Content Licensing Deals Between Publishers and AI Labs
9 min read · updated August 4, 2026
The headline number in a licensing story is usually a figure neither party confirmed. The structure of the agreement is far more informative and is often described openly, so this page works from structures: what is being bought, what is being granted, and which clauses are systematically absent from every announcement.
Five structures, often combined
| Structure | Description |
|---|---|
| Training corpus licence | A right to use an archive as training data, usually for a term, sometimes including the back catalogue. The distinguishing feature is that the publisher gets nothing ongoing from it: the value transferred is one-directional and permanent in effect, since a trained model cannot unlearn. |
| Retrieval and display licence | A right to fetch current content and quote it in answers, with attribution and a link. This is the structure that keeps producing referral traffic, and it is the one publishers with a live audience care most about. |
| Product integration | The publisher's material appears as a named source inside the product, or the publisher receives model access, API credits or engineering help. Value flows partly as services rather than cash. |
| Revenue share | The publisher receives a share of revenue earned on answers that cite it — advertising alongside the answer being the common form. Perplexity announced a publishers' programme along these lines in 2024. |
| Settlement converting to a licence | A dispute resolves into a forward-looking agreement. This is now a recognised path rather than an exception, and it means the litigation docket and the deal announcements are the same story. |
Most announced arrangements combine at least two. A deal reported as “a licensing agreement” will typically cover training rights, live retrieval with attribution, and some product integration, on different terms for each — which is why a single headline figure tells you almost nothing about what was actually traded.
What has been publicly announced
The following were announced publicly by the parties. Dates are the announcements, not the terms.
- OpenAI with news publishers. The Associated Press (2023), Axel Springer (2023), Le Monde and Prisa Media (2024), the Financial Times (2024), Dotdash Meredith (2024), News Corp (2024), Vox Media and The Atlantic (2024), Time (2024), Condé Nast (2024) and Hearst (2024). Publicly described terms typically cover both training and attributed display in ChatGPT.
- OpenAI with platforms. Reddit and Stack Overflow, both announced in 2024, which are different in kind: they license user-generated corpora rather than editorial output, and raise questions about the contributors’ position that the platforms answer in their own terms of service.
- Google with Reddit, announced in 2024.
- Meta with Reuters, announced in 2024, covering real-time news in Meta’s assistant.
- Amazon with The New York Times, announced in 2025 — notable because the Times was simultaneously litigating against OpenAI and Microsoft, which shows that a publisher can be a licensor and a plaintiff at the same time.
- Perplexity’s publishers’ programme, announced in 2024 with an initial group of publishers, structured as a revenue share rather than a fee.
The terms that are never in the press release
This is the useful part, because these are the clauses that determine whether a deal was good, and their absence from every announcement is itself a consistent, observable fact.
| Term | Description |
|---|---|
| Price and payment shape | Whether it is a one-off, an annual fee, per-article, or contingent on usage. A flat fee and a usage-linked fee behave completely differently as the product scales. |
| Whether past training is covered | Many deals settle the future and quietly resolve the past. Whether the agreement includes a release for content already ingested is the single most valuable clause and is essentially never stated. |
| Term and renewal | A three-year training licence over a permanent capability is a very different trade from a perpetual one, and the asymmetry is rarely acknowledged. |
| Exclusivity | Whether the publisher may sign with competing labs. Exclusivity commands a premium and forecloses the rest of the market. |
| Attribution obligations | Whether links are required, how prominent, and whether the obligation survives a product redesign. A display licence with no enforceable attribution clause is a training licence with extra steps. |
| Audit rights | Whether the publisher can verify usage at all. Without an audit right, a usage-linked payment is whatever the counterparty reports. |
| Indemnity | Who bears the risk if the content itself infringes something, and who bears it if the model reproduces it verbatim. |
| Termination and unwinding | What happens to a trained model when the agreement ends. Practically, nothing can happen, which is why the clause is uncomfortable and unpublished. |
Litigation as the other negotiating table
The deals cannot be understood apart from the cases, because the cases set the price of not doing a deal. A few that were publicly filed and publicly decided:
- The New York Times against OpenAI and Microsoft, filed in December 2023, alleging both training use and verbatim reproduction. The second allegation is the more consequential one legally, because it is about output rather than input.
- Thomson Reuters against Ross Intelligence. A 2025 ruling rejected a fair-use defence in a case about legal research headnotes used to build a competing product, on facts that included direct market substitution.
- Bartz and others against Anthropic. A 2025 ruling drew a distinction that has shaped the field: training on lawfully acquired books was treated differently from building a library from pirated copies. The case was subsequently reported as settled. The structural lesson — that how you acquired the corpus can matter separately from what you did with it — is the part worth carrying forward.
- Getty Images against Stability AI, filed in both the United States and the United Kingdom, raising trademark and database questions alongside copyright.
The background law differs sharply by jurisdiction. The EU introduced a text-and-data-mining exception with a rights-holder opt-out, which is the legal basis for treating a machine-readable reservation as meaningful in Europe. The United States has no equivalent statutory exception, and the question runs through fair use case by case. Neither framework is settled in application; see copyright and training data and the copyright questions around AI output.
The infrastructure layer joined in
The change that most affects sites too small to negotiate is that the intermediaries started acting. In July 2025 Cloudflare announced that it would block AI crawlers by default for new domains on its network, and introduced a pay-per-crawl mechanism in beta, using the HTTP 402 Payment Required status to signal a price for access.
That is structurally different from every other item on this page. Robots.txt asks; a CDN default enforces; and a 402 turns access into a transaction that can be priced without a contract per publisher. It also concentrates a policy decision in an infrastructure provider, which is a trade-off worth being aware of even if you like the outcome.
In parallel, standards work has been under way to give rights reservations a machine-readable form that is more expressive than a robots.txt disallow — an IETF working group on AI preferences among them. That work is in progress rather than deployed, and a proposal is not a control until crawlers act on it.
If you are not a large publisher
The honest position: individual licensing is not available to you. No lab is negotiating with a site that publishes a few hundred pages, and no realistic figure would cover the legal cost of the negotiation. What is available:
- Decide what you are actually optimising for. Being in the training corpus, being retrievable and cited, and being paid are three different goals, and the first two are usually in tension with the third.
- Express the decision where it is read. robots.txt for the compliant crawlers, the platform-level controls your CDN offers for the rest, and your terms of service for the legal layer.
- Put the terms in a place a machine can find — a visible licence statement and a
licenseproperty in your structured data. It costs nothing and it removes the ambiguity defence. - Watch collective arrangements. Trade bodies and collective licensing organisations are the mechanism through which small publishers have historically reached deals they could not reach individually, and that is the shape any solution for you will take.
- Keep the evidence. Logs showing which crawlers fetched what, and when, are the only record you will have if the question ever becomes concrete — the crawler audit produces exactly that.