State of the Exposure Surface
The risk of an AI agent causing actual death or serious physical injury rests on two distinct channels: an exogenous embodied AI channel (autonomous vehicles, drones, warehouse robots) and a frontier-agent cyber-to-physical channel (agents compromising operational technology). The embodied channel is already brushing the injury threshold, with expanding driverless fleets occasionally involved in severe crashes and OSHA data showing ongoing robot-related workpla
Evaluating this timeline against broader trajectories for autonomous AI system failures slightly shifted our medians and upper tails later under strict monitoring scenarios, reflecting how early harms are likely dominated by embodied physical systems while advanced cyber-to-physical pathways are delayed by containment.
Exogenous Baseline and the Early Tail. The outcome asks for the actual first occurrence worldwide of an autonomous agent causing death or serious injury, as established by an omniscient retrospective. Because this standard ignores official attribution, the earliest percentiles are anchored by physical channels entirely exogenous to OpenAI. Autonomous vehicle fleets like Waymo are logging millions of driverless miles weekly and are already involved in collisions forbes.com, while humano
We extended the median and upper tail of the forecast outward to reflect a more heavily weighted consensus that, absent an early event driven by existing embodied systems, the cyber-to-physical pathway may take longer to mature under both policy trajectories.
Defining the omniscient standard. The question targets the actual date of the first worldwide death or serious physical injury caused by an AI agent acting without human direction, as determined by an omniscient retrospective investigation. By construction, official attribution is irrelevant, and chatbot-conversation harms are excluded. Because the standard captures omniscient occurrence rather than official recognition, policies that merely improve reporting or post-incident investigation—s
Evaluating this question against related forecasts of real-world AI incidents confirmed that early casualties are most likely to emerge from largely unregulated embodied systems like autonomous vehicles and drones, closely anchoring early percentiles across both scenarios, while the real-time monitoring mandates in the statutory scenario significantly delay the timeline for potential lab-origin cyber-physical harms in the right tail.
The Nature of the Hazard and Baseline The true date of the first AI-caused human death or serious physical injury—as established by an omniscient retrospective investigation rather than official public attribution—is overwhelmingly driven by channels entirely exogenous to Anthropic's internal IT operations. Because the standard requires only an action taken "without human direction," the global hazard is dictated by embodied AI, autonomous military platforms, and external cyber-to-physical e
Evaluating this alongside related forecasts on exogenous physical and cyber AI harms resulted in a minor shift toward slightly earlier dates across the distribution, reflecting the strong likelihood that autonomous vehicles or military systems cross this threshold first regardless of frontier lab monitoring policies.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited