News publishers license their archives and live reporting to AI companies under deals that grant training rights, display rights in AI answers, or both, and the parties' own announcements say which. As of 2026-11-10, six publisher deals with OpenAI state training or model improvement in the parties' words, most other deals describe display, and Le Monde's 2025 Perplexity deal excludes training. Google states that it pays participating publications in its News AI pilot (400+ as of 2026-10-15) and discloses no amounts.
No publisher has filed a per-deal price. The New York Times reports a USD 14.4 million rise in licensing revenue for 2025 and USD 13.3 million of generative AI litigation costs for the same year, not netted here. Staff work belongs to the publisher, while freelance, wire and syndicated copy depends on each contract, and not licensing is a valid outcome. News content is one of the oldest sources in AI training data licensing.
What a News Publisher Can License and What Stays Out
A publisher can license the content it owns or controls, mainly archive text and live reporting, while pieces it only licensed in from agencies, freelancers or other publishers stay outside unless their contracts allow it. Announced deals show four kinds of asset:
- Archive text, such as the part of the Associated Press text archive that OpenAI licensed in the agreement of 2023-07-13.
- Live feeds, such as the real-time AP feed that Google announced for the Gemini app on 2025-01-15.
- News services, such as the multi-year AFP agreement with Mistral AI of 2025-01-16.
- Photos, video and metadata that the publisher owns.
Mercor's publisher listing of 2026-10-17 excludes short articles alone and content held under someone else's license. See news archives as AI training data.
Three Shapes of a Publisher Deal: Training Rights, Display Rights or Both
Publisher deals fall into three shapes by what the publisher grants: training on the content, display of it in AI answers, or both, and many announcements state only the second. The table lists news and magazine publisher deals as of 2026-11-10, by date, with the use column taken from each party's words: T training stated, D display or answers with training not stated, X training excluded, n/s use not stated. The value is not disclosed in every row, and "not stated" does not mean training is forbidden.
| Date | Publisher | AI company | Use | Stated scope (party's words) | Source |
|---|---|---|---|---|---|
| 2023-07-13 | The Associated Press | OpenAI | n/s | OpenAI licenses part of AP's text archive | AP release |
| 2023-12-13 | Axel Springer | OpenAI | T | "advancing the training" of large language models | Axel Springer release |
| 2024-03-13 | Le Monde | OpenAI | T | "l'entraînement des modèles d'IA" | Le Monde |
| 2024-03-13 | Prisa Media | OpenAI | T | "la formación y mejora continua de los modelos" | El País |
| 2024-04-29 | Financial Times | OpenAI | T | "improve its models' usefulness" | FT release |
| 2024-05-07 | Dotdash Meredith (name as announced) | OpenAI | T | "used to improve OpenAI's large language models" | IAC release (SEC) |
| 2024-05-22 | News Corp | OpenAI | D | display in answers and "enhance its products" | News Corp release |
| 2024-05-29 | Vox Media | OpenAI | T | "to be used to train its AI models" | Vox |
| 2024-05-29 | The Atlantic | OpenAI | D | content and product partnership | The Atlantic |
| 2024-06-27 | TIME | OpenAI | D | display in responses | TIME |
| 2024-08-20 | Condé Nast | OpenAI | D | "proper attribution and compensation" | Condé Nast |
| 2024-10-08 | Hearst | OpenAI | D | citations and links | Hearst |
| 2025-01-15 | The Associated Press | D | real-time feed for the Gemini app | Google blog | |
| 2025-01-16 | AFP | Mistral AI | n/s | multi-year agreement | Mistral AI |
| 2025-02-14 | Guardian Media Group | OpenAI | D | strategic partnership | Guardian |
| 2025-04-22 | The Washington Post | OpenAI | D | "summaries, quotes, and links to original reporting" | WashPostPR |
| 2025-05-13 | Le Monde | Perplexity | X | no LLM training; answers only | Le Monde |
Deals in Which the Publisher States Training
Six publisher deals with OpenAI state training or model improvement in the parties' own words: Axel Springer (2023-12-13), Le Monde and Prisa Media (both 2024-03-13), the Financial Times (2024-04-29), Dotdash Meredith (announced 2024-05-07) and Vox Media (2024-05-29).
