What Algorithmic Disgorgement Is and Is Not
Algorithmic disgorgement is distinct from standard data disgorgement. Data disgorgement requires deletion of a dataset obtained in violation of law or a prior order. Algorithmic disgorgement goes further: it requires deletion of any model, algorithm, or other derived work product that was built using that data, even after the source dataset itself has been deleted. The FTC's rationale, as expressed in staff commentary, is that a company should not retain the economic value created by unlawfully collected data simply because the raw data has been removed. If a model trained on that data continues to operate, the underlying violation continues to generate benefit for the company.
This matters for AI governance because model weights, embeddings, and fine-tuned checkpoints can persist and continue delivering value long after a dataset has been purged. A governance program that only tracks dataset deletion, without tracking which models were trained on that data, has not addressed the actual exposure the FTC has targeted in past enforcement matters.
The Enforcement Record Behind the Remedy
The FTC has applied algorithmic disgorgement in a small number of settlements, each grounded in specific factual conduct rather than a general AI risk theory. In the Everalbum matter, the FTC required deletion of facial recognition models built from photos collected without proper user consent. In the WW International/Kurbo settlement, the FTC required disgorgement of algorithms derived from children's personal data collected without verifiable parental consent under COPPA. The Rite Aid settlement included deletion of algorithms tied to facial recognition technology found to produce discriminatory outcomes.
In each case, the trigger was not the use of AI itself but a violation tied to how the underlying data was collected, represented, or used relative to an existing consent decree or statutory obligation. The legal authority in these matters has consistently been Section 5 of the FTC Act, applied to unfair or deceptive practices, not a standalone AI-specific statute. No enforcement action identified in current research has ordered destruction of a large language model or generative AI system; the confirmed precedents involve facial recognition, targeted advertising, and health or child data algorithms. This distinction matters for governance leaders assessing how directly these precedents apply to generative AI deployments.
Why There Is No Fixed Legal Test
The FTC has not published a codified standard specifying when a model, rather than only the data behind it, must be destroyed. Every instance of algorithmic disgorgement to date has been a negotiated settlement term, shaped by the specific facts of the case, including whether the company violated an existing consent order, misrepresented its data collection practices, or processed sensitive categories of data such as biometric or children's information without consent.
This means algorithmic disgorgement risk cannot be assessed against a fixed checklist. It is a fact-dependent determination made during settlement negotiations, and the absence of a bright-line rule is itself a governance challenge. Enterprises cannot point to a statute and confirm compliance; they must instead be able to demonstrate, with documentation, that the factual predicates the FTC has relied on in past cases (unlawful collection, absent consent, consent decree violations) do not apply to their own training data and model lineage.
Runtime Auditability as a Governance Gap
Model lineage documentation addresses the training pipeline, but enforcement scrutiny does not necessarily stop at training. If a deployed AI agent produces outputs or takes actions based on a model whose training data provenance is disputed, the organization needs a way to trace which model version, training data snapshot, and consent basis underlie that specific agent's behavior at runtime, not just at the point the model was built.
This is where audit logging and runtime monitoring intersect with algorithmic disgorgement risk. A governance program that can only answer provenance questions about a model in the abstract, without connecting that model to the specific agents and tool calls operating in production, has a gap between its compliance documentation and its actual deployed footprint. Runtime governance platforms that log agent identity, model version, and permission scope at the point of execution give compliance and legal teams a concrete record to produce if an inquiry extends beyond the training pipeline to deployed systems. Trussed AI provides runtime governance and audit logging for AI agents, which supports this kind of traceability between model lineage and live agent behavior, though it does not substitute for the underlying data provenance and consent verification work that must happen upstream.