What is accountability in AI?
Accountability in AI is the principle that a specific person or organization is answerable for what an AI system does, including its mistakes. Every major AI governance framework lists it, and it has a practical shape: a named owner for each system, a record of who approved its deployment and each change, and a route for an affected person to question a decision and receive a human answer. Software cannot be held accountable, so the question is always which human role carries the responsibility.
Accountability depends on traceability. If no one can reconstruct which model version, which data, and which prompt produced a decision, then no one can fairly be held responsible for it. Decision traceability and an AI audit log are the mechanisms that make accountability real.
Accountability in practice
A county government uses a model to prioritize housing assistance applications for caseworker review. The program director is the named owner. Each prioritization record stores the model version and the factors behind the score, caseworkers can override the ranking with a documented reason, and applicants can request a human review of their placement. When a resident appeals, the director can show how the application was handled and who made the final decision. The director also signs off on each model update, so every change has a human decision attached.