Deleting pirated training data when a lawsuit looms is a trust problem

If your training data is clean, why would you ever need to delete it? That is the question hanging over the latest twist in the Meta copyright saga, and "we were being careful" is not a satisfying answer.

What came out

A former Meta researcher, Stephen Roller, who worked at the company from 2017 to 2023, testified that he downloaded an AI training dataset that includes the pirated Books3 collection, using BitTorrent. According to a summary of his deposition, he deleted his copy as the lawsuits mounted. Hacker News picked it up under a headline that says it all: deleted "due to litigation."

The backdrop is a pile of cases in California federal court over Llama's training data. Plaintiffs allege Meta pulled roughly 80.6 terabytes from LibGen and about 35.7 terabytes from Anna's Archive, according to BigGo's summary of the filings. A magistrate judge has now ordered Meta to name the people who approved that sourcing. We're relaying allegations here, not findings. Meta has not been found liable for the torrenting.

Deleting the evidence is the wrong reflex, even if it's not the smoking gun

We can't say from the documentation whether this deletion broke any rule, and we won't pretend to. What we can say is how it looks, and in a trust business, how it looks is the product. An AI lab asks us to accept that its models were built responsibly. When a dataset at the center of the dispute gets erased as lawsuits pile up, the reasonable reaction is suspicion. Maybe the reason is innocent. The effect is the same.

There is also a deeper issue. We don't know what records Meta kept, but this episode shows how fragile the alternative is: if a dataset's origin lives in one researcher's download folder, it can disappear with one command. That is not a provenance system. It is an honor system, and honor systems tend to break precisely when the stakes get high.

The "blame the staff" defense makes this worse

Ars Technica reports that Meta has spent weeks blaming staff for the torrenting, and has now been ordered to name the executives who authorized it. Our reading: that is what tends to happen when decisions about data aren't clearly recorded. Responsibility defaults to whoever pressed the button. The researcher becomes the story, and the decision-makers stay in the fog.

That is bad for researchers, who risk being exposed for choices they may not have owned. It is bad for companies, which look evasive. And it is bad for authors, who can't easily reconstruct what was used on their work. Court records cited by Tom's Hardware show that a senior AI researcher voiced concerns as far back as October 2022. If that's accurate, a system that logged those concerns and the answers to them would have been worth more than any deletion.

What auditable by default would look like

This isn't exotic engineering. Labs can keep a manifest of every corpus: where it came from, who approved it, under what license or legal theory, and a cryptographic fingerprint of the contents. The files themselves can be purged if they must be. The record of what they were should not be. Retain the receipt, not necessarily the goods.

Labs will object that manifests become discovery fodder. Probably true, and we think that is the point. The legal picture is unsettled. In Kadrey v. Meta, per the Fiund tracker, Meta won summary judgment on fair use for training Llama on thirteen authors' books, but the judge framed it as a warning, not a blessing, and a separate claim over the torrenting remains unresolved. Meanwhile nearly 300 authors are seeking a jury trial and the maximum $150,000 per act of infringement, Publishers Weekly reports, and the attorney involved says many more cases are coming. In that environment, a lab that can show clean provenance has an asset. One that can't has a liability it tried to delete.

Our verdict

Fair use may protect training, as one ruling suggested for a narrow set of books. It does not protect an opaque supply chain from reputational damage. If the industry wants the benefit of the doubt, it should earn it with records that survive a lawsuit, not hard drives that don't. Regulators and customers should start asking for provenance manifests the way they ask for security audits. And any lab that treats deletion as its documentation strategy should expect people to assume the worst about what was deleted.