Somewhere between the term sheet and the signature, a buyer’s counsel sends a request list with a section most AI startups have never had to answer: for every dataset used to train or fine-tune the product, where it came from, under what rights, and who can prove it. That section is hard to answer not because the data is hidden but because the record was never kept.
One federal court separated what you did with the data from how you obtained it
That separation came on June 23, 2025, when Judge William Alsup of the Northern District of California granted partial summary judgment in Bartz v. Anthropic. As Akin Gump’s analysis of the order explains, the court held that training large language models on copyrighted books was fair use, calling the use “exceedingly transformative,” and that scanning lawfully purchased print books into searchable digital copies was fair use as well. The court drew a hard line at the other half of the record: downloading over seven million pirated books to build a permanent internal library was not fair use. That piracy question never reached trial; the parties settled it.
The settlement put a number on the line the court drew. In September 2025, Anthropic agreed to a class settlement that Kilpatrick Townsend summarized as at least $1.5 billion, an average of $3,000 per work, with a commitment to destroy the pirated datasets and a release limited to conduct through August 25, 2025. On July 20, 2026, Judge Araceli Martínez-Olguín granted final approval. The figure is a settlement rather than a damages verdict, but it is now a reference point — $3,000 per work — for data a company was not entitled to hold.
The United States Copyright Office reached a compatible position in the pre-publication version of its report on generative AI training, released in May 2025. In Lowenstein Sandler’s summary, the Office concluded that “some uses of copyrighted works for generative AI training will qualify as fair use, while others will not,” that commercial uses generating competing expressive content, especially through illegal access or pirate sites, are unlikely to qualify, and that the voluntary licensing market should be left to develop without immediate government intervention. The report itself answers a common founder defense in a single footnote: “‘Publicly available’ is not synonymous with ‘authorized.’”
Fair use may excuse what you did with the data. It does not settle how you got it, and buyers now know the difference.
Fair use may excuse what you did with the data. It does not settle how you got it, and buyers now know the difference.
The request list was rewritten before most buyers learned to read it
Transaction lawyers moved first, and U.S. counsel are explicit about it. Foley Hoag’s January 2026 alert says acquisition representations should address the ownership and licensing of AI technologies, including training data, “to ensure there are no undisclosed third-party rights or infringement risks,” and should cover the provenance of the data used in AI systems. A Canadian firm, Osler, reached the same checklist a year earlier, extending the standard data representations to “assurances regarding the quality, accuracy, and legality of the data used for training AI systems” and to “third-party data licensing agreements or datasets created through partnerships.” Each of those is a claim a seller signs.
Two in three dealmakers do not yet run those checks, and that is a trap for sellers, not a reprieve. KPMG’s August 2026 analysis of its 2025 Technology M&A Survey found that “only 33% of dealmakers prioritize investigating technical and AI debt during pre-deal evaluation,” even though 66% discuss it during deal planning, and argued that diligence “built for the SaaS world” evaluates the wrong signals when pointed at AI-shaped targets. A question a buyer skips in diligence does not disappear. It shifts into the representations the seller signs, and a representation that turns out wrong becomes an indemnity claim after the deal has closed.
A defensible answer is a manifest, and it is built before the data room opens, not after
What a sophisticated buyer wants is not a warranty but a document: a training-data manifest. For every dataset that touches every model version in production, it records the source, the acquisition method, the rights relied on, the date acquired, and the person who approved the use. In practice that is a table with one row per dataset and a column for each field, plus a link to the evidence — the license file, the purchase record, or a saved snapshot of the terms of use and robots.txt directives in force the day web data was collected. Four questions decide whether it survives review.
- Acquisition. Was each dataset licensed, purchased, generated internally, contributed by customers under contract, or collected from the web? For collected data, the manifest shows the terms of use and crawler directives that applied when it was gathered, because the Copyright Office has already said that availability is not authorization.
- Rights to the use made. Does the license or customer contract cover training, fine-tuning and retention of the resulting weights, or only the service the data was supplied for? Read the customer agreement first: a clause allowing use of customer data “to provide and improve the services” does not obviously reach training a model whose weights you keep, and a buyer’s counsel will read it the narrow way.
- Chain to the model. Can you tie each dataset to the specific model versions it shaped? If the weights shipping today hold data you can no longer justify, the buyer will ask what retraining without it would cost, and that cost tends to surface in the price.
- Base-model terms. Which open-weight or vendor model underlies the product, what does its license permit commercially, and does your fine-tuning or distillation comply with it?
Three steps put the manifest within reach.
- This week, the engineering lead lists every dataset and every model version running in production, one row each.
- Within 30 days, the founder and counsel audit customer and vendor contracts for language that permits training and retention of the resulting weights.
- Before the data room opens, the founder decides each dataset in turn: keep it with the evidence attached, relicense it, or retire it and retrain without it.
A federal district court has held that training on copyrighted works can be transformative fair use and that building a permanent library from pirated copies is not. The Copyright Office has said that availability is not authorization. U.S. transaction counsel are already writing data provenance into the representations a buyer asks a seller to sign. None of it is settled law for every case, but all of it is on the request list today. The founders who can hand over the manifest will spend diligence answering questions; the ones who cannot will spend it negotiating the price of a record they never kept.