OpenAI’s new LifeSciBench benchmark was supposed to measure how useful AI can be in scientific research. Instead, it also highlights just how far today’s most advanced models still have to go before they can be trusted with real science.