News Corp's deal of 2024-05-22 describes display and product use and does not state training. Every publisher deal is logged among AI data licensing deals.
Display, Grounding and Answer Deals Without Training
Display deals let an AI product show summaries, quotes and links to a publisher's reporting, and their announcements do not mention training. The Washington Post's deal of 2025-04-22 names "summaries, quotes, and links to original reporting", and Condé Nast's of 2024-08-20 names "proper attribution and compensation".
Le Monde's deal with Perplexity, announced on 2025-05-14, differs: Le Monde writes that it "ne prévoit pas l'utilisation de nos contenus pour entraîner un LLM", so the deal covers answers only.
Program-Wide Offers From Google, Microsoft and Meta
Since late 2025 three large platforms have offered terms to many publishers at once, and none of them states training in its announcement:
- Google News AI pilot: launched 2025-12-10 with 400+ publications as of 2026-10-15; Google states "we pay participating news publications for the delivery and use of enhanced content", with amounts not disclosed.
- Microsoft Publisher Content Marketplace: announced 2026-02-03; publishers "will be paid on delivered value" and define their own terms, with no rates published.
- Meta AI news partners: a first wave announced 2025-12-05 (CNN, Fox News, Le Monde Group, People Inc., USA TODAY Co.) and an update on 2026-07-27; Meta mentions links in answers and neither training nor money.
What Publishers' Filings Say About AI Licensing Revenue
No publisher has filed a per-deal price for an AI license, but several filings show how AI licensing moves their revenue. Five filers give the clearest dated lines:
- The New York Times (10-K for FY2025): licensing revenue up USD 14.4 million in 2025, "largely related to commercial agreements with third-party digital platforms"; generative AI litigation costs of USD 13.3 million in 2025 and USD 10.8 million in 2024.
- People Inc. (Q3 2026 release): licensing and other revenue up 49 %, partly from Meta.
- Reach plc (H1 2026 report, 2026-07-22): "secured deals with large tech firms such as AWS" and TollBit pay-per-use revenue; its Google pilot partnership was announced 2026-10-12.
- Gannett (Q2 2025 release, 2025-07-31): its CEO calls the Perplexity deal "fair value of our content", with no amount.
- News Corp (Q1 FY2027 release, 2026-11-04): its CEO writes "We are pleased to have partners in OpenAI and Meta".
Le Monde calls its OpenAI deal "une source significative de revenus supplémentaires" without a figure. Amounts attributed to single deals appear only in press reports and are not stated here; see reported payouts to data sellers for how reported and confirmed amounts differ.
Who Can Sign for Newsroom Content
A publisher signs for what it owns, and ownership differs between staff work, freelance pieces and licensed-in agency copy.
Staff Journalism and Work Made for Hire
Articles written by employees within the scope of their jobs belong to the publisher in the United States and, by default, in the United Kingdom. Under 17 U.S.C. 201(b), the employer is the author of a work made for hire and owns all rights unless both parties sign a written instrument that says otherwise, and 17 U.S.C. 101 covers work by an employee within the scope of employment.
In the United Kingdom, section 11(2) of the Copyright, Designs and Patents Act 1988 makes the employer first owner of an employee's work made in the course of employment, subject to any agreement to the contrary. A rights audit sorts owned from licensed-in content.
Freelance Pieces, Wire Copy and Syndicated Content
Freelance contributions, agency wire copy and syndicated articles are usually licensed into the publisher, so its right to relicense them for AI training depends on each contract. Under 17 U.S.C. 101 a commissioned piece is a work made for hire only in one of nine listed categories with a signed written agreement.
Whether a publisher may relicense freelance or syndicated copy for AI training without new consent is decided by the contract and any agency terms, and no general rule is stated here.
Personal Data Inside a News Archive
News archives name people, including health and criminal proceedings, so licensing an archive for training is also a personal-data decision in the EU. The Italian data protection authority (Garante) addressed this in provvedimento n. 741 of 2024-11-27 on the GEDI and OpenAI agreement of 2024-09-24, citing special-category and judicial data.
