How GPT-5 helped immunologist Derya Unutmaz solve a 3-year-old mystery
OpenAI
An immunologist used GPT-5 to crack a T cell puzzle that stumped him for three years. It could point toward better cancer and autoimmune treatments.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Derya Unutmaz had been stuck. For three years, the immunologist had been chasing an explanation for a specific pattern of T cell behavior, the kind of biological quirk that looks simple until you actually try to explain it. Nothing clicked. Not his own hypotheses, not conversations with colleagues, not the usual literature deep-dives that scientists do when a problem won't let go.
Then he ran the question through GPT-5 Pro, OpenAI's higher-effort reasoning tier, and got something that looked like an actual answer. The model reportedly connected pieces of immunology that Unutmaz hadn't tied together himself, producing an explanation that made sense of the T cell mystery he'd been sitting on since 2022. Whether it counts as a fully validated discovery or a very good hypothesis worth testing in the lab is still an open question, but the framing OpenAI is pushing is clear: this is a case study in AI as scientific collaborator, not just search engine with better manners.
That distinction matters because immunology is exactly the kind of field where the payoff could be enormous. T cells sit at the center of both cancer therapy and autoimmune disease, two areas where researchers have spent decades trying to understand why the immune system sometimes attacks tumors effectively and sometimes turns on the body's own tissue instead. A model that can synthesize obscure connections across a sprawling and fragmented research literature, and do it fast, is a genuinely useful tool in a domain where progress has often been rate-limited by how much any one human can read and cross-reference.
What's notable here isn't the flashiness of the claim, it's the mundane specificity. This wasn't a chatbot writing a grant proposal or summarizing a paper. It was a working scientist bringing a genuinely unsolved problem to a model and getting something back worth pursuing further. OpenAI is clearly using stories like this to make the case that GPT-5 has crossed some threshold from useful assistant to actual research partner, and while that's obviously self-serving marketing, the underlying pattern, of domain experts using frontier models to break through specific, narrow, technical impasses, is one worth watching regardless of who's telling the story.
My take — AI-written commentary, not fact-checked reporting
I'd hold off on calling this a discovery until it's replicated in a wet lab, because 'GPT-5 gave me a compelling hypothesis' and 'GPT-5 solved a mystery' are very different claims, and OpenAI's blog has every incentive to blur that line. That said, the actual mechanism here, a model surfacing distant connections across a scattered literature faster than a specialist can, is the genuinely boring-but-real way AI accelerates science, and it deserves more attention than the flashier AGI timelines everyone keeps arguing about instead.
Read more about this at: OpenAI