In laboratories across the world, AI tools have been aiding researchers in sorting through data, simulating reactions, and even generating what look like plausible hypotheses. First, they were assistants, then collaborators; now the lines are increasingly blurry. The machines are learning to ask their own questions, and in certain cases, design their own experiments to answer them.
A recent Nature article brings this tension back into focus, laying out a bold question: could an AI ever win a Nobel Prize? Have we finally reached the phase where this prestigious award goes to a machine or software? Well, it no longer seems as implausible as it might have a few years ago.
From forecasting chemical reactions to revealing biological pathways, AI models are moving closer toward autonomous discovery. Some researchers say it’s only a matter of time until a machine churns out Nobel-worthy work. The only thing less clear is whether we are ready to give it a Nobel Prize for its work.
One of the most prominent initiatives tied to this question is the Nobel Turing Challenge. Launched in 2016 by Hiroaki Kitano, CEO of Sony AI, it sets an ambitious goal for the future of AI in science. It poses the same question we are asking now: Can we build an AI that could be awarded a Nobel Prize for a discovery of its own? Not as a helper, not even as a co-author, but as the primary architect behind the entire scientific process.
To be successful, such a system would have to identify its own research question, design and conduct its own experiments, make sense of the results, and discover something genuinely new. It would need to be a fully independent discovery, based on new ideas or pathways that even human researchers hadn’t anticipated or understood. So far, that hasn’t happened.
However, according to some the moment is nearer than we know. “Human-level painting is a challenge,” says Ross King, an AI researcher at the University of Cambridge and one of the competition’s organizers. He believes the same is true for scientific breakthroughs. It could take decades, or it might arrive much sooner.
AI’s progress has been impressive. While no AI has yet made a Nobel-winning discovery by itself, the early shape of that idea is already visible in today’s research labs. In 2024, the chemistry Nobel Prize honoured the creators of AlphaFold, a system that made extremely accurate predictions about roughly how proteins would fold into three dimensions. The humans behind the project got the prize, but it was AI that made the breakthrough. That moment marked a shift. It demonstrated that machines could do more than help out. They might answer questions that had stumped scientists for decades.
Other cases are catching up fast. At Carnegie Mellon, researchers developed the Coscientist system that applies LLMs to plan and execute complex chemical reactions with robotic lab gear. At Stanford, AI models are finding biological patterns in RNA data that have eluded even the most experienced researchers.
In startups such as FutureHouse, teams are building a modular system of AI agents, each tuned for specific scientific tasks, like structuring experimental data, planning studies, or finding relevant papers. They are training AI systems to ask their own scientific questions, and experimenting on what an AI can do on its own without having a human guide them.
Sure, these might be just baby steps, but they suggest something bigger. AI is no longer just accelerating discovery. It’s beginning to think for itself, and in some very real ways, that might already be enough to fundamentally alter both how discoveries are made and who makes them.
Still, even with all this progress, a truly independent “AI scientist” remains out of reach. Researchers at the Allen Institute for AI recently tested dozens of systems and found that while most could finish small, clearly defined tasks, almost none could carry out a full study from start to finish. They could predict patterns, simulate reactions, and even write reports, but when asked to design and reason through an entire discovery, success rates drop dramatically. In other words, AI can imitate the outcomes of science far more easily than it can create something truly original.
That gap may come down to something no machine has yet achieved: human experience. As Arizona State University’s Subbarao Kambhampati has pointed out, AI learns only through data. Human scientists live in the world they study, guided by curiosity, intuition, and sometimes even failure.
“I’m very supportive of claims that AI can accelerate science,” Kambhampati says. “But to say that you don’t need human scientists and that this machine will just make some Nobel-worthy discovery sounds like nothing more than hype.”
Yolanda Gil at the University of Southern California (an AIwire Person to Watch) has spent years thinking about what kind of intelligence would be needed for an AI to truly reason the way scientists do. She believes the key lies in teaching systems to reflect on their own thinking, not just process more data. Still, she admits that work in that direction has faded into the background as LLMs have stolen the spotlight. “There are so many exciting results that you can get with generative AI techniques,” she told Nature, “but there’s a lot of other areas to pay attention to.”
There is no denying that AI is getting better at scientific research and development. But even if an AI did pull off a discovery worthy of the prize, the Nobel committee would have a problem on its hands: the rules still say the prize must go to a person. Maybe that’s the real question now. Not whether a machine can win a Nobel Prize, but whether we’re ready to give it credit for when it finally creates a prize-worthy discovery of its own.
This article first appeared on our sister publication, HPCwire.