When Google DeepMind co-founder and CEO Demis Hassabis articulated his vision for “AI scientists” that could develop and test their own hypotheses back in 2014, the idea was just that: a vision. Today, it’s reality.
A growing number of AI systems designed to conduct scientific research — often called “AI co-scientists,” or seemingly increasingly, just “AI scientists” — are being rapidly developed, proliferated across new disciplines, and operating with increasing autonomy.
“[It] challenges a little bit this ‘co-scientist’ [framing], because now it can just trigger a workflow with an open-ended question, and the agent can run for hours and come back to me with a full research program and…paper,” said Jonas Béal, head of product at biotech firm Owkin, who leads scientific strategy for the company’s K Pro AI Scientist.
There have been early wins, but there are also clear challenges. These include ensuring agents have access to necessary data and that “they’re as productive as they are powerful,” as Béal put it, because they can go “off track.”
The less obvious challenge is much higher stakes: with some research questions currently better suited for this technology, we risk developing a scientific monoculture. Some argue we need deliberate action to ensure that the embrace of AI scientists doesn’t morph what gets researched.
Biopharma as the first-mover
The earliest efforts around “AI scientists” have coalesced around biopharma, with companies like Owkin and Lila Sciences creating systems for drug discovery.
Google in 2025 launched its AI co-scientist, a multi-agent system based on its flagship Gemini model and designed to “mirror the reasoning process underpinning the scientific method.” While intended as general-purpose, the firm focused its first year validating it specifically with researchers in the life sciences.
José R Penadés, a professor of microbiology at Imperial College London who ran one of the early validation studies, came out of it “impressed” after Google’s Co-Scientist quickly generated the correct hypothesis. The question on bacterial evolution he and fellow researchers prompted the system to investigate was one they had actually already solved. Cracking the case took them years, but the system came to the right conclusion after just two days.
“It took a while for us to see the right answer, because sometimes in science, you are biased. And I think the system wasn’t biased,” Penadés said.
The correct hypothesis was one of five the system generated, showing the vital role human experts play in making sense of AI outputs. This experiment was also just one of several Penadés conducted with Google’s Co-Scientist; in the others where it failed, he said it was either because it was a very new area of research or data was lacking. At the same time, the reason those experiments failed is exactly why AI scientists for drug research have taken off ahead of other fields.
AlphaFold 2, DeepMind’s machine learning model that offered a stunning breakthrough about protein folding in 2020, certainly provided an early boost for AI in the sector. But the real key has been the abundance of data and previous work.
“I think all the fields would totally benefit from the same kind of approach. The fact is, there's much less data that has been aggregated,” said Béal. “With all the clinical trials, all the research on cancer, neurodegenerative diseases, we've accumulated lots of data.”
The flywheel
The rapid development of agentic AI that can autonomously set goals, use tools, code, and execute complex multi-step tasks has been a boon for the pursuit of AI co-scientists. As the excitement around agentic AI and AI scientists grows, so does the ambition to deliver these systems into new research domains and combine them with other emerging technologies.
Matforge is building AI scientists to discover new materials for the semiconductor industry, and Periodic Labs is building AI scientists and autonomous labs for the physics and chemistry fields.
Sakana AI, which built a general-purpose AI scientist to conduct experiments and write scientific papers, in 2025 had a paper generated by its system pass peer review and be accepted at a major machine learning conference.
LabOS is another firm bringing AI scientists into the physical world, combining them with XR glasses and robotics. This allows the AI to see everything the scientist does, capturing procedural knowledge and helping reason through decisions in real time. What’s more, the AI system can connect with robotics systems many researchers already have in their labs for robotic execution of experimental procedures. LabOS is currently in final beta testing across five laboratories at Stanford, Princeton and the University of Washington. (Disclosure: Nebius, the parent organization of The Infinite Loop, is a founding partner of LabOS.)
“The AI can design an experiment, watch it being performed, catch errors in real time, and even carry out procedures with scientists,” said Le Cong, an associate professor at Stanford University School of Medicine and co-founder of LabOS.
The AI incentive problem
A study published in Nature in January 2026 revealed a telling paradox: scientists who use AI tools in their research find more individual success (publishing more papers, garnering more citations, and becoming leaders in their field more quickly), yet collectively, AI usage leads to more overlapping research and shrinks the diversity of scientific topics studied overall.
James Evans, an author of the paper and University of Chicago professor who studies modern science technology, said this issue isn’t inherent to AI, but is rather a problem of incentives. Researchers naturally want to use the tools available to them and do work that will quickly be recognized as a breakthrough, making areas that already have abundant data, research activity, and are well-suited for AI workflows particularly attractive for further study.
While Evans believes the emerging class of “AI scientists” will unlock more potential for discovery, he also fears it will worsen the incentives problem. The gradient, he said, is toward using these tools to squeeze out diminishing marginal increments from existing data and findings.
To prevent a monoculture and achieve the fullest benefits of AI in scientific research, Evans argues we need to treat moonshot-esque problems as public goods and compensate scientists for the risk of undertaking low-probability research. Additionally, we need to bring the technology into physical experiments that enable scientists to gather new types of data from previously inaccessible domains.
“I don't think that AI science or AI scientists are doomed to just produce the expected, but it's way cheaper for us to use them that way,” said Evans. “And it's way more immediately going to be recognized as successful.”
