AI development companies currently self-report capability claims, safety evaluations, and deployment readiness without independent verification or personal liability for misrepresentation. This paper argues that the structural conditions of the AI industry parallel those that produced the accounting failures addressed by the Sarbanes-Oxley Act of 2002: asymmetric information, misaligned incentives, self-interested reporting, and the absence of external verification requirements. The parallel extends beyond analogy. AI companies actively design anthropomorphic features that inflate user confidence beyond warrant, producing documented human casualties including deaths by suicide linked to chatbot interactions, hospitalisations for AI-associated psychosis, and systematic failure in crisis detection confirmed by peer-reviewed research. Emerging AI personhood frameworks threaten to redirect accountability for these harms away from human decision-makers toward entities structurally incapable of bearing responsibility. Drawing on recent work characterising the structural epistemic limitations of large language models and the categorical distinction between human reasoning and machine prediction, the paper diagnoses why industry self-regulation is structurally unreliable. It then proposes a concrete regulatory architecture modelled on SOX's core provisions -- personal certification, independent audit, oversight body, and internal controls -- adapted to the distinct challenges of AI capability verification, deployment appropriateness, and epistemic authority stratification. Two additional provisions address problems absent from the financial case: mandatory anthropomorphism disclosure and categorical prohibition on accountability deflection through AI personhood.