AI Scientists Now Run Experiments Without Human Help

Scientists just got powerful new colleagues that never sleep, and they're already reshaping how research gets done.

Three groundbreaking AI systems described in Nature magazine can now perform tasks that once required entire research teams. ERA develops scientific software, The AI Scientist runs complete research projects from idea to manuscript, and MIRA makes medical decisions in simulated hospital environments.

ERA tackles one of science's biggest time drains: writing and fixing computer code. The system can inspect existing programs, spot problems, suggest improvements, run tests, and keep refining until the code works better. It already operates across fields from cell biology to disease forecasting, adapting to whatever scientific tools researchers need.

What makes ERA special is its feedback loop. The system doesn't just generate code and move on. It runs the program, checks if it performed well, and tries again if results fall short.

The AI Scientist takes things further by handling an entire research workflow. It proposes hypotheses, writes experimental code, runs simulations, creates charts, drafts papers, and even reviews its own work. This end-to-end approach mimics how human scientists work, though the quality still depends on the data and objectives humans provide.

Professor Guang-Guo Ying of South China Normal University, who analyzed these systems, notes they represent something fundamentally new. These aren't chatbots that answer questions or summarize existing papers. They're systems that take action, make decisions, and improve their own work through trial and error.

MIRA applies similar autonomy to medicine by working with simulated electronic health records. It orders tests, generates diagnoses, prescribes treatments, and decides whether patients need hospital admission. The system operates in safe simulations rather than with real patients, but it demonstrates how AI can handle complex decisions that unfold over time.

The Bright Side

The speed advantage is enormous. Tasks that once consumed weeks of a researcher's time can now happen in hours or days. In cell biology, where scientists must organize thousands of measurements from individual cells, AI agents can test far more analytical approaches than human teams. In disease forecasting, these systems can rapidly adjust models as new case data arrives.

This acceleration doesn't replace human scientists. Instead, it frees them from repetitive coding and data processing to focus on asking better questions and interpreting results. The AI handles the computational heavy lifting while researchers provide the scientific judgment that machines still lack.

Of course, speed brings risks. An AI might produce plausible-looking results that contain subtle errors in data handling or statistical assumptions. That's why Ying emphasizes that independent validation becomes more important, not less, as these systems grow more capable.

The breakthrough isn't just faster research, it's research that can explore more possibilities and adapt more quickly to what the data reveals.