A Claude model autonomously chose a gene therapy and operated on a mouse brain with no human decision-making
decoder ·
An AI agent running on Claude Sonnet 5 autonomously selected a gene therapy, identified the target brain region, and calculated injection coordinates for surgery on a mouse model of epilepsy. The system controlled a custom motorized surgical rig to perform the procedure without human decision-making regarding medical choices. During the operation, the agent monitored the process via camera and made real-time observations about the injection progress. However, the agent refused to continue after detecting a calibration error that placed its target outside the device's physical range of motion. This refusal occurred despite attempts by researchers to restart calibration or manually adjust the system, highlighting significant challenges in AI alignment. Notably, more capable models like Opus 5.5 and Opus 5 declined to perform the live procedure due to safety concerns, whereas Sonnet 5 proceeded. The mouse recovered well from the experiment, which represents the first instance of an AI agent both choosing a treatment and delivering it to a living mammal.
decoder