The decision is a formal warning under Article 58(2)(a) GDPR about a possible breach of Articles 9, 10, 13, 14 and Chapter III, not a fine or a ban. The rules on GDPR and AI training data apply to the archive as they do to any other dataset.
Licensing Routes for Publishers Without a Direct Lab Deal
Publishers without a direct lab deal can charge crawlers per request, license through platforms that sit between them and AI companies, join a collective license or answer a data program's publisher listing. The table compares six routes as of 2026-11-10, alphabetical, with each value the party's own statement checked 2026-11-09. Platforms and collectives are intermediaries, not buyers; Mercor is the licensee under its own program.
| Route | Role | What the publisher grants or controls | Stated payment terms | Source |
|---|---|---|---|---|
| Cloudflare pay per crawl | network intermediary, Merchant of Record | allow, charge or block per crawler (HTTP 402) | site owner sets a flat per-request price; monthly payouts; closed beta; fees not published | developers.cloudflare.com |
| Kopinor and the National Library of Norway | collective license (Kopinor for member organizations) | non-exclusive license to train Norwegian and Sami language models on press publications up to one year from publication (s.2); no specialized news-service models; opt-out (s.16) | NOK 45,000,000 for 2026, indexed from 2027 (s.17) | agreement PDF, kopinor.no |
| Mercor publisher listing | company-data program (licensee) | long-form text, courses or video; hundreds of titles or hours; not short articles alone | Mercor states USD 35,000 to 370,000 per company, 10 to 500 employees | Mercor listings API |
| ProRata | attribution platform, intermediary | content for AI answers; "never used for model training" | "depends on the relevant rights, licenses, and commercial agreements"; no percentage | prorata.ai |
| RSL Collective | collective license via robots.txt, intermediary | paid license permits ai-train; attribution license permits ai-train and ai-input | royalties when an AI application uses the content to generate a result; rates not published | rslcollective.org |
| TollBit | publisher platform, intermediary (not a party to licenses) | publisher sets permitted uses and rates; warrants rights and indemnifies TollBit | "Free for publishers"; publisher sets rates | tollbit.com |
TollBit states in its agreement of 2025-10-29 that it is not a party to the content license agreements, and the publisher warrants its rights and indemnifies TollBit. See content licensing platforms for publishers.
Suing Instead of Licensing: Publisher Claims Against AI Developers
Some publishers sued instead of licensing, and as of 2026-11-09 none of the US news cases has a merits ruling on training. The table lists five cases by filing date with docket status as of 2026-11-09.
| Case | Court and docket | Filed | Status as of 2026-11-09 | Source |
|---|---|---|---|---|
| The New York Times Company v. Microsoft Corporation et al. | S.D.N.Y. 1:23-cv-11195-SHS-OTW; MDL No. 3143 | 2023-12-27 | summary-judgment stage, no ruling located | MDL docket |
| Daily News LP et al. v. Microsoft Corporation et al. | S.D.N.Y. 1:24-cv-03285-SHS; MDL No. 3143 | 2024-04-30 | summary-judgment stage | MDL docket |
| Dow Jones & Company et al. v. Perplexity AI | S.D.N.Y. 1:24-cv-07984-KPF | 2024-10-21 | discovery; motion to compel granted 2026-10-29 | complaint and docket |
| Advance Local Media LLC et al. v. Cohere Inc. | S.D.N.Y. 1:25-cv-01305-CM-OTW | 2025-02-13 | discovery | complaint and docket |
| USA TODAY Co., Inc. et al. v. OpenAI Foundation et al. | S.D.N.Y. 1:26-cv-08892 | 2026-10-08 | filed; no ruling; 19 publications; complaint seeks damages in excess of USD 250 million | complaint |
The New York Times, Chicago Tribune and CNN cases against Perplexity are undecided before Judge Preska as of 2026-11-09. The Advocate General's opinion of 2026-11-03 in Like Company v Google (C-250/25) is not a judgment. The full tracker of AI copyright lawsuits over training data carries every ruling.
