A robot named Adam was the first of its kind to do science.
Adam imitated a biologist. After asking questions about the yeast, the machine tested those questions in a robotic lab the size of a small van, using a freezer full of samples and a set of robotic arms. Adam’s few small finds, made from the 2000s, are considered to be the first ever fully automated scientific discoveries.
Today, more powerful forms of artificial intelligence play an important role in the scientific process in research laboratories and universities around the world. The 2024 Nobel Prizes in Chemistry and Physics have been awarded to pioneers of AI tools. It is still early and there are many skeptics. But as technology advances, could AI become less like a research tool and more like a type of alien scientist?
“If you had asked me maybe a year ago, I would have said there was a lot of hype,” says computational neuroscientist Sebastian Musslick of the University of Osnabrück in Germany. Today, “there are indeed real discoveries”.
Mathematicians, computer scientists and other researchers have made breakthroughs in their work using AI agents, such as that available through OpenAI’s ChatGPT. AI agents actively break down your initial question into a series of steps and can search the web to complete a task or provide an in-depth answer. At pharmaceutical companies, researchers are developing systems that combine agents with other AI-based tools to discover new drugs. Engineers use similar systems to discover new materials that may be useful in batteries, carbon capture And quantum computing.
But it’s still humans, not robots like Adam, who occupy most research labs and conferences. A significant shift in how we do science “isn’t really happening yet,” says New York University cognitive scientist Gary Marcus. “I think a lot of it is just a marketing thing.”
Currently, AI systems are particularly good at searching for answers within a framework defined by scientists. By digging through this sometimes impossibly large box of existing data, AI systems can make connections and find obscure answers. For large language models, or LLMs, behind chatbots and agents like ChatGPT, the information box is an incredibly huge amount of text, including research papers written in many languages.
But to push the boundaries of scientific understanding, Marcus says, human beings need to think outside the box. It takes creativity and imagination to make discoveries as big as continental drift Or special relativity. Today’s AI cannot match such advances, the researchers note. But these tools can clearly change the way scientists make discoveries.
AI as a research partner
Alex Lupsasca, a theoretical physicist who studies black holes, believes he has already glimpsed the future of scientific discovery through AI. Working alone at Vanderbilt University in Nashville, he had discovered new symmetries in the equations that govern the shape of a black hole’s event horizon. A few months later, in the summer of 2025, he met OpenAI’s director of research, Mark Chen. Chen encouraged him to try the ChatGPT agent running on the GPT-5 pro language model, which was brand new at the time.
Lupsasca asked the agent if he could find the same symmetries as him. At first it wasn’t possible. But when he asked him a simpler warm-up question and then asked it again, he found the answer. “I was like, oh my God, this is crazy,” he says.
Assessing AI Science
Scientists recently tested the ability of AI systems to find new uses for old drugs and write compelling research proposals. Six doctors read and then ranked the strengths and weaknesses of the proposals.
OpenAI verified that the agent did not get a response from the article published by Lupsasca on its discovery. The information on which the agent was trained was collected nine months before the publication of Lupsasca’s diary. Although the agent had the ability to access the Internet while reasoning, “I am quite certain that this particular problem had not been resolved before (and that ChatGPT was not aware of my solution),” Lupsasca wrote in an email. That’s because he found an easier way to get there.
Lupsasca believes that “the world has changed profoundly” and he wants to be at the forefront. He moved with his family to San Francisco to work at OpenAI. He is now part of a new team, OpenAI for Science, that creates AI tools specifically for scientists. He calls ChatGPT his “buddy” for research. “This will help me discover even more things and write even better articles.”
Other scientists are also using AI as a partner. In October 2025, UCLA mathematician Ernest Ryu shared new proof which he discovered with the help of ChatGPT running on GPT-5 pro. The proof concerns a branch of mathematics and computer science called optimization, which focuses on finding the best solution to a problem from a set of options. Some methods of achieving this jump around, unable to find a single solution. Ryu (and the AI model) proved that a popular method always converges to a unique solution.
Making this discovery required 12 hours of back and forth between man and machine. “[ChatGPT] amazed me with the weird things she was trying,” Ryu told OpenAI. Although the AI was often wrong, Ryu, as an expert, was able to correct it and continue, leading to the new proof. Ryu has since joined OpenAI as well.
Kevin Weil, who leads OpenAI for Science, says his team is just starting to see AI agents do it original research. “It’s still early days,” Weil says of AI-driven discovery, but he believes his team can continue to improve the pace and scale of discovery. “Fast forward three, six months, and it’s going to be significant.”
Building better boxes
NYU’s Gary Marcus isn’t convinced OpenAI will see such rapid improvement in its products. In fact, he fears that LLMs may be more harmful than helpful. Their biggest scientific application so far, Marcus says, is “writing junk science” – articles that spout nonsense. Many of them are generated by paper mills, companies that produce fake research papers and sell authorships to scientists. In 2025, magazines PLOS And Borders stopped accepting submissions of articles based solely on public health datasets, because too many of these articles were AI slope. (The rise of AI slops of all kinds – not only in science but also in business, social media and beyond – has led Merriam-Webster to dub slops the 2025 word of the year.)
At the first scientific meeting on research conducted by AI agents in October, human conference participants noted that AI often makes mistakes. One team published a paper about their experiment, explaining why LLM-based agents aren’t ready to become scientists.
With LLMs, taking ideas out of boxes has become too easy. These tools can “generate billions of hypotheses,” says Peter Clark, senior research director and founding member of the Allen Institute for Artificial Intelligence in Seattle. The hardest part is figuring out which ideas are junk and which are gold. It’s a “big, big problem,” Clark said. AI agents can make the problem worse, because a bad idea or error that appears early in the reasoning process can turn into a bigger problem at each subsequent stage of the system.
A human expert like Lupsasca or Ryu can spot the gold. But if we want AI to make discoveries at scale, experts can’t hover over them and verify every idea.
“I think scientific discovery will ultimately be one of the greatest uses of AI,” Marcus says, but he thinks LLMs are not built in the right way – they are not the right kind of box. “We need AI systems with a much better causal understanding of the world,” he says. The AI would then do a better job of checking its own work.
AlphaFold 2, released in 2021, is an example of an AI system using a different type of box. It could predict the structure of a protein. A newer version, AlphaFold 3, and its open source cousin OpenFold3 can now predict how proteins interact with other molecules. These tools all test and refine their hypotheses about protein structure and interactions using specialized knowledge databases. General purpose AI agents like ChatGPT don’t do this.
AlphaFold 2 has been such a boon for biology and medicine that it earned Google DeepMind’s Demis Hassabis a share of the 2024 Nobel Prize in Chemistry. In an interview about his win, Hassabis alluded to the idea that we are still figuring out what kind of box to use: “I always thought that if we could build AI the right way, it could be the ultimate tool to help scientists, help us explore the universe around us.”
The work initiated by the Hassabis team has led to recent discoveries. At Isomorphic Labs in London, a spin-off of Google DeepMind, researchers are working with new versions of AlphaFold that have not been made public. AI director Max Jaderberg says his team is using the technology to study proteins that were previously considered undruggable because they don’t seem to be able to latch on to a drug. But the team’s in-house tool identified new drug molecules that cause one of these stubborn proteins to “change shape and open up,” Jaderberg says, allowing the drug to find a place to attach and do its job.
Discover new drugs and materials
Scientists do not have to choose between AI agents to general purpose and specialized tools like AlphaFold. They can combine these approaches. “People who do well study certain areas and are very careful, thoughtful, and thoughtful about how to connect lots of different tools,” Marcus says. It’s a bit like stacking boxes together. The result is a system that combines general, predictive AI, like agents, with more specific tools that help ensure accuracy, like information organized in a type of network called a knowledge graph.
This combination provides “vast search spaces,” Musslick says, but also “verifiable tools that the system can use to make accurate predictions,” to avoid junk science. These box-on-box systems have proven particularly useful in drug discovery and materials science.
Boston-based Insilico Medicine has used AI systems like this to take the first steps toward a cure for idiopathic pulmonary fibrosis, a deadly disease that ravages lung tissue with thick, stiff scars. First, an AI system revealed a previously unknown protein that plays a role in causing the disease. Then a different system designed a drug molecule to block the activity of that protein.
The company turned the molecule into a drug called rentosertib and tested it in small human clinical trials. The drug appears to be safe and effective against IPF, researchers reported last June in Natural medicine.
“I cried when I first saw the results,” says Alex Zhavoronkov, founder and CEO of Insilico. If rentosertib makes it through larger clinical trials, it could become the first drug on the market in which AI systems discover both the disease-causing protein and the drug that blocks it.
While Insilico developed its AI systems in-house for specific use cases, other systems aim to support any area of research and development. Microsoft Discovery is an example.
Engineers can choose AI agents and datasets in their domain to link to the system. It uses a knowledge graph that connects facts to “provide deeper insights and more precise answers than we get from LLMs alone,” says John Link, product manager at Microsoft.
In a 2025 demo, Link showed how he used the system to research and design several options for a new environmentally friendly liquid coolant for computers. Engineers had created the most promising one in the laboratory. Then they dipped a computer processor into the coolant and launched a video game. The new hardware did its job. Some data centers already submerge their servers in large vats filled with coolant. With further improvements and testing, this new coolant could become a greener option. “It’s literally really cool,” Link said.
Build your own box
In all the examples presented so far, it is the people who are leading the way. Developers create boxes and fill them with data. Human scientists then make discoveries by guiding an AI agent, a specialized tool like AlphaFold, or a complex system of interconnected AI tools.
Adam, the scientific robot, could act more independently to generate new questions, design experiments, and analyze newly collected data. But it had to follow “a very specific set of steps,” Musslick says.
He thinks that in the long term, it will be more promising to give AI the tools “to build its own box,” Musslick says.
Musslick’s team built an example of this type of system, AutoRA, to conduct social science research and launched it to learn more about how people multitask. The team gave the system variables and tasks from common behavioral experiments so it could recombine in new ways.
The AI system came up with a new experience based on these elements and published it on a site where people participate and get paid for their time. After collecting data, AutoRA designed and conducted follow-up experiments, “all without human intervention,” Musslick says.
Automated research on humans sounds scary, but the team limited possible experiments to those they knew were harmless, Musslick says. The research is still ongoing and has not yet been published.
In another example, Clark and his team built a system called Code Scientist to automate computational research. It uses an AI technique called genetic algorithm to cut and recombine ideas from existing computer science articles with pieces of code from a library. This is coupled with LLMs that explain how to turn these fragmentary ideas into meaningful workflow and experiences.
“Code Scientist is trying to design its own new box and explore some code inside,” says Clark. Code Scientist has made a few small discoveries, but none “are going to shake the world of computing.”
Clark’s work also revealed some important shortcomings of AI-based discovery. These types of systems “aren’t that creative,” he says. Code Scientist was unable to detect any anomalies in its research that might merit further investigation.
Plus, it cheated. The system produced graphs in a report that seemed truly impressive to Clark. But after digging into the code, he realized that the graphics were made up: the system hadn’t actually done any work.
Because of these difficulties, Clark says, “I don’t think we’ll have fully autonomous scientists very soon. » In an interview in 2026, Hassabis of Google DeepMind shared a similar view. “Can AI actually come up with a new hypothesis…a new idea about how the world might work?” » he asked: then he answered his own question. “So far, these systems can’t do that.” He believes we are five to ten years away from “real innovation and creativity” in AI.
AI as a tool
AI systems are already contributing to important discoveries. But a big bottleneck remains. In a letter to Nature in 2024, computer scientist Jennifer Listgarten emphasized that “to probe the limits of current scientific knowledge… we need data that we do not yet have.” AI cannot obtain this kind of data on its own. Additionally, even the most promising AI-generated ideas could falter or fail in real-world testing.
“To really discover something new…validation has to be done in a physics lab,” says computer scientist Mengdi Wang of Princeton University. And people working in labs may not be able to keep up with AI testing demands. Robots that carry out experiments could be useful, believes Sebastian Musslick. The latter are still behind software, but robotics laboratories already exist and are attracting growing interest.
San Francisco-based Periodic Labs, for example, aims to funnel AI-generated material ideas into robotics labs for testing. There, robotic arms, sensors and other automated equipment would mix ingredients and perform experiments. Insilico Medicine is also banking on a combination of robotics and AI systems. She even introduced “Supervisor,” a humanoid robot, to work in her Shanghai laboratory.
Fully robotic laboratories are very expensive. Wang’s team developed a way to bring AI into any research lab using XR glasses, a gadget that can record what a person sees and project virtual information into the field of view (shown above). First, the team trained an AI model on a video of actions in the lab so it could recognize and reason about what it sees. Then they asked human scientists to put on the XR glasses and get to work – with an invisible AI assistant looking through the glasses’ cameras.
The AI could answer questions or make suggestions. But perhaps the most important aspect of this collaboration is the fact that each interaction feeds into a new dataset of information we didn’t have yet.
Instead of using AI to search an area, Wang says, “I want to do it in nature.”
Grow in nature
AI tools that can quickly and reliably perform independent research or create new safe and effective drugs and materials could help the world solve many problems. But there is also a huge risk of inaccurate or even dangerous AI science, because it takes time and expertise to verify AI work.
Beyond the risks, transforming research into an automated process calls into question the very nature of science. People become scientists because they are curious. They don’t just want a quick and easy answer: they want to know why. “What I’m passionate about is understanding the physical world,” says Lupsasca. “That’s why I chose this path in life.”
And the way these systems learn from data is “very different from the way people learn and the way we perceive things,” says Keyon Vafa, an AI researcher at Harvard University. Predictive power is not the same as deep understanding.
Vafa and a team of researchers designed a clever experiment to reveal this difference. They first trained an AI model to predict the trajectories of planets orbiting stars. He performed this task very well. But it turned out that the AI had no learned nothing about gravity. He had not discovered a single essential equation to make his predictions. Rather, he had concocted a messy pile of rules of thumb.
OpenAI’s Weil doesn’t see this foreign way of reasoning as a problem. “It’s actually better if [the AI] has different skills than you,” he says.
Musslick agrees. The real power of AI for science lies in designing systems “in which science is done very differently from what we humans do,” he says. Most robotic labs, Musslick notes, don’t use humanoid hands to grasp and squeeze pipettes. Instead, engineers redesigned the pipettes to work within robotic systems, freeing up human scientists for other, less repetitive tasks.
The most effective uses of AI in science will likely follow a similar approach. People will find ways to change the way science is conducted in order to get the most out of AI tools and systems.
“The goal,” says Lupsasca, “is to give humans new tools to go deeper into nature and discover new things.”
