AI Is Being Tested on the Messy Reality of Lab Science, Not Just Perfect Data

OpenAI has introduced BioBasicsBench, a new benchmark designed to test whether AI agents can make sense of the kind of imperfect evidence scientists encounter in real wet labs. That distinction matters. AI systems often look impressive when they are given clean charts, polished papers, or neatly labelled datasets. Actual laboratory work is far less tidy: blurry images, failed controls, confusing measurements, handwritten notes, and results that may not support a clear conclusion.

BioBasicsBench contains 100 tasks across five common areas of lab work: protein methods, imaging and cytometry, assays and quantification, nucleic-acid methods, and molecular cloning. The benchmark gives models raw material such as gel images, microscopy images, sequencing files, instrument exports, photographs of equipment, and videos of bench procedures. The goal is not simply to get an answer—it is to determine whether the AI can spot mistakes, interpret uncertainty, and recognize when the data is too weak to justify a conclusion.

OpenAI developed the benchmark with enzyme-design company Ligo. Scientists created each task from real experimental data and reviewed it, including an external PhD-level wet-lab reviewer. Models receive credit not just for their conclusion but also for measurements, reasoning, and whether they correctly flag limitations in the evidence.

In OpenAI’s reported results, GPT‑6 Astra scored 66.7% at its highest reasoning setting, compared with 43.8% for GPT‑5.6 Sol. That is substantial progress, but it also leaves plenty of room for caution. A score of roughly two-thirds on routine scientific tasks is useful evidence of capability, not evidence that a model should be trusted to run research independently.

The broader story is that AI evaluation is becoming more realistic. Instead of asking, “Can a model answer a biology question?” BioBasicsBench asks, “Can it handle the ambiguity that makes real science difficult?” That is a much more valuable test for biotech teams, research labs, and health-tech builders.

The opportunity is enormous: AI that helps researchers troubleshoot experiments could reduce wasted time and accelerate discovery. The risk is equally clear: a confident but incorrect interpretation could send a team down the wrong path. The benchmark’s emphasis on uncertainty may be its most important contribution.