internal protocols spanning cloning, cell culture, and sequencing library preparation; a defined set of literature searches; and several analysis scripts already in use. Each workflow was run with and without model assistance, and the outputs were reviewed by lab members who had not prepared the original material.

We were not testing for elegance. We were testing for reproducibility. Could another scientist follow the resulting document, with the same reagents and instruments, and get the same result? That standard is less glamorous than a fluent paragraph, but it is the one that matters.

Three Workflows, Three Different Risks

Protocol Drafting

Protocol drafting showed the clearest time savings. When we supplied detailed bullet points, reagent concentrations, incubation times, and instrument settings, the models produced usable first drafts quickly. The savings were real, often on the order of a third of the drafting time. The risk was equally real: the models tended to smooth over gaps. A missing step became a plausible step. A vague instruction became a confident one. In one draft, a room-temperature incubation was rendered as a specific duration that no one in the lab had specified. It read better than the original. It was less true.

Literature Triage

Literature triage was useful at the level of sorting, not summarizing. The models helped group abstracts by theme and flag papers that might merit a closer look. They also produced citations that looked correct and were not, and summaries that blurred the distinction between a paper's findings and its discussion. We now require DOI-level verification for any citation that enters a lab document. A model may suggest a paper. It may not certify one.

Analysis-Code Review

Analysis-code review was the most instructive and the least forgiving. Models caught missing random seeds, hard-coded file paths, and a few fragile assumptions in scripts. They also proposed edits that changed filtering thresholds without announcing the change. In one case, a suggested simplification would have altered the number of retained cells in a single-cell analysis. The code ran. The result was different. That is the kind of failure that does not announce itself at the bench; it announces itself months later, when someone tries to reproduce the figure.

Where the Time Went

The time savings were uneven. Drafting benefited most. Literature triage saved reading time but added verification time. Code review was close to neutral once we accounted for the human review needed to catch silent changes. The lab did not recover a dramatic number of hours. It did recover attention for the parts of the work that require judgment: choosing controls, interpreting anomalies, and deciding what the data can support.

That is a modest claim, and we are keeping it modest. The models were useful assistants in narrow, well-specified tasks. They were poor substitutes for the scientist who knows why a step exists.

The Reproducibility Trap

The most important finding was not a particular error. It was a pattern we came to call fluent omission. Model-assisted text often reads more smoothly than the notes it replaces, and that smoothness can hide the absence of critical detail. A protocol can sound complete while leaving out the reagent lot, the centrifuge temperature, or the exact version of a software package. Reproducibility does not fail only when a step is wrong. It also fails when a step is missing and the prose does not miss it.

We also saw plausible normalization: the model would convert a lab-specific instruction into a generic one, or resolve an ambiguity that should have been preserved and discussed. In a methods section, ambiguity is not always a flaw. Sometimes it is a signal that the lab needs to make a decision. A model that resolves the ambiguity for you has made a decision on your behalf.

A Short Disclosure, Not a Scarlet Letter

We settled on a short disclosure in the methods section. It names the tool and version, the date and scope of use, the task the model performed, and the extent of human verification. It also states whether any model-generated text was adopted verbatim. The disclosure is not meant to stigmatize AI use. It is meant to give readers the same thing we ask for elsewhere in open science: enough context to judge the work.

That context matters for reproducibility. If a model helped draft a protocol, a reader should know which parts were machine-assisted and which were bench-verified. If a model helped review code, a reader should know that a human approved the final version. The disclosure is short by design. It should be easy to write and easy to read, not a second methods section.

What We Require Now

Our current practice is straightforward. Models may draft from detailed human notes, but no protocol enters the lab record without a bench-side review against the original source. Models may assist literature triage, but citations require independent verification. Models may comment on code, but they do not approve it, and any suggested change must be documented and tested. Every methods section includes the short AI-use disclosure when a model contributed to the text, the search, or the review.

None of this is a permanent policy. The tools change quickly, and our standards should change with evidence. But the principle is stable: the scientist remains accountable for the claim. A model can be a co-author in the loose sense that it helps produce a draft. It cannot be a co-author in the sense that matters to a reader, a reviewer, or a replication attempt. It cannot sign.

The bench is still the judge.