When an AI system first causes a human death

My median prediction for when an AI system will first directly cause a human death is 13th August 2029, driven by the compounding hazard of autonomous agent capabilities overcoming current physical-access constraints and oversight mechanisms.

Resolves to the date of the first human death directly caused by an AI system acting without human approval of the specific action.

An AI system acting on its own means software that plans and carries out a sequence of steps itself, using tools, code, or internet access, where no human checked or approved the particular action that caused the death. A human setting the overall goal or approving the deployment does not disqualify it. A human approving the specific action does.

Directly caused means the death would not have happened had the agent not acted. Contributing to it, or missing a warning, is not enough.

The agent must be the one that acts. A chatbot that tells a user how to kill someone does not count, even if the user then does it. But an agent that arranges the death through a third party, for instance by hiring or paying someone, does count, as does an agent that procures the means itself.

Excluded: self-driving cars, robots, and military autonomous weapons. Suicide does not count. Deliberate misuse of an agent counts, provided the agent executed the steps.