FTC algorithmic disgorgement: why buyers demand clean provenance

What is algorithmic disgorgement?#

Algorithmic disgorgement is a regulatory remedy that orders a company to delete models, algorithms or other work product developed with data it should not have collected or used. Ordinary data remedies stop at deleting the data. Disgorgement goes further, on the theory that a company should not keep the benefit of improperly obtained information once it is built into a model.

In the United States, the Federal Trade Commission has been the main source of these orders. They appear in settlements under the FTC Act's prohibition on unfair or deceptive practices, typically where a company collected personal information without the notice or consent that its own promises or the law required.

What have FTC deletion orders required so far?#

FTC orders with disgorgement terms have generally combined three duties: delete the improperly obtained data, delete models or algorithms developed with it, and certify compliance. Matters have involved photos used for facial recognition and information collected from children, among others. The orders are published on the FTC's website, and counsel should read the current ones rather than rely on summaries.

The pattern matters more than any single case. Each order tied the deletion duty to how the data was collected and what people were told at the time. That is the history a model developer cannot see when it licenses records from someone else, so it asks the supplier to show it.

Why one supplier's records can put a whole model at risk#

One supplier's records can put a whole model at risk because training mixes many sources into a single set of weights, and removing one source after the fact usually means retraining. For a model developer, the exposure from a deletion order is not the dataset; it is the model and everything built on it.

That changes how buyers weigh data provenance risk. Records with an unclear collection history are discounted or declined even when they are otherwise useful, and clean documentation becomes part of what makes a dataset licensable at all.

Disclosure laws add pressure. California AB 2013 requires developers of generative AI systems to post training-data documentation stating, among other things, whether datasets were purchased or licensed and whether they include personal information. A developer cannot answer those questions accurately without records from its suppliers.

The provenance file buyers now expect#

Buyers increasingly expect a provenance file that travels with the data. Industry work points the same way: the Data & Trust Alliance's Data Provenance Standards group dataset metadata into Source, Provenance and Use so that developers can select training data responsibly.

For operational records, a useful file answers six questions in plain language.

• 1Source: which systems the records came from, which legal entity held them and the date range covered.
• 2Collection basis: the privacy notices, terms and customer contracts in force when the records were created, with version dates.
• 3Rights review: which contracts, vendor terms and consents were checked, and what was excluded as a result.
• 4Preparation: which personal and confidential details were removed, and by what method.
• 5Release authorization: who approved the license for the supplier, and when.
• 6Change history: later corrections, withdrawals or replacement records.

Supplier warranties that disgorgement risk drives#

Disgorgement risk shows up in license agreements as warranties, indemnities and cooperation duties aimed at the supplier. Each is reasonable in principle; the negotiation is about scope.

Two requests deserve a firm no: a warranty that the data is fit for training any model, and an obligation to fund retraining. Both move the buyer's model risk onto the supplier without limit.

Illustrative: a software company checks its notice history#

Illustrative: a fictional B2B software company wants to license several years of Intercom conversations and linked Jira issues. During the rights review, its counsel pulls every archived version of the privacy policy and the customer terms.

An early policy version says conversation data is used only to provide the service. Counsel advises excluding conversations from that period pending further review, while later records, collected under broader notices, stay in scope. The team records each policy version and its effective dates in the provenance file and agrees a lawful collection warranty qualified by knowledge and tied to that file.

The company also decides against updating its policy and treating the change as covering older records. FTC staff warned in February 2024 that adopting more permissive data practices, such as using data for AI training, through a surreptitious, retroactive change to terms of service or a privacy policy may be unfair or deceptive.

Mistakes that weaken a supplier's provenance#

The most damaging mistake is losing the notice history. Companies that overwrite their privacy policy page without archiving old versions often cannot show what customers or users were told when older records were created, and a buyer will price that gap or walk away.

Other mistakes include mixing records from an acquired company into the main archive without its separate contracts and notices, licensing exports whose source system and date range nobody recorded, and treating a vendor's statement that the customer owns its data as proof that the customer may license it. Ownership under a vendor contract says nothing about the promises made to the people in the records.

How SourceX documents provenance#

SourceX builds provenance into the transaction rather than adding it at the end. Under the SourceX five-step transaction, the Rights step reviews the notices and contracts behind each record family, and the Preparation step removes personal and confidential details before anything is delivered.

The result is a SourceX Evidence Packet covering provenance, licensing rights, permitted use, the privacy record and release authorization. It gives the buyer the documentation that disgorgement risk calls for, and it gives the supplier a written basis for the warranties it signs.