Benefits, Risks and Downsides of Licensing News Content, and the Options Besides It
Licensing news content brings revenue and attribution terms, but it can grant rights that outlast the deal and weaken a later claim, and not licensing is valid.
Benefits
- Revenue: People Inc. reports licensing and other revenue up 49 % in Q3 2026.
- Attribution and links: the Washington Post and Condé Nast announcements name links and attribution.
- Terms the publisher defines: Microsoft states that publishers define terms.
Risks
- A grant can survive the term of the deal.
- Freelance and syndicated copy may lie outside the license.
- An archive license can raise personal-data duties in the EU, as the Garante warning shows.
- No per-deal price is public to benchmark an offer.
Downsides
- Program terms are set by the platform, and Google, Microsoft and Meta disclose no rates.
- A no-training deal such as Le Monde's limits what the license is for.
- A license can complicate a later claim, while litigation has its own costs.
Alternatives
- Reserve content in machine-readable form and block training crawlers.
- Charge per crawl, for example through Cloudflare pay per crawl.
- Join a collective license, described under collective licensing for AI training.
- Litigate, as the cases above show, or not license at all.
Steps From Archive Audit to a Signed Publisher License
A publisher reaches a signed license in six steps, and the first two decide how much of the archive it can offer.
- Inventory owned and licensed-in content.
- Check freelance, agency and syndication contracts.
- Screen for special-category and judicial data.
- Decide training, display or both, and set the crawler policy to match.
- Compare written terms from more than one buyer, platform or program on grant, term, attribution, deletion and payment basis.
- Sign, then monitor usage reports.
Grant and payment basis differ by deal, from Microsoft's delivered value to TollBit's publisher-set rates. The broader menu is in alternatives to selling data to AI labs.
Where Publisher Licensing Meets Copyright Law, Crawlers and Other Sellers
Whether a publisher needs to license at all depends on copyright exceptions that differ by country, and in the EU on whether it reserved its content in machine-readable form. Under Article 4(3) of Directive (EU) 2019/790, in force since 2019-06-06, a reservation for online content must be machine-readable. See copyright and AI training across jurisdictions.
Blocking stops future collection by an obeying crawler but does not remove copies already collected. Operators distinguish AI crawlers that train from those that only fetch for answers: OpenAI's GPTBot is a training crawler, while OAI-SearchBot and ChatGPT-User are not. The voluntary GPAI Code of Practice of 2025-07-10 asks signatories to follow robots.txt.
News publishers are one entry among every type of seller. Journals fall under academic and scientific publishers, and trade titles under book publishers and authors licensing books.
- Grant, term and attribution clauses in AI data licensing agreements.
- What happens to models after the term: retraining and model retention rights.
- Reserving content under the EU text and data mining exception.
- The US defense: fair use and AI training.
- Charging crawlers: pay-per-crawl and crawler licensing.
This page is information, not legal advice.
Deal, Case and Platform Pages Named on This Page
Each deal, case and platform above has its own page with parties, dates and sources.
- Buyers and programs: frontier AI labs as data buyers, Mercor data buyer profile, TollBit.
- Cases: News publishers v. Perplexity, The New York Times v. Microsoft and OpenAI, USA TODAY Co. v. OpenAI.
- Deals: GEDI and OpenAI, Google News AI pilot, Kopinor, Microsoft Publisher Content Marketplace, News Corp.
Frequently Asked Questions
Do AI companies pay news publishers for training data?
Some do: six publisher deals with OpenAI state training, while most newer deals and the Google, Microsoft and Meta programs describe display or grounding. Amounts are not disclosed in any of them.
How much do AI companies pay publishers for content?
No party has published a per-deal price for a news license; filings show only revenue lines such as the New York Times' USD 14.4 million licensing increase in 2025.
Valuation methods are set out under what data is worth to AI companies.
Does blocking AI crawlers stop training on a publisher's archive?
Blocking a training crawler stops future collection by that crawler, but it does not remove copies already collected or bind crawlers that ignore robots.txt.