Accountability Gap

The distance between the humans whose choices (build, deploy, approve, configure) shaped what an agent could do, and the specific harmful action the agent eventually takes on its own, with no human reviewing that specific action before it happens. The gap widens as autonomy increases and narrows as meaningful human checkpoints are added. The gap can also hide inside a shared responsibility: when several value-chain roles each assume another role covered a safeguard, the result is an action nobody specifically owns, an "everyone and therefore no one" pattern that looks like distributed accountability but functions as none at all.

the unclear chain of responsibility when an AI agent acts and causes harm; under precedent such as Moffatt v. Air Canada, organizations, not agents, bear legal and professional accountability.

The shift in where responsibility must be engineered once a system moves from producing outputs a human must act on to taking actions on its own, with no human checkpoint at the moment of action. Named here, developed fully in Topic 0.3. (see Topic 0.3)