DataPacts

Text Corpora for AI Training: What Buyers Pay For

Checked

At a glance

AI developers pay for text when they need rights, freshness or a corpus they cannot crawl: newspaper archives, books, scholarly articles, forum and Q&A posts and reference text, usually under a license rather than a sale.

Independent reference. Every buyer is covered on the same terms, every figure carries its source and check date, and no buyer pays to be listed or placed higher.

Request an independent introduction

Tell us what data your company holds. We reply with the buyers that fit and what their terms leave out. Your details reach a buyer only after you agree to a named introduction.

IndependentYour details stay with us until you agreeNot selling is a valid outcome

Few of these deals state a value: Reddit's S-1 (filed 2024-02-22) states USD 203.0 million aggregate contract value for arrangements with unnamed licensees, and the Authors Guild states USD 5,000 per title for HarperCollins's backlist license (2024-11-19).

This page compares text as one family within AI training data licensing: the kinds of text, what a license buys, the values on record, the programs that take text, and whether to license at all. Not licensing is a valid outcome.

The DataPacts Editorial Team explains how we source, date and check every figure.

Kinds of Text AI Developers License

Text reaches AI developers in eight kinds, each with its own holders and deals: news, books, scholarly articles, forum and Q&A posts, reference text, reviews, parallel and multilingual text, and course content.

The table gives one example for each kind, as of 2026-11-16, plus three separate data types.

Kind of text Typical holder Example on record (date, stating party) Source Page
Books publishers, authors HarperCollins nonfiction backlist, opt-in per author (2024-11, Authors Guild) authorsguild.org Books
News archives newspapers, wire services OpenAI licenses part of AP's text archive (2023-07-13, AP) ap.org News archives
Scholarly articles and research academic publishers Wiley GenAI content rights projects (2024, Wiley filings) sec.gov Scholarly articles
Forum and Q&A posts platforms, communities Stack Overflow OverflowAPI with OpenAI (2024-05-06, Stack Overflow) stackoverflow.co Forum posts
Reference text encyclopedias, wikis Wikimedia Enterprise API partners (2026-01-15, Wikimedia Enterprise) enterprise.wikimedia.com n/a
Reviews and local data review platforms Yelp agreement with OpenAI, "recently signed" (letter 2026-02-12, Yelp) sec.gov n/a
Parallel and multilingual text translation firms, language bodies Kopinor collective license for Norwegian language models (2025-12-19, agreement); multilingual text is a separate page kopinor.no Translation text
Course and education content publishers, education companies Mercor publisher or education listing (checked 2026-11-09, Mercor) Mercor listing feed Education content
Legal text (separate data type) law firms, legal publishers privilege and the ROSS ruling are on the legal page n/a Legal text
Workplace documents (separate data type) operating companies covered on the documents page n/a Documents
Code (separate data type) software companies Pangea per-token rates for code (checked 2026-11-09, Pangea) pangea.ai Code

Workplace text and code have their own pages, and the parent page compares every one of the data types AI companies buy.

How Much Text Models Train On, and the Open Corpora a License Competes With

Frontier models train on trillions of text tokens, most of it crawled or taken from open corpora, so a paid text license competes with a free baseline.

SpaceXAI's training content summary for Grok 4.7 (2026-09-21) states more than 10 trillion text tokens and declares commercial licensing agreements with rightsholders for text and image.

Epoch AI (2024-06-06) estimates about 300 trillion tokens of effective public human-generated text (90 % interval 100 to 1,000 trillion), fully used between 2026 and 2032 (80 % interval); this is an estimate.

The table lists open corpora with the size each publisher states, as of 2026-11-16, in alphabetical order.

Corpus Stated size Publisher, date Source
Common Corpus 2,267,302,720,836 tokens (card); the paper says about two trillion Pleias, first release November 2024 huggingface.co/datasets/PleIAs/common_corpus
Common Pile v0.1 8 TB from 30 sources EleutherAI and co-authors, 2025-06-05 arxiv.org/abs/2506.05209
DCLM-Pool 240 trillion tokens DataComp-LM card, checked 2026-11-09 huggingface.co/datasets/mlfoundations/dclm-baseline-1.0
Dolma 3 Mix about 6 trillion tokens Allen Institute for AI card, checked 2026-11-09 huggingface.co/datasets/allenai/dolma3_mix-6T
Institutional Books 1.0 983,004 volumes, 242 billion tokens Institutional Data Initiative, 2025-06 arxiv.org/abs/2506.08300
Institutional Newspapers (Boston Public Library) 1.47 million pages, 16.3 billion tokens Institutional Data Initiative, 2026-09-02 institutional.org

The open baseline has limits: Longpre et al. (2024-07-20) find that about 5 % or more of all C4 tokens and 28 % or more of the most critical sources were fully restricted by robots.txt (2023 to 2024), and 45 % of C4 by terms of service, across about 14,000 domains. The corpora, their sources and licenses are set out in major AI training datasets.

What a Text License Buys: Training Rights, Real-Time Access or Display

A text deal buys one of three things: a right to train on an archive, ongoing access to fresh text through an API, or the right to display and quote text in answers, and many announced deals buy only the last.

The table groups announced deals by what the buyer gets, as of 2026-11-16.

What the buyer gets Examples on record Payment basis on record Source
Training right over an archive or catalog Axel Springer and OpenAI ("advancing the training"); Dotdash Meredith and OpenAI ("used to improve OpenAI's large language models"); Wiley content rights projects; HarperCollins backlist one-time project fee (Wiley); per title (HarperCollins); otherwise not disclosed axelspringer.com; sec.gov; authorsguild.org
Ongoing access to fresh text through an API Reddit Data API (Google: "display, train on, and otherwise use"); Stack Overflow OverflowAPI; Wikimedia Enterprise on-demand, snapshot and realtime APIs aggregate contract value over two to three years (Reddit S-1, licensees not named); others not disclosed blog.google; sec.gov; stackoverflow.co; enterprise.wikimedia.com
Collective license for model training Kopinor and the National Library of Norway; CLA Generative AI Solution NOK 45,000,000 a year (Kopinor); CLA fees not published kopinor.no; cla.co.uk
Display, quotation and grounding in answers TIME, Guardian, Washington Post, Hearst, Atlantic, Condé Nast with OpenAI; Le Monde with Perplexity; Microsoft Publisher Content Marketplace; Google News AI pilot usage-based "paid on delivered value" (Microsoft); "we pay participating news publications" (Google); amounts not disclosed about.ads.microsoft.com; blog.google; lemonde.fr

Training or model improvement is stated for Axel Springer, Le Monde (with OpenAI), Prisa, Dotdash Meredith, Vox Media, the Financial Times, Google and Reddit, while for TIME, the Guardian, the Washington Post, Hearst, the Atlantic, Condé Nast and the Meta partners only display, or nothing on training, is stated.

Training is explicitly excluded for Le Monde's agreement with Perplexity, which "ne prévoit pas l'utilisation de nos contenus pour entraîner un LLM". How licensed text differs from scraped text is compared in licensed vs scraped training data.

Text Deals and Revenue Lines With a Value on Record

Few text deals carry a stated value, and the ones that do come from filings, a collective agreement or an authors' organization, not from the buyers.

The table lists every value on record with its stating party, as of 2026-11-16, in each value's own unit.

Parties Date Text licensed Value on record (stating party) Source Page
Reddit, Inc. and unnamed licensees 2024-01 (S-1 filed 2024-02-22) posts and comments via access, search and analysis USD 203.0 million aggregate contract value, two to three years; minimum USD 66.4 million expected in 2024 (Reddit S-1) sec.gov Reddit
John Wiley & Sons, Inc. and an unnamed "large tech company" disclosed 2024-06-13 book and learning content rights for training USD 23 million (Wiley) sec.gov Wiley
John Wiley & Sons, Inc. and an unnamed "large tech company" disclosed 2024-09-05 content rights for training USD 21 million project; Q1 amount USD 16 million (release) vs USD 17 million (presentation) (Wiley) sec.gov n/a
Taylor & Francis (Informa PLC) and unnamed AI buyers 2024 academic content and data access USD 75m+ AI-related revenues in 2024, "non-recurring" (Informa) investegate.co.uk (Informa RNS) Taylor & Francis
HarperCollins Publishers L.L.C. and an unnamed AI company 2024-11 nonfiction backlist, opt-in per author USD 5,000 per title, split 50-50 with the author (Authors Guild) authorsguild.org HarperCollins
Kopinor and the National Library of Norway 2025-12-19 Norwegian press publications up to one year from publication NOK 45,000,000 for 2026, index-adjusted from the 2027 remuneration (agreement s.17) kopinor.no Kopinor

Press reports carry other figures for text deals; none is repeated here, because no filing, agreement or party statement confirms them.

Some revenue lines mix licensing with other products and are not pure data licensing revenue. Reddit's "Other revenue" is USD 140.026 million for FY2025 and USD 47.961 million for Q3 2026 (10-K and 10-Q), and covers content licensing plus products sold directly to users.

Yelp states "Other revenue" of USD 74 million for 2025, including data licensing (letter 2026-02-12), and People Inc. states licensing and other revenue up 49 % in Q3 2026, driven by Apple News+, syndication and the Meta content partnership.

