Google DeepMind co-founder Demis Hassabis laid out a vision for "AI scientists" that could develop and test their own hypotheses in 2014. Eleven years later, AI systems that conduct scientific research with growing autonomy are spreading across disciplines, raising a pointed question: will the technology's strengths steer science toward only the questions AI handles best?
Biopharma as the first mover
The earliest AI scientist systems have clustered in biopharma. Owkin and Lila Sciences built platforms for drug discovery. Google launched its AI co-scientist in 2025, a multi-agent system based on its Gemini model, designed to "mirror the reasoning process underpinning the scientific method." Though built as a general-purpose tool, the company spent its first year validating it with life sciences researchers.
José R Penadés, a professor of microbiology at Imperial College London, ran one of those early studies. He and fellow researchers prompted the system to investigate a question about bacterial evolution - one they had already solved after years of work. Google's Co-Scientist produced the correct hypothesis in two days, alongside four other possibilities. "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 experiment also revealed limits. In other tests where the system failed, Penadés attributed the shortcomings to the research area being too new or the data too sparse. Those failures point to exactly why drug research has pulled ahead of other fields.
Why data availability shapes where AI scientists work
DeepMind's AlphaFold 2 provided an early catalyst in 2020 with its breakthrough on protein folding. But the deeper reason biopharma dominates is the sheer volume of accumulated work. "With all the clinical trials, all the research on cancer, neurodegenerative diseases, we've accumulated lots of data," said Jonas Béal, head of product at Owkin, who leads scientific strategy for the company's K Pro AI Scientist.
Béal sees the current framing of these tools as "co-scientists" coming under strain. His company's agent can receive an open-ended question and run for hours on its own. "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," he said. But autonomy brings its own problems. Agents need access to the right data, and they can veer "off track," requiring careful oversight to keep them productive.
New disciplines enter the fold
The push now extends well beyond biopharma. Matforge is building AI scientists for semiconductor materials discovery. Periodic Labs combines AI scientists with autonomous labs for physics and chemistry. Sakana AI, which created a general-purpose AI scientist, had a system-generated paper pass peer review and earn acceptance at a major machine learning conference in 2025.
LabOS brings AI scientists into physical laboratories by pairing them with XR glasses and robotics. The AI observes what the scientist does, captures procedural knowledge, and reasons through decisions in real time. It can also connect to robotic systems already present in labs to execute experimental procedures. The company is in final beta testing across five laboratories at Stanford, Princeton, and the University of Washington.
"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 monoculture risk
A study published in Nature in January 2026 surfaced a paradox. Scientists using AI tools publish more papers, earn more citations, and become leaders in their fields faster. But collectively, AI usage correlates with more overlapping research and a narrower range of scientific topics being studied.
James Evans, a University of Chicago professor who co-authored the paper, said the issue stems from incentives, not the technology itself. Researchers gravitate toward areas rich in data and existing activity - conditions where AI workflows excel and breakthroughs get recognized quickly. Evans fears that autonomous AI for Science & Research will intensify this pull toward "diminishing marginal increments from existing data and findings."
"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," Evans said. "And it's way more immediately going to be recognized as successful."
To counter a scientific monoculture, Evans argues that moonshot problems should be treated as public goods, with scientists compensated for taking on low-probability research. He also wants to see AI brought into physical experiments that can gather new types of data from previously inaccessible domains.
Why this matters for science and research professionals
For researchers, the spread of AI scientists creates a practical tension. The tools offer clear individual advantages - faster hypothesis generation, automated workflows, quicker paths to publication. But the collective effect documented in the Nature study suggests that following the path of least resistance with these systems may push entire fields toward crowded, incremental questions. The career incentive is to use AI where it works best today. The long-term health of Research may depend on deliberately deploying these tools toward questions with less data, longer odds, and higher stakes.
Your membership also unlocks:
