The UK government is putting fresh money behind a fast-moving idea in research: AI systems that can do more than assist scientists, and instead carry out large parts of the scientific workflow themselves.
Through ARIA, the Advanced Research and Invention Agency, 12 projects have been chosen for funding from a field of 245 proposals. The selected teams are building systems described as AI scientists, including robot biologists and chemists designed to plan, run, and interpret experiments in the lab.
What ARIA Means By An AI Scientist
ARIA defines an AI scientist as a system able to handle an entire research loop. That means forming hypotheses, designing experiments to test them, running those experiments, and analyzing what comes back.
In some cases, the results are not the end of the process. The system can feed what it learns back into itself and repeat the cycle, creating a loop in which each round of work informs the next.
The human role does not disappear in this model. Scientists still set the initial research questions and oversee the process. But the routine, repetitive, and time-consuming parts of laboratory work can shift toward automated systems.
“There are better uses for a PhD student than waiting around in a lab until 3 a.m. to make sure an experiment is run to the end,” says Ant Rowstron, ARIA’s chief technology officer.
That framing explains why the competition matters. The goal is not simply to attach AI to science as another analysis tool. It is to test whether AI can coordinate meaningful stretches of practical research, including work that normally depends on instruments, lab protocols, troubleshooting, and repeated execution.
Why The Competition Drew Attention
ARIA originally planned to spend less on the competition, but it doubled the amount because of the number and quality of submissions. The agency selected 12 projects from 245 proposals, a sign that many groups are already trying to automate substantial parts of lab work.
Each selected team will receive around £500,000 (around $675,000) for nine months of work. By the end of that period, the teams are expected to show that their AI scientist can produce novel findings.
The winners include a mix of university and industry teams. Half are from the UK, while the rest are from the US and Europe. That mix matters because the work is not confined to one type of institution or one narrow research field.
One selected team is Lila Sciences, a US company building what it calls an AI nano-scientist. Its system is focused on discovering ways to compose and process quantum dots, nanometer-scale semiconductor particles used in medical imaging, solar panels, and QLED TVs.
“We are using the funds and time to prove a point,” says Rafa Gómez-Bombarelli, chief science officer for physical sciences at Lila: “The grant lets us design a real AI robotics loop around a focused scientific problem, generate evidence that it works, and document the playbook so others can reproduce and extend it.”
Another funded team, from the University of Liverpool, UK, is working on a robot chemist. It runs several experiments at once and uses a vision language model to help identify and address problems when the robot makes an error.
A London startup, still in stealth mode, is developing ThetaWorld. That AI scientist uses LLMs to design experiments on physical and chemical interactions that matter for battery performance, with the experiments then carried out in an automated lab.
A Small Bet Designed To Measure A Big Shift
Compared with the £5 million projects over two or three years that ARIA usually funds, the £500,000 awards are deliberately modest. Rowstron describes the program as an experiment for ARIA as well as for the research teams.
By backing several projects for a short period, the agency is trying to understand where the frontier really is. The result will help establish a baseline for future larger-scale funding decisions.
That caution reflects a broader problem in AI research. Rowstron notes that there is considerable hype, especially now that leading AI companies have teams focused on science. When claims appear through press releases rather than peer review, it becomes harder to judge what these systems can actually do.
“That’s always a challenge for a research agency trying to fund the frontier,” he says. “To do things at the frontier, we've got to know what the frontier is.”
For now, the most advanced systems in this area are agentic systems that call on other tools as needed. Rowstron says they use large language models for ideation, other models for optimization, and automated methods to run experiments before feeding results back into the process.
He sees AI scientists as sitting above tools such as AlphaFold. Those tools can help scientists move faster through difficult parts of research, but they may still leave many months of lab work to verify results. The AI scientist concept aims to automate more of that remaining work.
The Limits Are Still Real
ARIA is not funding systems that can create entirely new scientific tools on their own. Rowstron says the current projects rely on existing tools rather than generating new ones as part of solving a research problem.
There are also reliability concerns. Agentic systems can struggle to stay on task when they run for long periods without human correction. That limits how far they can operate independently before errors start to compound.
A study titled “Why LLMs aren’t scientists yet,” posted online last week by researchers at Lossfunk, an AI lab based in India, reported that LLM agents failed three out of four times when asked to complete a scientific workflow. The researchers pointed to problems including “deviation from original research specifications toward simpler, more familiar solutions” and “overexcitement that declares success despite obvious failures.”
Rowstron is clear that the technology remains early. “Obviously, at the moment these tools are still fairly early in their cycle and these things might plateau,” he says. “I’m not expecting them to win a Nobel Prize.”
Still, ARIA’s bet is that the direction of travel may be important even before the systems are fully mature. If AI scientists make some forms of experimentation faster, research organizations will need to understand how to work with them, how to supervise them, and how to judge their results.
“But there is a world where some of these tools will force us to operate so much quicker,” he continues. “And if we end up in that world, it’s super important for us to be ready.”