OpenAI licenses part of AP's text archive (2023-07-13), Google receives an AP real-time feed for the Gemini app (2025-01-15), and the Kopinor agreement (2025-12-19) covers Norwegian press text for training.
This page sorts news as a data type within AI training data licensing by archive versus feed, deal scope, clearance, opt-outs and lawsuits. No deal on record states a per-article price, and not licensing is a valid outcome.
DataPacts Editorial Team dates every figure on this page and names the party that states it; how we source, date and check every figure sets out the rules.
What News Data Means for AI Developers
News data is more than article text: it covers the archive of past stories, the live wire or feed, the photos and video that run with them, and the metadata that dates and attributes each piece. The table lists six kinds as of 2026-11-16, each with one dated example from the stating party.
| Kind of news data | What it gives an AI developer | Example on record (date, stating party) | Source |
|---|---|---|---|
| Archive text | past reporting for training | OpenAI licenses part of AP's text archive (2023-07-13, AP) | ap.org |
| Current and archived titles | a publisher's portfolio for answers and products | News Corp titles including WSJ, The Times and The Australian; display and "enhance its products" (2024-05-22, News Corp) | newscorp.com |
| Live wire or feed | current facts for answers (grounding) | AP real-time feed for the Gemini app (2025-01-15, Google); AFP 2,300 wires a day in six languages for Le Chat (2025-01-16, Mistral AI) | blog.google; mistral.ai |
| National press under a collective license | press text for training national-language models | Norwegian press publications up to one year from publication; images, audio, video, magazines and weeklies excluded (2025-12-19, agreement s.2) | kopinor.no |
| Digitized historic newspapers | long-run text from library collections | Institutional Newspapers (Boston Public Library): 1.47 million pages, 16.3 billion tokens (2026-09-02, Institutional Data Initiative) | institutional.org |
| News photos and video | images and footage that run with stories | no news-specific deal on record; see the image and video pages | n/a |
Metadata such as bylines, dates and sections travels with the text, and no deal on record prices it separately. Each kind is one of the data types AI companies buy.
Archive Depth Versus Live Feeds: Why News Is Licensed Twice
An archive teaches a model how events were reported, while a live feed keeps answers current, so the same publisher can license its back catalog once and its new reporting on an ongoing basis. News Corp's agreement with OpenAI (2024-05-22) names both current and archived content of its titles.
Feeds serve a different purpose. AP delivers a feed of real-time information to enhance results in Google's Gemini app (2025-01-15, Google), and Mistral AI states that Le Chat integrates "AFP's daily production of 2,300 wires in six languages" so that responses are "grounded in reliable, up-to-date information" (2025-01-16). Neither mentions training.
Scope limits shape what an archive license covers. Kopinor and the National Library of Norway agreed a license for press publications "up to one year from the date of publication" (s.2). Mercor's publisher listing likewise asks for long-form text, courses or video, "not short articles alone" (checked 2026-11-09).
Training on News Versus Showing It in Answers
A news license grants training rights, display rights in AI answers, or both, and the parties' own words decide which; many news deals state only display.
Axel Springer's announcement of 2023-12-13 names both answers with attribution and links and "advancing the training" of OpenAI's models, while Le Monde states that its Perplexity agreement of 2025-05-14 "ne prévoit pas l'utilisation de nos contenus pour entraîner un LLM".
The same split runs through every kind of text, and text corpora for AI training sets out what each right covers. For news, the test is the grant in the parties' own announcement or contract, not the headline.
News Data Deals on Record, by Scope and Freshness
The news deals with a primary source differ less in price, which none discloses, than in what they cover: an archive, a live feed, a set of titles or a whole national press.
The table compares 14 deals as of 2026-11-16, with the use each party states (T training, D display or answers with training not stated, X training excluded, n/s not stated).
| Date | Parties | News data covered | Archive or live | Use stated | Value on record | Source | Page |
|---|---|---|---|---|---|---|---|
| 2023-07-13 | The Associated Press and OpenAI | part of AP's text archive | archive | n/s | not disclosed | ap.org | AP and OpenAI |
| 2023-12-13 | Axel Springer SE and OpenAI | selected global news including paid content; Politico, Business Insider, Bild, Welt | both | T and D | not disclosed | axelspringer.com | Axel Springer and OpenAI |
| 2024-05-22 | News Corp and OpenAI | current and archived content of named titles | both | D | not disclosed | newscorp.com | News Corp and OpenAI |
| 2024-09-24 | GEDI Gruppo Editoriale S.p.A. and OpenAI | news archive (editorial content) | archive | n/s | none on record | garanteprivacy.it | GEDI and OpenAI |
| 2024-10-01 | Microsoft (Copilot Daily) with Reuters, Axel Springer, Hearst Magazines, USA TODAY Network, Financial Times | "authorized content sources" for a news summary | live | D | not disclosed | blogs.microsoft.com | Microsoft and publishers |
| 2024-10-08 | Hearst and OpenAI | 40+ newspapers and 20+ magazine brands, with citations and links | both | D | not disclosed | hearst.com | Hearst and OpenAI |
| 2025-01-15 | The Associated Press and Google | real-time feed for the Gemini app | live | D | not disclosed | blog.google | n/a |
| 2025-01-16 | AFP and Mistral AI | 2,300 wires a day in six languages for Le Chat | live | D | not disclosed | mistral.ai | AFP and Mistral AI |
| 2025-05-14 | Le Monde and Perplexity | content to answer questions only | live | X | not disclosed | lemonde.fr | Le Monde and Perplexity |
| 2025-12-05 | Meta and news partners (first wave CNN, Fox News, Le Monde Group, People Inc., USA TODAY Co.; list updated 2026-07-27) | real-time news with links in Meta AI | live | D | not disclosed | about.fb.com | Meta and news partners |
| 2025-12-10 | Google News AI pilot (more than 400 publications as of 2026-10-15) | enhanced content for AI features in Google News | live | D | Google states it pays participating publications; amounts not disclosed | blog.google | Google News AI pilot |
| 2025-12-19 | Kopinor and the National Library of Norway | Norwegian press publications up to one year from publication | archive window | T | NOK 45,000,000 for 2026, indexed from 2027 (s.17) | kopinor.no | n/a |
| 2026-02-03 | Microsoft Publisher Content Marketplace (co-design publishers including AP, Condé Nast, Hearst Magazines, People Inc, USA TODAY Co., Vox Media) | premium publisher content for grounding Copilot | live | D | "paid on delivered value"; no rates | about.ads.microsoft.com | n/a |
| 2026-07-22 (disclosed) | Reach plc and AWS | news content; "secured deals with large tech firms such as AWS" (Reach H1 2026 report) | n/s | n/s | not disclosed | Reach H1 2026 report | Reach plc |
Revenue lines, signing rights and routes to a buyer are the publisher's side, covered by news publishers licensing content to AI companies.
What a News Archive Is Worth on the Record
No AI developer or publisher has published a per-article or per-year price for a news license, so the record holds one collective fee, one program range and a set of undisclosed amounts.
The collective fee is Kopinor's: NOK 45,000,000 for 2026 (s.17), paid by the National Library of Norway, not by a commercial AI developer and not per publisher.
The program range is Mercor's. Mercor states USD 35,000 to 370,000 for a publisher or education company, "Depends on catalog size and rights" (checked 2026-11-09, unchanged 2026-11-16), and Google states that it pays pilot publications "for the delivery and use of enhanced content" without amounts.
No other news deal on record carries a value from a party, so a seller has no public benchmark; what company data is worth to AI companies explains the valuation methods.
Clearing a News Archive Before Licensing It
Clearing an archive means checking who owns each article, whose personal data it holds and which rights reservations apply; this section is information, not legal advice.
Whose Articles Are in the Archive
A publisher can license the reporting it owns, while wire copy, freelance pieces and syndicated articles inside the archive depend on each contract.
Under 17 U.S.C. 201(b) an employer is the author of a work made for hire and owns all rights unless a signed written instrument says otherwise, and under CDPA 1988 s.11(2) an employer is first owner of a work an employee makes in the course of employment, subject to agreement.
Each outside piece needs its own contract check, which a rights audit organizes before any license is signed.
Names, Health and Court Reporting in Old Articles
A news archive names real people, including in health and court reporting, so in the EU licensing it for training is also a personal-data decision.
