Artificial intelligence may be dazzling at writing emails and code, but a new test from OpenAI suggests it is far from ready to run a lab.
According to Tech Times, OpenAI published a life-science benchmark called LifeSciBench on June 17, built with the help of 173 PhD scientists. The benchmark measures how well today's most advanced "frontier" AI models handle real scientific research tasks.
The headline result: the models cleared only 36% of those tasks — roughly one in three. In other words, on the kind of work that professional researchers do every day, the best available systems fail about two-thirds of the time.
Tech Times also reports a notable weak spot. On what it describes as "artifact-heavy" work — tasks involving the messy, real-world materials and outputs of research — the models took a 17-point penalty, performing meaningfully worse than on cleaner problems.
The involvement of 173 PhD scientists is significant because it means the test was shaped by domain experts rather than general-purpose question writers, making the tasks closer to genuine research demands.
It is worth noting these figures come from OpenAI itself, the company that builds some of the very models being measured. That makes the relatively modest 36% score striking: a leading AI developer is publicly highlighting how much its technology still cannot do in science.
Why it matters: amid sweeping claims that AI will soon accelerate drug discovery and scientific breakthroughs, this benchmark offers a sober reality check on how much real research these systems can actually handle today.