An FAQ on selling data to AI labs
Patronus AI teamAs of November 202610 min read
Patronus AI publishes this FAQ and licenses data for its own research. Every claim links its source, read on Nov 19, 2026.
AI companies license data they cannot easily get from the open web, or want in a structured, licensed form, such as news archives, online discussions, image metadata and, for agent training, private codebases and business records. Public deals include OpenAI with the AP, News Corp and the Financial Times, and Google with Reddit. Few prices are public, but Reddit’s Feb 22, 2024 S-1 reports data licensing arrangements with an aggregate contract value of $203.0 million over two to three years. Buyers set limits on third-party and personal data, and FTC staff have warned that quietly changing terms of service or a privacy policy to allow AI training on data already collected may be unfair or deceptive.
Snapshot as of November 2026. Last updated
What kind of data do AI companies buy?
AI companies buy data they cannot easily get from the open web, or want in a structured, licensed form. When OpenAI launched OpenAI Data Partnerships on Nov 9, 2023, it said it wanted “large-scale datasets that reflect human society and that are not already easily accessible online to the public today”, in any modality, and especially data that “expresses human intention (e.g., long-form writing or conversations rather than disconnected snippets)”. It named partners including the Icelandic Government and Miðeind for Icelandic language data and the Free Law Project for legal documents.
News archives and online discussion are two categories with public deals. OpenAI licensed part of the AP text archive (AP, Jul 13, 2023) and received access to current and archived content from News Corp publications (News Corp, May 22, 2024). Google gained access to Reddit’s Data API in a deal both companies announced on Feb 22, 2024. Shutterstock’s 2024 annual report says it has seen “increased demand for access to our metadata for machine learning and generative artificial intelligence model training.”
Agent training uses a newer category of data, the working records of a business. A Scale AI posting for a Data Acquisition Lead, Frontier Environments describes RL environments “built from the operating data of the companies that actually hold it”, and asks the hire to find which companies hold the data for the next domain. SemiAnalysis (Jan 6, 2026) wrote that demand for coding environments is high enough that it believes defunct startups are being acquired “specifically for the value of another private GitHub repo to make environments out of.”
Who buys training data from companies?
Model developers buy directly. OpenAI has signed content agreements with publishers such as the AP, News Corp and the Financial Times (FT, Apr 29, 2024), and runs its own data partnership program. Google licensed access to Reddit’s data. Each of these was announced by the companies involved.
Companies that build training data and environments for labs also buy source material. Scale AI’s Data Acquisition Lead posting describes a role that will “Own it from cold outreach to signature” and work with Scale’s legal team “to define the first version of how these transactions get priced.” SemiAnalysis reports that environment companies hire developers to replicate the interfaces of sites such as DoorDash and Uber Eats and sell the mockups to labs. Patronus AI licenses private codebases, company records and released games through data.patronus.ai, and its page says it licenses data directly for its own work.
How do data licensing deals with AI companies work?
In Google’s deal with Reddit, the buyer gets ongoing access through an API. Google’s post on the Reddit deal says Google now has access to Reddit’s Data API, which gives it “efficient and structured access to fresher information” that Google can “display, train on, and otherwise use”. In the AP deal, OpenAI licensed an archive. The AP release says OpenAI is “licensing part of AP’s text archive”, and the AP gets access to OpenAI’s technology and product expertise in return. The News Corp and FT deals cover display to users as well. News Corp says OpenAI may display content from its mastheads in response to user questions and use it to enhance its products, and the FT says ChatGPT users will see attributed summaries, quotes and links to FT journalism.
Contract terms are rarely public, but Reddit’s S-1 filing gives the length of its early agreements. It says the data licensing arrangements it entered in January 2024 have terms ranging from two to three years, and describes the business as “allowing third parties to access, search, and analyze data on our platform.” The filing also calls data licensing for AI training “a novel business model without an established track record”. OpenAI’s program offers two routes, an open-source archive that anyone can use and private datasets used to train OpenAI’s own models.
Patronus AI’s data partnerships page describes the steps of a deal with one training data buyer. The owner shares a few details, Patronus reviews a sample under NDA, and it makes a written offer with the price, what is included and how the full delivery will be checked. Its page gives a first answer within two business days and a sample review within five business days. After acceptance, it has up to 30 days to check the full dataset against the agreed criteria and pays on acceptance. If it does not proceed, it deletes the material and confirms in writing.
How much is a dataset worth to an AI company?
Few prices are public. The AP, News Corp, FT, Google and Reddit announcements cited here state no price. Some figures are public in securities filings. Reddit’s S-1 (filed Feb 22, 2024) says that in January 2024 it entered data licensing arrangements “with an aggregate contract value of $203.0 million and terms ranging from two to three years”, and expected at least $66.4 million of that to be recognized as revenue in 2024. The same filing says substantially all of the contract value came from one partner.
Reddit’s later 10-K for 2025 reports other revenue, which it defines as revenue from content licensing and products sold directly to users, of $140.0 million in 2025, $114.7 million in 2024 and $15.2 million in 2023. Shutterstock reports a Data, Distribution, and Services line that includes metadata licensing, its Giphy business and Shutterstock Studios. That line grew 28% to $175.3 million in 2024, and the filing credits growth in its data offering as a main driver.
For training tasks built from data, Epoch AI (Jan 12, 2026) reports that multiple interviewees agreed on a ballpark of $200 to $2,000 per task, and SemiAnalysis reports website replicas often cost about $20,000 each.
Do I need to own the rights to license data for AI training?
Buyers set limits on rights. OpenAI’s data partnership page says it is “not seeking datasets with sensitive or personal information, or information that belongs to a third party”.
Platform terms can also limit reuse. Reddit’s announcement says its Data API terms state that content accessed through the API “cannot be used for commercial purposes without Reddit’s approval.”
Licenses on data that is already shared are often wrong. The Data Provenance Initiative audited more than 1,800 text datasets and reported “license omission of 70%+ and error rates of 50%+” on widely used dataset hosting sites. Scale AI’s Data Acquisition Lead posting lists experience with data licensing, IP or regulated data among its preferred qualifications.
Should a data license for AI training be exclusive?
The deal announcements cited on this page do not say whether the licenses are exclusive, though some say what the deal leaves unchanged. Google’s post on the Reddit partnership says it “does not change Google's use of publicly available, crawlable content for indexing, training, or display in Google products”, and Reddit says API access stays free for non-commercial use under its published threshold.
Exclusivity raises the price in the related market for RL environments and tasks sold to AI labs. Epoch AI reports that two founders independently said exclusive deals there are “roughly 4-5x more expensive than non-exclusive ones.”
What about personal data and privacy when selling data to AI companies?
Personal data brings legal risk. In a Feb 13, 2024 post, FTC staff wrote that it may be unfair or deceptive for a company to start using consumers’ data for AI training and “to only inform consumers of this change through a surreptitious, retroactive amendment to its terms of service or privacy policy.” The post says a business that collects data under one set of privacy commitments “cannot then unilaterally renege on those commitments after collecting users’ data.” A company planning to license customer data should check what its privacy policy promised when the data was collected.
Buyers set their own limits too. OpenAI says it is not seeking datasets with sensitive or personal information and can help partners remove it. Reddit’s S-1 says its data licensing programs may implicate “complex and developing data privacy and data protection, misappropriation, and intellectual property laws, rules, and regulations.”
What happens to data after an AI company licenses it?
It depends on the license. Under OpenAI’s data partnership program, data in the open-source archive “would be public for anyone to use in AI model training”, while private datasets train OpenAI’s proprietary models and are handled with the access controls the partner prefers. Google says its Reddit access lets it display, train on and otherwise use Reddit content.
Data used to build training environments can be reused across model generations. SemiAnalysis reports that website replicas bought for agent training are a one-time purchase reused for future models, and that trajectories and logs from earlier training runs are fed back into various stages of training, such as mid-training. Patronus AI says it uses licensed material to build AI training and evaluation tasks and environments, that its customers may use the work it creates from that material, and that it does not resell the code or records themselves. The license sets out how the material may be used and shared.
Why are AI companies paying for data instead of scraping the web?
More of the open web is closed to AI crawlers than before. The authors of Consent in Crisis audited 14,000 web domains and found that in a single year, 2023 to 2024, new restrictions made about 5% or more of all tokens in the C4 corpus, and 28% or more of its most actively maintained critical sources, fully restricted from use. Counting terms of service, they found 45% of C4 restricted. They write that, if respected or enforced, these limits are “rapidly biasing the diversity, freshness, and scaling laws for general-purpose AI systems.”
Some data was never public. OpenAI’s program asks for data “not already easily accessible online”, and the Data Provenance Initiative found that commercially closed datasets dominate categories such as lower resource languages, creative tasks and newer synthetic data. SemiAnalysis reports that Google has lowered scraping rate limits on products such as Gmail, which makes them harder to copy for training environments.
What makes a codebase valuable for AI training?
Coding environments are built from a repository’s history and its tests. Epoch AI describes a common coding task as a git repository at a specific commit with a failing test suite, where the grader runs the tests and checks that they pass. SemiAnalysis describes how the public SWE-rebench pipeline picks pull requests that were resolved, merged to the main branch, described well enough to state the problem, touch more than one file, and add or change test files, since the tests are used for grading. It reports that the pipeline filtered 450,000 pull requests down to 21,336 valid tasks.
About this FAQ
Patronus AI publishes this FAQ. It licenses data for its own research and environment work, so it appears as one example among the buyers described. The format follows Epoch AI’s FAQ on RL environments.
This page is general information, not legal advice. The Sources list gives the full reference for every numbered source, read on Nov 19, 2026. To report an error, write to us through the contact page.
Sources
- 1 OpenAI, “OpenAI Data Partnerships”, Nov 9, 2023
- 2 Google, “An expanded partnership with Reddit”, The Keyword, Feb 22, 2024
- 3 Reddit, “Expanding our Partnership with Google”, Upvoted, Feb 22, 2024
- 4 Reddit, Inc., Form S-1 registration statement, filed with the SEC Feb 22, 2024
- 5 Reddit, Inc., Form 10-K for the year ended Dec 31, 2025, filed with the SEC Feb 6, 2026
- 6 The Associated Press, “AP, Open AI agree to share select news content and technology in new collaboration”, Jul 13, 2023
- 7 News Corp, “News Corp and OpenAI Sign Landmark Multi-Year Global Partnership”, May 22, 2024
- 8 Financial Times, “Financial Times announces strategic partnership with OpenAI”, Apr 29, 2024
- 9 Shutterstock, Inc., Form 10-K for the year ended Dec 31, 2024, filed with the SEC Feb 25, 2025
- 10 Longpre et al., “The Data Provenance Initiative: A Large Scale Audit of Dataset Licensing & Attribution in AI”, arXiv 2310.16787, Oct 25, 2023 (abstract read)
- 11 Longpre et al., “Consent in Crisis: The Rapid Decline of the AI Data Commons”, arXiv 2407.14933, Jul 20, 2024 (abstract read)
- 12 Federal Trade Commission, Office of Technology, “AI (and other) Companies: Quietly Changing Your Terms of Service Could Be Unfair or Deceptive”, Feb 13, 2024
- 13 Scale AI, “Data Acquisition Lead, Frontier Environments”, job posting, first published Aug 21, 2026, read Nov 19, 2026
- 14 AJ and Dylan Patel, “RL Environments and RL for Science: Data Foundries and Multi-Agent Architectures”, SemiAnalysis, Jan 6, 2026 (free portion read)
- 15 Jean-Stanislas Denain and Chris Barber, “An FAQ on reinforcement learning environments”, Epoch AI Gradient Updates, Jan 12, 2026
- 16 Patronus AI, Data partnerships, read Nov 19, 2026
Patronus AI · November 2026 snapshot · More resources