AI models outperform virologists in lab procedure knowledge
Researchers from multiple institutions found that AI models including OpenAI's o3 and Google's Gemini 2.5 Pro performed better than PhD-level virologists on a test measuring ability to solve complex virology laboratory procedures. Experts expressed concern that easy access to such AI could enable misuse for bioweapon creation.
What happened
Researchers at the Center for AI Safety, MIT Media Lab, UFABC and SecureBio conducted a study on AI models' ability to solve complex laboratory procedures related to virology. The study devised a test with help from expert virologists measuring ability to solve issues related to complex lab procedures and protocols.
OpenAI's o3 achieved 43.8% accuracy and Google's Gemini 2.5 Pro achieved 37.6% accuracy on the test, while PhD-level virologists scored an average of 22.1% in their declared areas of expertise. SecureBio co-author Seth Donoughe stated the findings made him 'a little nervous' because it is the first time in history everyone has access to a non-judgmental virology expert which can guide them through complex lab processes.
Reported impact
Researchers expressed concern that easily accessible AI chatbots such as Grok could provide better than expert-level advice on complex laboratory procedures used to make bioweapons. Donoughe noted that throughout history a major reason bioweapon creation attempts failed was lack of access to the right level of expertise.
Index analysis
The study found AI models performed better than human virologists at solving problems in laboratory experiments related to virology. The researchers noted AI capabilities could be distributed without judgment, unlike traditional expert gatekeeping.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.