Jay Chooi explains why text-only safety evaluations and narrow robotics demonstrations cannot describe the full capabilities of physical AI. He discusses controlled doll-based safety tests and differences in model refusal behavior, while arguing for a wider set of tasks rather than a single definitive benchmark. The examples concern test scenarios, not actual injuries.
Jay Chooi proposes measuring tasks with consequences for manufacturing, supply chains, jobs and public understanding. His expectation of general-purpose robotics by 2028 is a forecast he attributes to researchers and industry contacts, not a demonstrated capability or a guaranteed deployment timeline.
Jay Chooi separates robotics benchmarking from his research on possible human disempowerment. He describes indicators such as the share of organizational spending devoted to human labor, and explores AI-mediated political coordination. These frameworks identify questions to measure; they do not establish that lost economic leverage or any particular political outcome is inevitable.
Watch on YouTube
