USA TODAY sued OpenAI over training data, but proving *which* data is the hard part

The lawsuit and the allegation

In October 2026, USA TODAY and a group of affiliated newspapers filed a lawsuit against OpenAI in a New York federal court ^[1]. The papers alleged that OpenAI had used copyrighted news articles to train its AI systems without obtaining permission or paying for the use.

The complaint is the latest in a series of cases where publishers, authors and other rights holders have challenged AI companies over training data. What makes each case hard is not whether copyright applies—it generally does—but proving what the model learned from.

OpenAI, like other large model providers, does not routinely publish exhaustive lists of its training data. That means the plaintiffs face a steep hill: they must show, through discovery or other means, that their content was in the training set, and that it contributed to the model's output in a way that infringes their rights.

Why proving training data is difficult

Think of it like proving someone read your diary two years ago. You know they had access to the room, and they seem to know things only your diary contains, but they never took a photo of the page and the diary itself doesn't keep a visitor log.

Training a large language model works in a similar way. Data is ingested, processed and then discarded, leaving behind only the statistical patterns the model learned. The original articles are not stored inside the model, and there is no automatic ledger of what went in.

That puts the burden of proof on the claimant. In the USA TODAY case, the plaintiffs will likely rely on circumstantial evidence: that OpenAI had access to the web, that their articles were publicly available, and that the model can reproduce passages that are strikingly similar to their work. Whether that is enough depends on the court and on what OpenAI's internal records show.

What the AI Act requires from December 2027

In the EU, the picture is about to change. Article 10 of the AI Act requires providers of high-risk AI systems to use training, validation and test data that is relevant and representative, and to manage it under documented data governance, including the data's origin ^[2].

For systems listed in Annex III, that duty applies from 2 December 2027. The rule does not prescribe exactly how to document the origin, but it does mean a provider must be able to show, when asked, where the data came from.

Auditors, notified bodies and, in a dispute, courts, will want to see that documentation. A provider who cannot produce it, or whose records were created long after the training run, starts on the back foot.

What this means for you

If you train AI models, or commission someone else to do it, here are three questions worth asking now:

• Do you keep a manifest of each data set before it goes into training, with enough detail to show where it came from?
• Could you prove, if challenged a year from now, that the manifest you have today is exactly the one that existed on the day the training started?
• If a third party questions whether a certain data set was used, can you show them a record they don't have to take your word for?

The USA TODAY case is a reminder that these questions are no longer theoretical. Courts in the US are already hearing them, and under the AI Act they will soon be mandatory in the EU as well.

How a digital witness makes data sets checkable

Here, the thing to seal is the manifest or snapshot of a data set when it is frozen for training. That might be a list of file names, sources and timestamps, or a structured record of licences and scraping logs—whatever shows the origin.

Sealing takes a fingerprint of the manifest, not the data itself, so the underlying content never leaves your systems. The fingerprint is recorded on WitniumChain, where any later change would show. If the question comes up months or years later, anyone you share the manifest with can recalculate the fingerprint and check that it matches the sealed one.

What it shows is which version of the manifest existed when the data set was frozen, and that it hasn't changed since. What it doesn't show is that the data was lawful to use, that it was actually fed into the model, or that the model's output infringes anyone's rights. Those are separate questions. But when the first question is "which data did you use?", a sealed manifest from the day of training is a straightforward answer.

How it would work

1. Before training, you finalise the data set manifest and seal it with one API call.
2. WitniumChain returns a witness ID, which you store alongside the model version.
3. Later, if someone asks what data was used, you share the manifest and the witness ID. They check the fingerprint and see that the manifest is unchanged.

The first 2,000 witnesses a month are free. Read more in the docs.

What to watch

The USA TODAY case is ongoing, and its outcome will turn on US copyright law and what OpenAI's internal records reveal. But the broader question—how do you prove what a model learned from?—will recur wherever AI is trained.

In the EU, the clock is ticking. High-risk systems under Annex III must meet the data governance requirements, including documented origin, from 2 December 2027. Providers who have not started keeping those records are now 13 months away from a legal duty that is hard to retrofit.