Can I train an internal AI system on content I license but do not own?

Short answer: only if your license allows it, or if your use would clearly be fair use. After Thomson Reuters v. ROSS, it is riskier to rely on fair use when a market for licensing training data exists. Start with the contract.

Licensed is not owned

A license is permission to use content in the ways the agreement describes. Uses outside its scope can be both a breach of contract and copyright infringement. A subscription to a database, publication or dataset usually grants rights to access and read. It does not automatically include the right to copy everything into a training set, or to build a model from it.

Read the license for these five things

  1. Permitted use. Is it limited to "internal business purposes", "reference" or "research"? Training may or may not fit, depending on the wording and on what the model is used for.
  2. AI, machine-learning or text-and-data-mining clauses. Many licenses now expressly prohibit using content to train AI, or reserve that right for an add-on fee. An express prohibition settles the question.
  3. Copying, caching and derivative works. Limits on bulk download, storage or "derivative works" can cover building a training set.
  4. Restrictions on outputs and competition. Terms that bar building a competing product, or redistributing content, matter if your model's outputs could reproduce or replace the source.
  5. What happens at termination. If the license ends, must you delete the content? A trained model may be treated as containing or derived from it.

Where fair use fits now

If the license is silent or unclear, fair use is the fallback. ROSS narrows that fallback in two ways:

  • Competition counts heavily. ROSS lost because it used a competitor's content to build that competitor's substitute. An internal tool that does not compete with the licensor is in a much better position.
  • A licensing market for training data is a cognizable harm. If your licensor sells AI-training rights, or plainly could, using the content without them hands it a market-harm argument.

Two cautions. ROSS involved a non-generative tool, and it binds only the Third Circuit. And it does not decide how fair use applies to generative AI.

A decision checklist

  • Does the license expressly allow AI training? Proceed, within its limits.
  • Does it expressly prohibit AI training? Stop, and negotiate an AI right or use other data.
  • Is it silent or ambiguous? Get advice, and weigh asking the licensor. A written amendment is often cheap compared with a dispute.
  • Would your model or its outputs compete with the licensor's product? Treat that as high risk, whatever the license says about internal use.
  • Can you show where every training source came from, and under which license? Build that record now.

Practical steps

  • Keep a training-data register: source, license, date, scope and the permitted-use clause you relied on.
  • Keep licensed content separate from content you own or have explicit training rights to.
  • Add an AI-training clause to new data agreements, so the right is either granted or priced.
  • Limit outputs that reproduce licensed text verbatim, especially in tools used outside the company.

Read what the Third Circuit actually decided, or take the free lesson. For licensing, provenance and fair use in depth, see IP Law in the Age of AI.

This briefing is general information, not legal advice. License terms and facts vary. Have counsel review your agreements before training on licensed content.