GPT-5 Cracks a Three-Year Immunology Puzzle, AI Emerges as a Research Partner

For three years, immunologist Derya Unutmaz had been circling the same stubborn question. His lab studies T cells, the immune system's frontline soldiers, and a particular behavior in their data refused to add up no matter how the team approached it. According to OpenAI, the breakthrough did not come from a new experiment or a fresh dataset, but from a conversation. Working with GPT-5 Pro, Unutmaz laid out the problem the way he might to a knowledgeable colleague, and the model offered a mechanistic explanation that finally made the pieces fit. What had resisted years of human reasoning yielded, in part, to an exchange with a machine.

What makes the episode notable is less the answer itself than the nature of the help. This was not a case of retrieving a buried citation or summarizing a stack of papers. GPT-5 reasoned across competing biological possibilities, weighed which mechanism best matched the observed data, and proposed a hypothesis specific enough to be tested. That is the kind of cognitive work scientists have long assumed belonged exclusively to trained experts. The model functioned less like a search engine and more like a sparring partner capable of holding a complex problem in mind and pushing back on it.

The shift matters because hypothesis generation has always been the hardest part of science to automate. Machines can crunch numbers and flag patterns, but framing the right question and imagining the underlying cause has stayed firmly human. A growing number of researchers now report using frontier models this way, not to replace their judgment but to expand the space of ideas they can explore quickly. When a model can suggest a plausible mechanism in minutes, a scientist can spend their time deciding which suggestions are worth chasing in the lab, compressing a cycle that once took months.

None of this means the model replaces the experiment. Unutmaz still has to validate the hypothesis at the bench, and the value of an AI suggestion ultimately rests on whether it survives contact with reality. But the case offers a concrete glimpse of where the technology is heading. As these systems grow more capable of structured reasoning, their role in research looks less like a faster reference desk and more like a tireless collaborator, one that can be wrong, but that is increasingly worth arguing with. }