In an era where artificial intelligence is widely used in scientific research, a new international study issues an important warning: today’s leading AI models remain unreliable in correctly identifying potential laboratory dangers. According to research published in Nature Machine Intelligence, none of the 19 most advanced AI systems were able to provide fully accurate laboratory safety recommendations.
The study used a new testing platform called LabSafety Bench, consisting of 765 multiple‑choice questions and 404 open‑ended scenarios simulating real laboratory conditions. The goal was to assess how accurately AI models can identify chemical hazards, explosion risks, and poisoning threats.
In multiple‑choice questions, some models — including DeepSeek‑R1 and GPT‑4o — achieved around 85% accuracy, correctly handling straightforward tasks such as how to dispose of broken glass contaminated with hazardous chemicals.
However, performance dropped sharply when the models faced open‑ended questions requiring analysis of realistic lab situations. None of the systems exceeded 70% accuracy. These questions involved identifying all risks in a given setup or predicting the consequences of incorrect human actions.
According to the researchers, many AI systems:
- misjudge hazards,
- overlook critical risks,
- and sometimes generate completely incorrect information.
This demonstrates that AI is not yet a reliable advisor in high‑risk laboratory environments. Scientists acknowledge AI’s usefulness in research but emphasize that safety‑critical tasks still require human oversight.
