In science, there is a silent, nearly undetectable moment when an idea that no one has yet considered is already concealed within the data. Beneath decades’ worth of published papers and citations, it waits patiently for someone to make the necessary connections. Only humans were able to connect in that way for the majority of human history. It might no longer be the case.
Researchers at the University of Chicago created an AI model that predicted scientific discoveries before human scientists did, which is still a little unsettling to comprehend. The study was published in Nature Human Behaviour. Not a guess. not an estimate. Predicted—with sufficient accuracy to identify the precise substances and, in certain situations, the researchers who are most likely to discover them.
| Category | Details |
|---|---|
| Key Researcher | Prof. James A. Evans |
| Title | Max Palevsky Professor, Department of Sociology |
| Institution | University of Chicago |
| Lab | Knowledge Lab |
| Published In | Nature Human Behaviour |
| Study Focus | AI predicting future scientific discoveries and generating “alien” hypotheses |
| Key Achievement | 400% improvement in predicting future discoveries vs. content-only models |
| Discovery Predicted | CsAgGa2Se4 as thermoelectric material — identified by AI before humans confirmed in 2012 |
| Reference Website | University of Chicago Knowledge Lab |
The director of the Knowledge Lab and the Max Palevsky Professor in the Department of Sociology, Prof. James A. Evans, has spent years considering how scientific knowledge spreads throughout human societies. His team’s goal went beyond simply creating a more intelligent research paper search engine.
They sought to comprehend the nature of human discovery itself, including its acceleration, stalling, and—most importantly—complete blindness. The end product is what Evans refers to as a “digital double” of the scientific system: a model of what the research community as a whole is probably going to do next and what it most likely won’t.
The mechanics are truly intriguing. The group created models that performed millions of random walks across scientific literature. These walks began with a property, such as COVID vaccination, and then jumped through related papers, shared authors, and cited materials, following threads in a manner similar to that of an inquisitive graduate student, but at a scale that was beyond the capabilities of the human mind.
In comparison to methods that solely examined research content, that process yielded a model that increased predictions of future discoveries by 400%. Because the model knew which scientists were most likely to make those discoveries due to their professional connections and expertise, it was even able to identify them with more than 40% accuracy.
One is particularly noteworthy. A material known as CsAgGa2Se4 was recognized by the AI as a potential thermoelectric material. That discovery wasn’t verified by scientists until 2012. The papers used by the AI were published prior to 2009. That’s a three-year advantage that remains undiscovered by any researcher who passes by it, sitting undisturbed inside the literature.
It’s difficult to ignore that for a little while. Scientists were working toward a discovery that a machine had, in a sense, already made somewhere in a lab that was most likely operating on too much coffee and insufficient funding. That is not a critique of science. It’s merely a striking example of how limited human attention is, even for the world’s most gifted researchers.
Evans takes care to avoid portraying this as AI outperforming scientists. He intentionally employs augmentation rather than replacement in his language. “If you build in awareness to what people are doing, you can improve prediction and leapfrog them to accelerate science,” he stated. However, the second part of that sentence is just as important. “You can also figure out what people can’t currently do, or won’t be able to do for decades or more into the future.” This is where things become somewhat bizarre and philosophically fascinating.
In a second experiment, the team essentially asked the model to predict human behavior in the opposite way. It looked for things that scientists most likely wouldn’t find next, not because those theories were incorrect but rather because there wasn’t a vibrant community of researchers in a position to do so. These were referred to as “alien” theories. They weren’t outlandish concepts either.
The researchers found that these extraterrestrial conclusions were, on average, more robust from a scientific standpoint than those that humans were pursuing. According to Evans, this is because human scientists typically explore every option within a well-known framework before venturing into truly uncharted territory. That isn’t always the best course of action for exploration, but it makes sense for a career.
The research contains a structural critique that is worth considering. According to Evans, the model implies that graduate education isn’t especially well-suited for scientific advancement as it stands. It is adjusted to generate workers with credentials. “They do not optimize discovery of new, technologically relevant things,” he said. “To do that, each student would be an experiment — crossing novel gaps in the landscape of expertise.” From a sociology professor at one of the most prestigious research universities in the world, that is a subtly radical statement.
This way of thinking relates to a larger phenomenon occurring in both technology companies and research institutions. Targeted therapy development and the identification of promising materials have already been aided by AI systems trained on published findings. However, the majority of those systems completely disregarded the human factor, including the scientists’ identities, backgrounds, and training. The fact that Evans’s model views the scientific community as a social system rather than merely an information archive proves to be crucial.
Perhaps the reframing Evans is suggesting is more important in this case than the precise predictions the model made. He challenges the Alan Turing-era notion that artificial intelligence should mimic human intelligence. He contends that no one can solve more difficult problems with that imitation game.
Creating systems that explicitly map the areas where human cognition tends to cluster and then move elsewhere could be beneficial. “It’s about changing the framing of AI from artificial intelligence to radically augmented intelligence,” he stated. It sounds like a subtle distinction. It isn’t.
As this field has grown over the last few years, there is a feeling that science is in a unique position—not because machines are going to take the place of researchers, but rather because the tools available to researchers are starting to change in kind rather than just degree. It’s not really about whether AI can forecast a discovery. It has already done so.
What scientists decide to do with that data and whether the organizations founded on human research are adaptable enough to use it are the more intriguing questions. At least that part is still totally up to us.
