Frontier AI as Strategic Infrastructure: Sovereign Governability Under Dependency

Abstract

When personal data enters an artificial-intelligence development lifecycle, deleting the original record may not end the consequences of an unlawful, deceptive, or otherwise prohibited data practice. The data may have informed a model, algorithm, embedding, feature set, evaluation corpus, or other operational artifact. This working paper examines a recurring remedial pattern in Federal Trade Commission enforcement matters: in appropriate cases, the Commission has required deletion or destruction not only of covered source information but also of specified information, work product, models, algorithms, or data products derived from it. The pattern begins earlier than most AI-governance accounts acknowledge. In the 2019 Cambridge Analytica final order, entered after the respondent failed to answer the FTC’s administrative complaint, the Commission required deletion or destruction of covered information and any information or work product, including algorithms or equations, that originated in whole or in part from that covered information. In Everalbum (2021), the Commission used the defined term “Affected Work Product” for models or algorithms developed in whole or in part using covered biometric information. In Kurbo/Weight Watchers (2022), a stipulated order required destruction of affected work product developed in whole or in part using children’s personal information collected through the Kurbo program. In Ring (2023), a federal court order required deletion or destruction of reasonably identifiable affected work product, subject to a sworn certification process where destruction was technically infeasible. In Rite Aid (2024), the FTC paired a five-year facial-recognition-use prohibition with deletion or destruction of covered photos and videos and data, models, or algorithms derived in whole or in part from them, along with third-party notification and deletion-direction obligations. Amazon Alexa (2023) occupies a related but distinct role: it illustrates the enforcement significance of retention, deletion commitments, and restrictions on future use of information subject to deletion directions, rather than serving as a general precedent for deletion of already-trained models. These matters do not establish a universal doctrine of “tainted AI.” They do not hold that every downstream artifact associated with problematic data is automatically unlawful, that every customer of a model vendor inherits liability, or that every deletion request always requires erasure of model parameters. The applicable remedy remains dependent on statutory authority, procedural posture, operative order language, covered data, derivation nexus, technical feasibility, and proportionality. The matters nevertheless reveal a consequential enforcement logic: a firm may not always retain specified informational or algorithmic value created from data it had no right to collect, retain, or use in the relevant manner. This working paper develops that logic into a governance framework for verifiable AI remediation. It proposes a Full-Lineage Remediation Standard, a Training Provenance Record, a derivation-nexus analysis, a material-contribution inquiry for large mixed training corpora, and an evidence hierarchy for deletion, model retirement, retraining, and machine unlearning. It also compares the U.S. remedial pattern with the European Data Protection Board’s Opinion 28/2024 on AI models and personal data, which recognizes that unlawful processing in model development may affect subsequent deployment and that supervisory authorities may impose proportionate corrective measures extending from dataset erasure to retraining or, in appropriate cases, erasure of an AI model. The paper concludes that AI governance must be capable not only of recording what data a system holds, but of evidencing what that data helped build, where the resulting artifacts reside, and what action the organization took when the underlying data practice failed.