Public AI transparency records are widely understood as governance instruments: structured disclosures through which governments account for their use of algorithmic tools. This working paper argues that they should also be understood as media forms. Like a book that passes through interface, database, algorithm, and protocol before reaching a reader, a public decision increasingly passes through AI systems, administrative workflows, reporting templates, and transparency schemas before it reaches the public as an accountable object. Drawing on media theory — including Flusser’s apparatus, Fuller’s conditions of appearance, Blanchot’s withdrawal, and Derrida’s supplement — and on close reading of two UK Algorithmic Transparency Recording Standard records, the paper argues that ATRS records produce a specific kind of readability: institutional, systemic, and population-level. What they do not produce is the capacity to reconstruct a specific decision event. This is not treated as operator failure, but as a structural effect of the apparatus. The paper terms this structure the schema’s reading path and asks what kind of accountability can be performed through a record that was never designed to preserve the event it names. It contributes to ongoing work on algorithmic transparency, public records, media theory, AI governance, and decision-event reconstructability.
Hon Bor So (Thu,) studied this question.