The Institutional Production of Causation: Predictive Classification and the Limits of Self-Attested Transparency in AI Governance

<p>Predictive classifications do not merely describe the world. Once embedded in institutional decisions, they alter access, scrutiny, opportunity and the evidence from which their own validity is later assessed. This Perspective extends that account of institutional causation to the governance of AI systems. Providers and oversight bodies classify incidents as authorised or unauthorised, material or immaterial, in scope or out of scope, and verified or unverified. These labels allocate investigative attention and determine which records are preserved, which populations are examined and which explanations remain available. When the institution issuing a classification also controls the evidence capable of testing it, a closed evidentiary loop can result: the initial category defines the inquiry, the inquiry defines the evidence population, and that population ratifies the category. Detailed disclosure may then remain self-attested rather than independently verifiable. The discovery disputes in The New York Times Company v. Microsoft Corporation et al. illustrate the institutional difference between disclosure and contestability. Court-ordered preservation, negotiated sampling, de-identification and controlled access did not decide the merits of the litigation, but they created an external forum in which evidence scope and negative assurances could be challenged. Transparent AI governance therefore requires more than publication. It requires institutional separation between the authority to classify and the authority to determine what evidence may test the classification.</p>