Extension: Accountability Capture: How Record-Keeping to Support AI Transparency and Accountability (Re)shapes Algorithmic Oversight
Accountability is not a single property of an agent runtime. Chappidi and her coauthors define an accountability relationship through an actor who owes an account, a forum entitled to receive it, and information that permits review and response. They then distinguish records from a system, which describe its operation, from records about a system, which describe how it was commissioned, built, deployed, and monitored. Records may support investigation, oversight, challenge, repair, and redress. The paper also identifies accountability capture. The apparatus used to observe conduct can reconfigure the conduct, induce evasion, and distort the accountability regime itself. This structure independently supports separate diagnosis. Attribution fails when operational records cannot reconstruct what occurred. Authority drifts when the actor, forum, governing requirement, or design decision is not preserved across commissioning and use. Evaluation becomes captured when the measurement apparatus changes behavior or serves interests that defeat the accountability purpose. Recourse disappears when records do not reach a forum capable of challenge, response, and remedy. The correspondence is functional, not terminological. The paper does not use Kaal's four labels, prove that the taxonomy is exhaustive, or study sovereign local agent runtimes. Its survey concerns organizational record keeping around algorithmic systems. It nonetheless shows why a general appeal to trust is inadequate. Each failure has different evidence and a different institutional repair. A runtime assessment should therefore test traceability, delegated authority, evaluator independence, and enforceable recourse separately.
governance-designinstitutional-designrisk-and-incentivesai-and-agentsaccountabilitytraceabilityauthorityevaluationrecourse