Wiley states FY2026 AI revenue of USD 49 million (release), with Research USD 33.1 million and Learning USD 16.0 million in the 10-K. Dated price points for every data type are collected under AI training data prices per work and token.

Programs and Intermediaries That Take Text

Outside direct lab deals, text moves through three kinds of channels: buyer programs that pay a holder, corpus distributors that license catalogs to developers, and collective or platform intermediaries that license on rights holders' behalf.

The table compares program terms as each program states them (checked 2026-11-09), as of 2026-11-16, grouped by role.

Program Role Text it takes Terms as stated (checked 2026-11-09) Source Page
Defined.ai Supplier Program buyer program (licensee and marketplace) text datasets among audio, image, video no revenue share published; no exclusivity, no minimums (s.3.5); Defined.ai states several partners earn over USD 1 million a year (release 2026-01-27) defined.ai Defined.ai
Mercor publisher or education listing buyer program long-form text, courses or video; hundreds of titles or hours Mercor states a deal range of USD 35,000 to 370,000, "depends on catalog size and rights" Mercor listing feed Mercor
GSK (Gengo-Shigen-Kyokai) distribution service corpus distributor text corpora, speech data, lexicons non-members 33,000 yen per resource (buyer-side fee); provider compensation not stated gsk.or.jp GSK
Linguistic Data Consortium membership corpus distributor speech, text and lexical corpora catalog for-profit Standard USD 34,000, Subscription USD 40,000 (buyer-side fees); perpetual rights to data licensed during membership years ldc.upenn.edu LDC
CLA Generative AI Solution collective licensing body (licenses on opted-in rights holders' behalf) opted-in books, journals, magazines covers generative AI training, fine-tuning and RAG; "indemnity protection"; fees not published cla.co.uk n/a
CCC Annual Copyright License, internal AI re-use collective licensing body published content for internal AI uses internal-only AI re-use rights, launched July 2024; payout not published copyright.com n/a
Created by Humans intermediary for authors books training, reference and transformative rights; payouts not published createdbyhumans.ai Created by Humans
Microsoft Publisher Content Marketplace grounding marketplace premium publisher content for grounding publishers "will be paid on delivered value"; training not mentioned about.ads.microsoft.com Microsoft
ProRata attribution and Gist attribution platform publisher and creator content for AI answers "never used for model training"; payment "depends on the relevant rights, licenses, and commercial agreements" prorata.ai ProRata
RSL Collective collective license via robots.txt web content of opted-in publishers and creators paid license: royalties "whenever an AI application uses your content to generate a result"; rates not published rslcollective.org n/a

The LDC and GSK amounts are fees a corpus buyer pays, not payouts to a holder. Created by Humans, CLA, CCC, ProRata and RSL Collective act on rights holders' behalf and do not buy text.

Platforms that license publisher content for AI answers are compared under content licensing platforms for publishers.

What Text Buyers Ask For in a Corpus

Buyers ask for long-form, rights-cleared text in quantity, delivered in a structured format, and they exclude open-access text and content held under someone else's license.

Mercor's publisher listing (published 2026-10-17, checked 2026-11-09) asks for content the company has the rights to license, long-form text rather than short articles alone, and hundreds of titles or hours; it lists content distributed under someone else's license and content already open access as not a fit.

Stack Overflow states 83 million or more Q&A pairs for licensing and a free sample of 1,000 (checked 2026-11-09).

The HarperCollins license (Authors Guild, 2024-11-19) is a nonfiction backlist, opt-in per author, with outputs limited to 200 consecutive words or 5 % of a book.

The Kopinor agreement (2025-12-19, s.2) covers press publications "up to one year from the date of publication", so older archives sit outside it. Valuation methods for a corpus are covered separately.

Buyers ask for five things:

  • Rights record: proof that the holder can license for AI training
  • Volume: hundreds of titles or hours, or millions of Q&A pairs
  • Structure: Q&A pairs, XML or EPUB, API or snapshot delivery
  • Freshness: ongoing access, as in the Reddit and Wikimedia APIs
  • Exclusions: open-access text and content under another license

Who Can License a Text Corpus: Authors, Users and Reservations

The holder of a text corpus can license only the rights it holds: a publisher's archive often includes freelance and wire copy, a platform's corpus was written by its users, and an author may hold the AI right in a book.

Automattic states (2024-02-27) that only public content from WordPress.com and Tumblr sites that have not opted out is shared, that partners are updated about new opt-outs and asked to remove content from past sources, and it names no AI company and no value.

A website owner can reserve text from mining: Directive (EU) 2019/790 Article 4(3) allows a reservation "in an appropriate manner, such as machine-readable means in the case of content made publicly available online", and Article 53(1)(c) of the AI Act (Regulation (EU) 2024/1689, applying from 2025-08-02) requires general-purpose model providers to have a policy that identifies and complies with such reservations.

