The price of intelligence

For years, artificial intelligence has been sold to us as the tool that would make science faster, cheaper, and, above all, more productive. Much of this is true: AI already helps researchers with literature reviews, writing code, and analyzing genomic data. But the picture is more complex than press releases suggest. James Zou, a data biologist at Stanford,tells Nature in that he has spent "well over $100,000" on AI subscriptions in the past year. At his university, this figure is roughly equivalent to the cost of maintaining a postdoctoral researcher.The commitment of major private players to science is becoming increasingly explicit. In October 2025, Anthropic launched Claude for Life Sciences, a version of its model geared towards biomedical research, with connections to platforms like Benchling, PubMed, and 10x Genomics, and collaborations with pharmaceutical companies such as Sanofi, Novo Nordisk, and AstraZeneca. In April 2026, OpenAI responded with GPT-Rosalind (a tribute to Rosalind Franklin, discoverer of the DNA structure). Google DeepMind, for its part, has deployed AI co-scientist in collaboration with the U.S. Department of Energy, and it is already credited with experimentally validated hypotheses on liver fibrosis and antimicrobial resistance.The results are real. But inStanford HAI's annual report presents two data points that coexist uncomfortably: on the one hand, the number of scientific publications mentioning AI has multiplied by almost thirty between 2010 and 2025; on the other, humans still outperform the best AI agents in complex tasks where reasoning and originality are key. In fact, this increase in productivity declared to be associated with the use of AI is obligatorily linked to a layer of human curation, verification, and correction that rarely appears in headlines. Matteo Niccoli, a geoscientist cited in the same article in Nature, says it bluntly: the bottleneck is not the tool, it is "the thinking and the discussion" around it. One must know when the model drifts, when it hallucinates, when it has lost context. It is useful, yes, but it is not exactly a labor-saving device.And when the work is not saved, the price, on the other hand, does go up. GitHub Copilot announced at the end of April that it was moving from a fixed subscription to pay-as-you-go billing. And a recent commentary in Nature recalls that in 2025, Google, Amazon, Microsoft, and Meta spent $380 billion on AI, with packages of up to $250 million for individual researchers. If the science of the future is built on these infrastructures, it also inherits their inequalities.The question is not whether AI is useful for doing science. It is. The question is who can afford it, who reviews its work, and who is left out when the bill comes.