Who Sells Data to AI Companies: Every Type of Seller and Their Options

Data for AI training is sold or licensed by anyone who holds rights in records or content an AI developer wants: companies, publishers, platforms, creators, collectives, public institutions and individuals. This index sorts them into 18 seller types in six groups as of 2026-11-09, each with one dated example from a filing, docket, statute or the seller's own page.

Seller types differ in who owns the rights and who may sign, the first question in AI training data licensing for any holder, whether it later licenses, sells or declines.

Who sells data to AI companies?

Six groups of sellers license data to AI companies, and they differ less in the data they hold than in who owns the rights and who may sign. The six groups are listed below, each with one dated example.

1. Companies: Tempus AI licensed a de-identified oncology dataset to Pathos AI for USD 200 million over three years (Tempus 10-K, fiscal 2025).
2. Publishers: News Corp and OpenAI announced a multi-year partnership on 2024-05-22, value not disclosed (News Corp).
3. Platforms and nonprofits: Reddit's S-1 of 2024-02-22 reports USD 203.0 million in aggregate contract value for its January 2024 data licenses, with counterparties it does not name.
4. Creators, artists, stock libraries and music rights holders: Shutterstock issued its first AI dataset earnings to contributors in Q4 2022.
5. Collectives and public institutions: Kopinor and the National Library of Norway agreed NOK 45,000,000 for 2026 (agreement signed 2025-12-19).
6. Individuals: Kled takes an exclusive, perpetual and irrevocable license for AI uses of uploads (terms revised 2026-07-21).

Brokers, shutdown services and pooling platforms are not sellers in this index, because they place a holder's data with buyers and do not own it.

Every seller type at a glance

The index lists 18 seller types, each with what it licenses and one dated example with its stating party, as of 2026-11-09. Rows follow the six groups above, not size or fees.

The seller type says who may sign, while the data type decides which buyers want the records; data types have their own section.

Sellers, intermediaries and buyers: which side of a deal you are on

A seller holds the rights, an intermediary places the data, and a buyer trains on it or resells it. SimpleClosure, for example, sells or licenses the code and workspace data of winding-down startups on their behalf, so the startup stays the seller. Buyers are listed under who buys AI training data, every buyer and program compared without fees.

The table compares the six seller groups by rights holder, signer and the rule to check first, as of 2026-11-09.

In chapter 11 the trustee or debtor signs with court approval, and a collective signs for its members.

What each seller group licenses, with dated examples

Each group below names the rights question first and a dated example second, in the order of the seller-type table.

Companies selling records they created while operating, closing or in court

Operating companies, startups, closed companies, acquired companies and estates sell the same kind of asset, workplace and product records, but the person who may sign changes at each stage. Operating companies sign through their officers, often with one of the company data partnership programs that license chat, email, tickets and code.

Three filed or listed examples follow.

• Tempus AI's Neuro order with Pathos AI, disclosed 2026-10-22, is worth up to USD 55.0 million, mostly at USD 100 per delivered record for 500,000 records over five years (Tempus 10-Q).
• Mercor's Game Studio listing states USD 100,000 to 1,000,000 to the company (published 2026-10-17, checked 2026-11-09).
• Wiley acquired Emerald Publishing on 2026-06-01 for a preliminary USD 462.7 million, "to strengthen its proprietary content advantage" in AI (Wiley 10-Q).

After closure, the wind-down party signs. Keelan Miskell, founder of a defunct crypto startup, states net proceeds of roughly USD 25,000 for its code base (MIT Technology Review, 2026-10-13). SpaceX's compute and option agreement with Anysphere (Cursor), dated 2026-04-19 (SpaceX 424B4), covers developer interaction data but states no data price.

In insolvency, the estate signs and the court approves. At the Spirit Airlines auction of 2026-08-14, Google LLC made the successful bid of USD 10,000,000 and Mercor.io Corporation the alternate bid of USD 7,500,000 for de-identified data (Bankr. S.D.N.Y. 25-11897 (SHL), Dkt. 1463), a sale not approved as of 2026-11-09. The Character.AI stalking horse bid filed 2026-10-13 excludes training data and models (Bankr. D. Del. 26-11603 (BLS)). How trustees sell data assets in insolvency is covered separately.

Newsrooms and book, journal and textbook houses as licensors

Publishers license archives and live content, and many of their deals are display or grounding deals in which training is not mentioned. The publisher signs, but what it may license depends on its contracts with authors and journalists.

Four publisher examples follow.

• AP and OpenAI: OpenAI licenses part of AP's text archive (AP, 2023-07-13).
• Wiley disclosed GenAI content rights projects of USD 23 million (2024-06-13) and USD 21 million (2024-09-05), buyers unnamed.
• Informa reports AI-related revenues of USD 75m+ for 2024 at Taylor and Francis, as non-recurring data access revenue.
• HarperCollins licensed its nonfiction backlist, opt-in per author, to "an unknown AI company" (Authors Guild, 2024-11-19).

Training is stated for Axel Springer, Le Monde with OpenAI, Prisa, Dotdash Meredith, Vox Media and the FT, but not for TIME, the Guardian, the Washington Post, Hearst, The Atlantic or Condé Nast, and Le Monde's Perplexity deal of 2025-05-14 excludes it. Microsoft's Publisher Content Marketplace, announced 2026-02-03, pays publishers on delivered value for grounding. Each publisher deal is logged among AI data licensing deals with parties, dates and sources.

Online communities and nonprofit projects licensing what users wrote

Platforms license content their users posted, usually through a paid API, and the users are not parties to the deal. Reddit's 10-Q for Q3 2026 states that substantially all its licensing contract value comes from two counterparties whose agreements are approaching the end of their terms, and Google announced its Reddit Data API agreement on 2024-02-22 without a value.

Four further examples follow.

• Stack Overflow states 83M+ Q&A available for licensing (checked 2026-11-09); its OpenAI partnership of 2024-05-06 carries no published value.
• Automattic announced on 2024-02-27 that public posts on WordPress.com and Tumblr go to select AI companies it does not name, excluding sites that opted out.
• Wikimedia, LLC, a subsidiary of the Wikimedia Foundation, sells access through its paid Enterprise API and named five new AI customers on 2026-01-15.
• Yelp's shareholder letter dated 2026-02-12 states that it "recently signed" an agreement with OpenAI, value not disclosed.

Works by creators, artists, photo libraries and the music industry

Creators and artists sell either directly, through a platform that pools their work, or through a stock library or label that licenses on their behalf. The route decides who negotiates: a library or label signs under its contributor or artist contracts, while a pooling platform licenses under terms each creator accepts.

Payout figures are each platform's own claim, and a single platform can contradict itself. Troveo's two pages, both checked 2026-11-09, state different totals (shown in the seller-type table), and no buyer confirms either figure.

Four library, platform and music examples follow.

• Getty Images states in its 10-K for fiscal 2025 that it compensates creators on a recurring basis for content used as training data, with no published per-contributor rate.
• Wirestock states a 50 % creator commission on commission-based projects (blog, 2026-07-13).
• UMG and Udio settled and signed licenses on 2025-10-29, and UMG and ElevenLabs announced a multi-year agreement on 2026-09-10, values not disclosed (UMG).
• Warner Music Group states in its 2026 letter that "all of our deals are based on variable economics".

Collectives, research institutions, memory institutions and public bodies

Collectives sign for many rights holders at once, while universities, libraries and governments often publish data openly or keep commercial use subject to permission. GEMA has offered an AI licensing model since September 2024, and STIM signed its first opt-in collective AI license, covering training and output, in September 2025 with a licensee it does not name. The Kopinor agreement is a payment by a public institution, not by a commercial AI developer.

Research and memory institutions mostly control access. The Institutional Data Initiative at Harvard Law School Library requires commercial users of Institutional Books 1.0, including for AI model training, to contact it, and it released 1.47 million Boston Public Library newspaper pages on 2026-09-02. US schools releasing de-identified student records follow 34 CFR 99.31(b)(1), which requires removal of all personally identifiable information.

Public bodies in the EU follow the Open Data Directive (EU) 2019/1024, whose recital 55 excludes documents in which third parties hold intellectual property. Article 12 of the DSM Directive (EU) 2019/790 allows extended collective licensing with rightholders free to exclude their works, and health data under Regulation (EU) 2025/327 reaches AI developers from 2029-03-26 only through health data access bodies. These EU rules on AI training data, including the TDM exception, are set out by instrument.

People licensing their own photos, video and voice

Individuals sell their own media and voice through apps and marketplaces whose terms decide how long the license lasts and whether it is exclusive. Three platforms show the range, all checked 2026-11-09.

• Kled, operated by Nitrility, Inc., states 500,000+ users, and its exclusive license appears in the seller-type table.
• ElevenLabs' Voice Library accepts Professional Voice Clones only, with a notice period of 30 days to 2 years.
• Mozilla Data Collective states that uploaders "keep 100% of revenue".

Selling one's own data differs from paid contributor work such as rating, writing or recording to a brief, which is a job rather than a license. That trade is covered under human data for AI, with contributor work and pay.

Benefits, risks, downsides and alternatives for any seller, including not selling

Selling or licensing data can earn money from records that already exist, but a license to train usually cannot be undone once a model is trained.

The three benefits are listed below.

• Revenue from existing archives, as the filed Tempus values show.
• Control through negotiated scope, term and deletion clauses.
• Recurring income where pay follows delivery, as in the Tempus Neuro order.

The four risks are listed below.

• Consent: Ifeoma Ajunwa of Emory University states that employers should have express consent before selling workers' data (MIT Technology Review, 2026-10-13).
• Privacy: the Spirit sale requires de-identification under Cal. Civ. Code 1798.140(m).
• Third-party rights: archives hold works owned by others.
• Litigation: Bartz v. Anthropic (N.D. Cal. 3:24-cv-05417) ended in a USD 1.5 billion settlement, with final approval on 2026-07-20 (Dkt. 680).

Disputes are tracked among AI copyright lawsuits over training data, with rulings and settlements.

The three downsides are listed below.

• Little visibility: EU AI Act training content summaries disclose licensed data only as Yes or No plus modality (Art. 53, Regulation (EU) 2024/1689, from 2025-08-02).
• Creditors first: insolvency proceeds go to the estate.
• Lasting terms: licenses such as Kled's are perpetual and irrevocable.

The four alternatives, including not selling, are listed below.

• Not selling or licensing at all.
• Opting out: YouTube's third-party training setting is off by default, and YouTube states it "isn't facilitating payments" to creators (checked 2026-11-09).
• Licensing through a collective such as STIM.
• Publishing openly under CC0 or CC BY 4.0.

The guide on how to sell or license company data to AI companies covers each step and treats declining as a valid outcome.

How this seller index is built: dated examples, stating parties, no fees

Every example names who stated it and when: a filing, a docket entry, a statute or the party's own page, never a press estimate. Press-only deal values, such as figures reported for News Corp and OpenAI, are left out.

Payout and listing figures appear as the party's own claim with the date they were checked, and contradictions stay visible. DataPacts charges the company nothing for an enquiry or introduction and takes no cut of any sale.

The rules on how we source, date and check every figure apply to every row of this index.

Where sellers go next: buyers, process, law and data types

This index answers who sells; the pages below answer how to sell, under which law, to whom and at what price.

• Law: AI training data law by jurisdiction.
• Market: the AI training data market, with sizes and prices.
• Terms: payouts and program terms compared.
• Funding: how DataPacts is funded.
• Sourcing: how AI companies get training data, from crawling to synthetic data.
• Ethics: ethics and risk of selling data for AI training.

Frequently asked questions

Can a company that has shut down still license its data?

Yes, through the closing company, its wind-down party or its estate. SimpleClosure states that 350+ companies offered assets on AssetHub since its April launch (blog 2026-09-03), selling on each startup's behalf. The defunct-company row in the seller-type table covers this route.

Has a court approved a bankrupt company's data sale to an AI buyer?

Not in the Spirit Airlines case as of 2026-11-09. Google LLC holds the successful bid (Dkt. 1463), and the hearing is set for 2026-11-17 (Dkt. 1779). The 23andMe sale to a nonprofit was not an AI training license; court sales are tracked under bankruptcy estates selling data assets.

Do individuals get paid for AI training data?

Some platforms pay individuals, on terms and figures each platform states. ElevenLabs' blog headline of 2024-10-28 reads "We crossed $1M in Voice Library Payouts". Paid rating and writing work is a job, compared among AI training jobs and platforms.

Who gets paid when a publisher licenses books for AI training?

The publisher, and the author where the contract shares the income. For HarperCollins, the Authors Guild states USD 5,000 per title split 50-50 with the opted-in author (2024-11-19). The Guild also recommends that authors receive 75 to 85 % of AI training license income, an advocacy position.

Is a data seller the same as a data broker?

No, a data seller holds rights in its records or content, while consumer data brokers trade personal profiles and are not covered here. Intermediaries such as SimpleClosure place a holder's company data with buyers without owning it.

How much do sellers receive?

No rate card exists, and every figure is the stating party's own filing or claim. Filed values and court bids differ by seller type, as the examples above show. Methods and public price points are set out under what company data is worth to AI companies.