Swedish researchers at Chalmers University of Technology have built an “AI scientist” that can generate hypotheses, design experiments, run them using lab robotics, and interpret the results—all in one automated loop. The system, detailed in a new study, combines multiple large language models (LLMs) with a database of ~60,000 biological relationships and physical laboratory robots. It is a concrete step toward fully autonomous scientific discovery.

This isn’t another LLM that writes plausible papers. It’s a closed-loop system that actually does wet-lab work. The researchers tested it on baker’s yeast (S. cerevisiae), but the architecture is general. If this scales, it could change how quickly we find new treatments, understand biology, and solve complex problems.

What Happened

The AI scientist works by first analyzing a massive database of phenotypical, physiological, and metabolic relations for yeast. Using that knowledge, it generates plausible biological questions, then recommends specific experiments to test them. The system then instructs a robot named Eve—a physical lab automation platform—to carry out the experiments. After the robot runs the protocol, the AI evaluates the outcomes and refines its own understanding based on the new evidence.

“It is too much information for a human to analyse, but our AI scientist could identify promising biological questions, recommend experiments to test them, evaluate experimental outcomes and iteratively refine its understanding based on new evidence,” said Ievgeniia Tiukova, a Chalmers researcher and co‑author of the study.

The team deliberately chose yeast because it is a well‑studied model organism with a rich dataset. But the approach is designed to be transferable to other organisms and domains. The key innovation is the tight integration of LLMs (which handle reasoning and hypothesis generation) with robotic hardware (which handles execution) and a feedback loop that closes the gap between thinking and doing.

Read the full announcement →

My Take

This is the real deal—not just a language model trained on papers, but a system that can physically interact with the world and learn from the outcomes. The most impressive part is the closed‑loop nature: the AI gets to see the results of its own experiments, not just a human‑curated summary. That is what separates this from earlier “AI researchers” that only produce text.

For developers, this means the era of autonomous scientific labs is beginning. The infrastructure—LLMs for reasoning, databases for priors, robots for execution—is already here. The bottleneck is no longer the technology but the integration and validation. We will likely see similar systems for drug discovery, materials science, and synthetic biology within a year.

But there are risks. An AI that runs experiments on its own could produce large volumes of low‑quality data if the hypothesis generation is flawed. Reproducibility will be harder to audit. And the system’s decisions are only as good as its priors—biased databases will produce biased science. Still, the potential to accelerate discovery by orders of magnitude outweighs the risks, provided we build guardrails.

What to Watch

  • Lab automation vendors like Opentrons and Automata will see huge demand as AI‑driven experiment design becomes standard.
  • Expect a wave of open‑source “AI scientist” frameworks that let research labs plug in their own databases and robots.
  • Ethical oversight boards will need to update guidelines for autonomous experimentation, especially when the system proposes experiments on human cells or pathogens.
  • The biggest short‑term impact will be in antibiotic discovery and cancer biology, where large phenotypic datasets already exist.