On 2024-11-27 the Italian Garante (provvedimento n. 741) issued a formal warning under Art. 58(2)(a) GDPR to GEDI over its OpenAI content-sharing agreement, stating that the processing could breach Arts. 9, 10, 13 and 14 and Chapter III; it was a warning, not a fine or a ban.
GDPR and AI training data sets out the rules on special-category and criminal-offence data.
The EU Press Publishers' Right and AI Answers
In the EU, news publishers hold a related right in their press publications, and the text and data mining exception covers that right as it covers copyright. Art. 4(1) of Directive (EU) 2019/790 extends the exception to the Art. 15(1) rights.
The Court of Justice case Like Company v. Google (CJEU C-250/25) tests this for training and chatbot answers.
Advocate General Szpunar's opinion of 2026-11-03 treats the training questions as inadmissible and adds that training is as a rule permitted under the exception, while answers that reuse press publications may infringe the press publishers' right in certain circumstances; it is an opinion, and the judgment was pending as of 2026-11-09.
Keeping News Out of AI Training Without Licensing It
A news site can keep its reporting out of training by blocking the training crawlers by name and, in the EU, by reserving its rights in machine-readable form, but neither step pays anything or recalls copies already collected. Three tokens matter most:
- GPTBot (OpenAI): OpenAI states that disallowing it indicates a site's content should not be used in training generative AI foundation models; OAI-SearchBot and ChatGPT-User are not training crawlers.
- Google-Extended: a robots.txt token, not a separate crawler, that controls use for training future Gemini models and for grounding, and that Google states does not affect inclusion in Google Search.
- Applebot-Extended: does not crawl; disallowing it opts content out of training Apple foundation models while pages remain eligible for Search (Apple, checked 2026-11-09).
In the EU, a reservation under the EU text and data mining exception must be machine-readable for online content (Art. 4(3) of Directive (EU) 2019/790), and the voluntary GPAI Code of Practice of 2025-07-10 asks signatories to follow robots.txt under RFC 9309.
The AI Act adds transparency: the copyright policy and training-content summary duties of Art. 53(1)(c) and (d) of Regulation (EU) 2024/1689 apply from 2025-08-02, and the summary template has non-SME providers list the top 10 % of domain names by content scraped.
Rightsholders can also act outside crawlers. Kopinor's agreement lets a rightsholder opt out with effect for future training (s.16), and ANJ states that on 2026-11-03 it, ABERT and ANER obtained a preliminary injunction in Brazil barring Meta from using journalistic content without authorization under its Commercial Terms.
Blocking stops future collection only by a crawler that obeys it, so opting out of AI training works across platform settings with the same limits.
What News Lawsuits Say About How Articles Were Used
News lawsuits against AI developers rest on three theories about the data: that articles were copied for training, that copyright management information was stripped from them, and that answer engines copy articles at query time. The table lists eight cases by filing date, with allegations as pleaded and status as of 2026-11-09.
| Case | Court and docket | Filed | Theory about the data (as alleged) | Status as of 2026-11-09 | Page |
|---|---|---|---|---|---|
| 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 | copyright infringement, DMCA 1202, misappropriation, trademark dilution | pending; summary-judgment stage, no ruling located | The New York Times v. Microsoft and OpenAI |
| The Intercept Media, Inc. v. OpenAI, Inc. et al. | S.D.N.Y. 1:24-cv-01515; MDL No. 3143 | 2024-02 | removal of copyright management information from articles used in training, 17 U.S.C. 1202(b) | pending in MDL 3143 | The Intercept v. OpenAI |
| Raw Story Media, Inc. and AlterNet Media, Inc. v. OpenAI, Inc. et al. | S.D.N.Y. 1:24-cv-01514-CM; 2d Cir. 25-1756 | 2024-02 | removal of copyright management information, 1202(b)(1) | dismissed for lack of standing; appeal briefed, no decision located | Raw Story v. OpenAI |
| 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 | same claims as The Times | pending; summary-judgment stage | Daily News v. Microsoft and OpenAI |
| Dow Jones & Company, Inc. et al. v. Perplexity AI, Inc. | S.D.N.Y. 1:24-cv-07984-KPF | 2024-10-21 | copying news for an answer engine; trademark claims | discovery; motion to compel granted 2026-10-29 | Dow Jones and NY Post v. Perplexity |
| Advance Local Media LLC et al. v. Cohere Inc. | S.D.N.Y. 1:25-cv-01305-CM-OTW | 2025-02-13 | infringement in training and retrieval-augmented outputs; trademark claims | discovery | Advance Local v. Cohere |
| Like Company v Google Ireland Limited | CJEU C-250/25 | 2025-04-03 (request lodged) | press publishers' right in training and chatbot answers | AG opinion 2026-11-03; judgment pending | n/a |
| USA TODAY Co., Inc. et al. v. OpenAI Foundation et al. | S.D.N.Y. 1:26-cv-08892 | 2026-10-08 | content of 19 publications used to train; DMCA 1202; damages in excess of USD 250 million sought | filed; no ruling | USA TODAY v. OpenAI |
No US news case has a merits ruling on training as of 2026-11-09, and every theory above is an allegation.
