We hear a lot about AI writing emails and code. Far rarer — and far more interesting — is AI helping make a genuine scientific discovery. This week, OpenAI shared exactly that: how GPT-5 Pro helped immunologist Dr. Derya Unutmaz finally crack a puzzle his lab had been stuck on for three years.
The kicker: the model correctly predicted a result the lab had measured but never published — meaning it couldn't have just looked it up online. It had to reason its way there. Here's the story, and an honest look at what it does (and doesn't) prove.
What Happened
Unutmaz, a doctor and immunologist who has followed AI closely for years, revisited an old, unsolved experiment with GPT-5 Pro. When he asked the model to reason through it — and even simulate the experiment — it produced a specific, correct prediction about how certain immune cells would behave, along with a plausible explanation. For a long-stalled problem, it was an "aha" moment.
The 3-Year Mystery
The puzzle dates back to 2022. Unutmaz's lab was studying a basic but important question: how does glucose — a sugar — affect the way T cells develop and specialize?
T cells are frontline immune cells that help the body fight cancer and infections, so understanding what shapes them matters enormously. But the 2022 experiment produced results the team couldn't fully make sense of at the time — so the question quietly sat unresolved for years.
How GPT-5 Helped
When Unutmaz asked GPT-5 Pro to simulate the experiment, the model correctly predicted a striking outcome: a boost in CD8+ T cells' ability to kill lymphoma cells. Crucially, his lab had observed this but not yet published it — so the answer wasn't sitting on the internet for the model to retrieve.
It went further, proposing a concrete mechanism:
- That disrupted N-linked glycosylation during T-cell "priming" could explain the observations.
- That memory T cells — not naïve T cells — were the key population driving the effect.
These were specific, testable hypotheses that fit the lab's real data — the kind of insight you'd hope for from a sharp scientific collaborator. It's a more rigorous cousin of the AI-for-biology work we covered when AlphaFold's creator joined Anthropic.
Why This One Is Different
AI models are famous for summarizing what's already known. What makes this case stand out is that GPT-5 Pro appeared to reason to a novel, unpublished result rather than recall it. Predicting something that wasn't yet in any paper is a meaningful signal that today's frontier models — the same ones topping the 2026 model leaderboards — can do real hypothesis-driven thinking, not just retrieval.
That's the difference between a search engine and a collaborator.
The Human Still Mattered Most
It's important not to over-read this. This was AI augmenting a human expert, not replacing one. Unutmaz:
- Framed the problem and supplied the deep experimental context.
- Knew which questions were worth asking.
- Validated the model's predictions against his lab's actual results.
Without an expert to pose the right question and check the answer, a confident-sounding model is just that — confident-sounding. The magic here was the pairing: human judgment plus machine speed.
Why It Matters
- AI as a research copilot. Models that generate and test hypotheses could compress months of stuck research into days.
- Reasoning, not just recall. Predicting unpublished results hints at genuine scientific reasoning.
- Big fields stand to gain. Cancer, autoimmune disease and infections all hinge on exactly this kind of cellular puzzle.
- Keep the skepticism. One vendor-shared example isn't peer-reviewed proof — validation and replication still rule.
Frequently Asked Questions
What did GPT-5 actually do in this discovery?
Immunologist Derya Unutmaz asked GPT-5 Pro to reason about a stalled experiment on how glucose affects T cells. The model correctly predicted an outcome his lab had measured but not yet published — a boost in CD8+ immune cells' ability to kill lymphoma cells — and proposed a plausible mechanism. Because the result was unpublished, the model couldn't have simply looked it up; it reasoned to the answer.
What was the 3-year mystery?
Back in 2022, Unutmaz's lab ran an experiment on how glucose (a sugar) influences the way T cells — key immune cells that fight cancer and infections — develop and specialize. The results were puzzling and the team couldn't fully explain what they were seeing, so the question sat unresolved for about three years.
What mechanism did GPT-5 propose?
GPT-5 Pro suggested that disrupted N-linked glycosylation during the 'priming' of T cells could explain the observations, and that memory T cells — rather than naïve T cells — were the key population driving the effect. These were testable, specific hypotheses that matched the lab's data, not vague guesses.
Did the AI replace the scientists?
No. This was AI augmenting human expertise, not replacing it. Unutmaz framed the problem, supplied the experimental context, and his lab validated the AI's predictions against real data. The model acted like an exceptionally fast research collaborator that can generate and test hypotheses — but the human scientists set the questions and confirmed the results.
Why is this case significant?
Predicting an unpublished experimental result is a meaningful signal that advanced models can do genuine hypothesis-driven reasoning, not just summarize existing knowledge. If reliable, that could accelerate research in cancer, autoimmune disease and infectious disease by helping scientists form and prioritize hypotheses far faster than before.
Should these results be treated as proven?
Treat them as promising but still subject to normal scientific scrutiny. A single high-profile example, shared by the model's maker, isn't the same as peer-reviewed, independently replicated proof. The exciting part is the demonstrated potential; the responsible part is remembering that lab validation and peer review still decide what's actually true.
What does this mean for the future of research?
It points toward AI becoming a standard 'research copilot' — helping scientists design experiments, interpret confusing data, and generate testable ideas. The biggest gains may come not from AI making discoveries alone, but from pairing fast machine hypothesis-generation with rigorous human experimentation and judgment.
Final Thoughts
This is the kind of AI story worth getting excited about — not a chatbot trick, but a model helping a scientist see something he'd missed for three years. And the most encouraging part isn't that AI "did science" on its own. It's that a great researcher plus a powerful model together cracked something neither might have alone.
The honest caveats remain: it's one example, shared by the model's maker, and real proof still runs through the lab bench and peer review. But as a glimpse of how discovery might work in the AI era — fast hypotheses, expert validation — it's a genuinely hopeful one. We'll be watching where AI-for-science goes next.