The scope of the Article 4 reservation for text online has its own page. The Kopinor agreement lets rights holders opt out with effect for future training (s.16).

An author may hold the AI right in a book separately from the publisher, which is why HarperCollins's license is opt-in per author.

Newsroom rights and platform user terms are covered on the seller pages below. This page is information, not legal advice, and a rights audit before licensing text records what the holder can license.

Text Disputes That Shape License Talks

Licensing and litigation run side by side for text: the same publishers that sign deals sue developers that trained without one.

News Corp's CEO Robert Thomson says in the Q1 FY2027 release (2026-11-04) "we will woo and we will sue", with OpenAI and Meta named as partners.

The Times case, The New York Times v. Microsoft and OpenAI (S.D.N.Y. 1:23-cv-11195-SHS-OTW, filed 2023-12-27), is part of MDL No. 3143 and pending at summary judgment as of 2026-11-09, with cross-motions filed 2026-09-04.

In Bartz v. Anthropic (N.D. Cal. 3:24-cv-05417), final approval of the USD 1.5 billion settlement came on 2026-07-20 (Dkt. 680); it pays for past acquisition of books, not a forward license.

Rulings and status for every case are on the tracker of AI copyright lawsuits over training data.

Licensing a Text Corpus: Benefits, Risks, Downsides and Alternatives, Including Not Licensing

Licensing text can earn from an archive that already exists, but a model trained on it cannot be untrained, and several disclosed text deals were one-time payments.

Benefits:

  • An existing archive can earn: Wiley states USD 23 million and USD 21 million for content rights projects (2024).
  • Authors can share: HarperCollins's USD 5,000 per title is split 50-50 (2024-11-19).
  • A collective agreement can pay a rights holder body: Kopinor's NOK 45,000,000 for 2026 (agreement of 2025-12-19).

Risks:

  • Revenue can end: Reddit states that its two main licensing agreements are approaching the end of their terms (10-Q filed 2026-10-30).
  • Revenue can be one-time: Informa calls its USD 75m+ of 2024 data access revenue "non-recurring".
  • Revenue can fall: Wiley's AI license revenue was USD 13.7 million in Q1 FY2027 against USD 28.9 million a year earlier (10-Q).

Downsides:

  • A model that has trained on text cannot be untrained: HarperCollins's terms limit outputs to 200 consecutive words or 5 % of a book rather than provide a recall (Authors Guild, 2024-11-19).
  • A training grant and a display grant are different, and most news deals state display only or nothing on training, so each deal's scope has to be read.
  • Open corpora set a free baseline: Common Corpus is openly licensed, while Institutional Books requires contacting its publisher for commercial use including AI training.

Alternatives:

  • Grant a display or grounding license only, as in the Google News AI pilot and Microsoft's marketplace.
  • Join a collective license, such as CLA's.
  • Reserve rights in machine-readable form (EU Article 4(3)).
  • Not license at all, as set out in the decision page on whether to license your text at all.

The rights held and the terms offered decide which option fits.

Text Sellers by Type and the Seller Steps for a Corpus

Each kind of text holder has its own page on rights, deals and steps, while this page compares the text family as a whole.

Seller pages by type of holder:

Seller steps for a text corpus:

  1. Read the complete seller guide for the whole process.
  2. Agree the license using every clause of a licensing agreement.
  3. Check how payment works, covered under upfront, milestones and revenue share.

The law that applies to a text corpus differs by country and is compared in AI training data law by jurisdiction.

Frequently Asked Questions

How much do AI companies pay for text data?

No rate card exists; values are stated deal by deal. Reddit's S-1 states USD 203.0 million aggregate contract value (2024-02-22), Wiley states USD 23 million and USD 21 million, and the Authors Guild USD 5,000 per title for HarperCollins (2024-11-19). Every deal is in AI data licensing deals.

Do AI companies pay to train on news articles?

Some news deals state training, and many cover display only. Axel Springer and Dotdash Meredith state training or model improvement, and Google states (2025-12-10) that it pays participating publishers for enhanced content. The terms are set out for the Google News AI pilot program.

Is text on the open web free to use for AI training?

Crawling is not a license: reservations, terms and copyright limit it, and the limits differ by jurisdiction. In the EU, Article 4(3) of Directive (EU) 2019/790 lets a holder reserve text. The wider position is covered in is web scraping for AI training legal.

Is there a per-word or per-token price for text?

None is on record for prose. The only per-token rates found are for code: Pangea states USD 100 per million tokens of source code (checked 2026-11-09).

Advisers and agents can use the introducer form.