AI Training Data and Copyright Law: What the Thomson Reuters Appeal Ruling Means for Your Contracts

Why the Thomson Reuters Appeal Ruling Is a Turning Point for AI Copyright Law

The US Court of Appeals has upheld Thomson Reuters' landmark win in its case against Ross Intelligence, confirming that using copyrighted content to train an AI system is not automatically protected by the fair use doctrine under US copyright law. For any business that either licenses data to third parties or builds AI products that ingest external content, this ruling is the clearest judicial signal yet that AI training data copyright risk is real, enforceable, and commercial.

The case centred on Ross Intelligence's use of Westlaw headnotes, which are editorial summaries authored by Thomson Reuters' legal editors, to train a competing AI-powered legal research tool. The court found that this use was not transformative enough to qualify as fair use, a conclusion that has immediate consequences for how the AI industry thinks about data sourcing.

What "Not Transformative" Actually Means in Practice

Fair use analysis in US copyright law turns on four factors, with transformativeness carrying the most weight in modern case law. The court's finding that training an AI on copyrighted material, and then selling a competing product, does not constitute transformation is significant. It suggests that the mere act of converting prose into model weights does not change the legal character of what was copied.

This is not a blanket prohibition on AI training. Courts will still assess the nature of the original work, the amount copied, and the effect on the market for the original. But the ruling firmly rejects the argument, popular in some AI development circles, that machine consumption of text is categorically different from human reading and therefore exempt from copyright constraints. Businesses reviewing their AI training data licensing agreements should treat this ruling as confirmation that the legal exposure they may have assumed was theoretical is, in fact, litigable.

The Contract Clauses That Now Matter Most

For data licensors and licensees, this ruling reshapes the negotiating landscape around several specific clause types.

Permitted use definitions are the first battleground. If a data licence grants rights to "research and internal use" without explicitly addressing AI training, that ambiguity is now dangerous for the licensee and potentially valuable for the licensor. Licensors should review their standard templates to ensure AI training is either expressly permitted (with appropriate pricing) or expressly excluded.

Downstream use and sublicensing restrictions matter equally. A company that trains a model on licensed data and then makes that model available to third parties may be creating a second, distinct infringement event, separate from the training itself.

Indemnification carve-outs in AI product contracts are increasingly common, but many were drafted before courts had weighed in meaningfully on training data. Standard IP indemnities often exclude liability arising from the customer's own data inputs. Vendors and customers alike need to re-read these provisions with fresh eyes, because the line between "your data" and "third-party data your model was trained on" is blurring.

How This Ruling Affects AI Vendors and Their Customers Globally

Although this is a US federal ruling, its influence will extend well beyond American courts. The European Union's AI Act and the UK's ongoing consultation on text and data mining exceptions are both live policy questions, and judicial reasoning from US appellate courts carries persuasive weight in common law jurisdictions and informs regulator thinking elsewhere.

For AI vendors selling into global enterprise markets, the ruling creates a disclosure pressure. Customers procuring AI-powered products are increasingly asking where training data came from and whether it was lawfully obtained. That question, once a due-diligence nicety, is now a genuine contractual risk allocation point. Vendors who cannot demonstrate clean provenance for their training data will face harder negotiations and potentially wider indemnification demands.

What Businesses Should Do With Their Contracts Right Now

The practical response is not panic but precision. Three actions are worth prioritising.

First, audit existing data licences for AI training language. Silence on the point is not safe for either side. If you are a data provider, silence may mean you are undercharging. If you are a data consumer, silence may mean you are exposed.

Second, update your standard vendor terms and procurement templates to address AI training data provenance explicitly. Ask vendors to warrant that training data was obtained with appropriate rights, and ensure that warranty is backed by a meaningful indemnity.

Third, if you are building AI products that ingest customer data, review your permitted use and model improvement clauses carefully. The Thomson Reuters ruling reinforces that courts will look at commercial substitution, meaning whether your AI product competes with or displaces the market for the original content. If it does, fair use becomes very difficult to argue.

AI training data copyright law is no longer a theoretical field. It has appellate precedent, and contracts that were drafted in the pre-ruling era need to catch up.