Recursive automates the scientific method for AI

Published on May 16, 2026 | Translated from Spanish

Startup Recursive, led by Professor Tim Rocktäschel from UCL, has unveiled a platform that aims to automate the scientific method in artificial intelligence research. Its goal is to transform large-scale calculations into open discoveries, allowing systems to safely experiment with their own self-improvement.

A digital laboratory where an AI system, represented as a robot with mechanical arms, manipulates test tubes and floating mathematical equations, while a human scientist observes from a monitor displaying self-improvement graphs.

How controlled self-experimentation works 🤖

Recursive's system uses a feedback loop where AI models design hypotheses, run experiments, and analyze results without direct human intervention. Rocktäschel has implemented safeguards to prevent unwanted behaviors during self-improvement, prioritizing reproducibility and transparency. Results are stored in open repositories for third-party validation.

The robot scientist that never takes a vacation ☕

Now researchers can let an AI do the fieldwork while they enjoy their coffee. The machine formulates hypotheses, tests them, and publishes results without complaining about salary or asking for a day off. It just needs to learn how to write papers with academic jargon so no one notices the difference at conferences.