LLM assistance significantly increases novices' success at wet-lab molecular biology and virology tasks.
Assessed doubtful no decomposition
First pass on a small budget. Kept the claim atomic: its evidence is a small number of direct trials, so there are no reusable subclaims worth creating yet. Recorded Hong et al. 2026 as a denying instance (confidence 0.8, since the paper reports a modest non-significant benefit). Assessed doubtful at credence 0.3, in line with the parent's seed. A future pass should read the Los Alamos trial (Romero-Severson et al.) and check for RCTs using newer models.
Assessed Doubtful
verdict confidence 0.65 · credence 0.30
The best direct evidence does not show a significant increase. The largest test is a preregistered, investigator-blinded randomized trial run by Active Site in a BSL-2 lab in Cambridge, Massachusetts (Hong et al., 2026). It enrolled 153 novices over eight weeks, June to August 2025, on five tasks modelling a viral reverse genetics workflow, using mid-2025 frontier models. Full workflow completion was 5.2% with LLM access and 6.6% with internet access only (P = 0.759), and no single task differed significantly. The LLM arm did score numerically higher on four of five tasks, most clearly cell culture (68.8% vs 55.3%, P = 0.059). A post-hoc Bayesian pooling put the gain at about 1.4-fold per typical task, with a 95% credible interval of 0.74 to 2.62. A smaller Los Alamos trial on protein expression in E. coli was also reported as non-significant. So the evidence points to a modest benefit at most, and the end-to-end success the claim implies was not found. "Significantly" is the sticking point: a 1.4-fold gain that the trial cannot tell apart from zero does not meet that bar, whether the word is read statistically or practically. Two things could change this. The trial tested mid-2025 models used by participants with little LLM skill (about 40% never uploaded images), so newer models, agentic or vision-guided assistance, or trained users could produce real uplift. And the trial was underpowered for modest effects. The status therefore describes the evidence as it stands for 2025-era systems, not a ceiling on what LLMs can do.
Claim entered the graph