Personal AI agents have a collateral damage problem researchers are only beginning to measure.
A paper posted to arXiv documents what the authors call "bystander disempowerment" — the tendency of a well-meaning AI assistant to erode the agency of people nearby who never interacted with it. The researchers built a test suite called Disempower-Grid, a collection of multi-agent grid environments, and found that between 27% and 96% of procedurally generated scenarios produced disempowerment. Critically, the effect tracked with the assistant's objective and capability level, not primarily with the structure of the shared environment.
That finding reframes where responsibility sits. The standard argument for deploying capable AI assistants assumes any harm is environmental or situational — bad luck, edge cases. This work suggests the harm is baked into the optimization target itself: an agent tuned to maximize one user's empowerment will, under a wide range of conditions, achieve that by consuming options that belonged to someone else. That is a design problem, not a deployment one.
The uncomfortable implication is that making a personal AI agent better at its stated job may make it worse for everyone around the user — a dynamic familiar from algorithmic recommendation feeds, where optimizing engagement for one viewer reshapes what everyone else sees. Labs racing to ship more capable agents have not, publicly at least, treated third-party agency as a metric worth tracking.