The Times, Daily News and The Intercept cases sit in In re OpenAI Copyright Infringement Litigation (MDL 3143), where Emmerich Newspapers v. Microsoft (S.D. Miss. 3:26-cv-00791) was conditionally transferred by order CTO-4 on 2026-10-27, with finality not checked.
Licensing News Content: Benefits, Risks, Downsides and Alternatives, Including Not Licensing
Licensing news can earn from an archive that already exists and keep reporting attributed in AI answers, but a training grant cannot be recalled and no public price exists to judge an offer.
Benefits
- Money from an archive that already exists, as in the Kopinor fee of NOK 45,000,000 for 2026.
- An ongoing feed, as in AP's real-time feed for Gemini.
- Citations and direct links to the publisher in answers, as Hearst's announcement of 2024-10-08 states.
Risks
- Personal data in the archive, which the Garante warning of 2024-11-27 shows regulators examine.
- No public benchmark, since press-reported news deal values are not primary-sourced.
- Scope limits, such as Kopinor's one-year window, that decide what is actually licensed.
Downsides
- A no-training deal limits the use: Le Monde's agreement is for answering questions only.
- The law is unsettled, as the pending CJEU judgment in C-250/25 shows.
Alternatives
- A display-only license, as in Google's News AI pilot and the Microsoft Publisher Content Marketplace.
- A collective license, as in Kopinor's agreement; see collective licensing for AI training.
- Reserving rights and blocking training crawlers.
- Not licensing at all.
News Beside Other Text, Images and Video, and the Seller Steps for an Archive
This page covers news as a data type; the wider text family, news photos and footage, and the publisher's own perspective have their own pages.
- Books as AI training data
- Selling photos and images for AI training (news photos)
- Selling video for AI training (news footage)
- Scientific data and papers as AI training data (journals)
- Forum and community archives
Content licensing platforms for publishers meter and place news for AI crawlers, AI crawlers lists every bot, the AI data licensing deals and AI copyright lawsuits over training data trackers hold the full record, and copyright and AI training across jurisdictions compares the law.
- Read how to sell or license company data for the full process.
- Remove or assess names and sensitive details, see PII removal and de-identification.
- Check grant, term and attribution against AI data licensing agreements.
- Decide with whether to sell or license your data, where not licensing is an outcome.
Frequently Asked Questions
Do AI companies pay for news archives?
Some deals pay, but few disclose amounts. Kopinor's NOK 45,000,000 for 2026 is paid by the National Library of Norway, and Google states that it pays participating pilot publications without giving amounts. Other deals on record state no value.
Can a news site stay in AI answers but out of training?
For OpenAI, yes, by token. GPTBot is the training crawler and OAI-SearchBot is not (see GPTBot and OpenAI's other crawlers), while Google-Extended covers grounding as well as training.
Does blocking Google-Extended affect Google Search?
No. Google states that the token Google-Extended "does not impact a site's inclusion in Google Search nor is it used as a ranking signal".
Are digitized historic newspapers used as AI training data?
Yes. Institutional Newspapers, with the Boston Public Library, holds 1.47 million pages and 16.3 billion tokens (post of 2026-09-02), and the post states no license terms. Open corpora are listed under major AI training datasets.