The original experiment compared T cells exposed early in development to either low levels of glucose or deoxyglucose, a glucose-like compound that disrupts a cell’s ability to use glucose. Researchers expected the two environments to produce similar outcomes because both limited glucose availability. Instead, the results diverged sharply.
Cells exposed to deoxyglucose overwhelmingly developed into inflammatory-response cells, while low-glucose conditions produced far fewer of those cells. The effect also persisted after deoxyglucose was removed, suggesting that reduced energy availability alone could not explain the outcome. Unable to identify the underlying mechanism, the research team eventually set the project aside.
After GPT-5 Pro became available in late 2025, Unutmaz uploaded the experimental results and asked the model to analyze them. According to Unutmaz, the model proposed that deoxyglucose was disrupting production of IL-2, a protein that can prevent T cells from developing into an inflammatory-response cell known as Th17. Under that explanation, deoxyglucose effectively removes a constraint on Th17 development, providing a possible reason the two experimental conditions produced different outcomes.
“GPT-5 came up with this really remarkable insight that retrospectively, makes perfect sense,” Unutmaz said.
The experience prompted him to test the model in another context. Using an experiment involving CD8+ T cells that target a type of lymphoma, Unutmaz asked GPT-5 Pro to simulate the outcome. His laboratory results had shown that the cells exhibited improved ability to kill lymphoma cells. According to Unutmaz, the model independently predicted the same effect, despite the findings not yet being published.
“That was the moment that I felt like, okay, these models have now come to a point where they really, truly understand,” he said.
Unutmaz said he increasingly views advanced AI systems as research collaborators rather than simple tools. In his view, the technology can help scientists process large volumes of newly published research, identify unanswered questions, and refine experimental strategies before committing resources in the laboratory.
“The number of things you can do to address your hypothesis is vast,” Unutmaz said. “You have countless approaches, and you don’t know which one will be the best strategy.”
He uses GPT-5 Pro to simulate experiments and evaluate possible outcomes, a process he says can reduce the time required to identify promising research directions. Even so, he emphasized that scientific expertise remains essential because researchers must determine whether AI-generated insights are meaningful and scientifically plausible.
The growing ability of AI systems to generate scientific insights also raises safety considerations. While the technology could accelerate progress in biology and medicine, similar capabilities could potentially be misused by actors seeking to develop biological or chemical threats. OpenAI said its Preparedness Framework is intended to monitor those risks and establish safeguards around capabilities that could cause severe harm.
Unutmaz remains optimistic about the technology’s future role in science. More recently, he has been using tools including Codex and GPT-5.2 Deep Research to assemble large cancer mutation datasets and produce research materials focused on T cells, with the goal of advancing precision immunotherapy research.
Reflecting on the pace of change, Unutmaz said he feels fortunate to contribute during a period of rapid scientific and technological development.
“To not only be able to witness it historically but participate a little bit, I feel truly lucky and privileged to do that.”
This analysis is based on reporting from OpenAI.
Image courtesy of Akadeum Life Sciences.
This article was generated with AI assistance and reviewed for accuracy and quality.
