DataPacts

AI-Generated Data as Training Data: Provider Terms, Distillation and Resale

Checked

At a glance

model-generated data, meaning model outputs, chat logs with AI and synthetic text, can be offered to AI developers only to the extent that the provider's terms for the model that produced it, the holder's rights in the inputs and the buyer's own requirements allow, and the providers' written terms differ from one another.

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

Five provider documents read on 2026-11-17 each assign or disclaim ownership of outputs and each bar using the service to build competing models. None of them addresses licensing a dataset of ordinary outputs to a third party; Google's terms bar reselling or training on Grounded Results from its Search grounding service.

This page covers model outputs offered as AI training data licensing material: the clauses, one provider's statement on distillation, the Copyright Office's January 2025 conclusions and what buyers state about generated material.

How the quotations were taken is set out in how we source, date and check every figure. The OpenAI pages returned HTTP 403 on 2026-11-17, so OpenAI is quoted from archived copies of 2026-11-16 (Terms of Use) and 2026-11-10 (Services Agreement).

Three Kinds of Model-Generated Data a Holder Might Offer

A holder may have three kinds of model-generated data: outputs it generated itself by prompting a model, logs of its own users' conversations with an AI product, and a person's own history of chats with a consumer chatbot. The three differ in whose terms apply and who holds the inputs.

  • Outputs the holder generated itself: the terms of the model used govern, as the next section shows.
  • Logs of a company's AI product with real users: Mercor asks for these in its AI Product listing, and the logs carry both the company's user terms and the terms of the models behind the product.
  • A person's own chat history: the consumer terms of the chatbot apply, and no buyer program for such history is on record as of 2026-11-17.

How synthetic data is generated and how good it is are questions for the page on synthetic data in AI training, which also covers the labs' own use of synthetic data. This page asks only whether outputs and logs already held can be offered.

Output and Competition Clauses in Provider Terms: What Five Documents Say

Each of the five documents read on 2026-11-17 assigns or disclaims ownership of output and bars using the service to build competing models, but the wording and scope differ.

OpenAI's individual terms bar using Output to develop models that compete with OpenAI, its Services Agreement adds a defined exception, Anthropic's commercial terms bar access to build a competing product or service including competing AI models, and Google's Gemini API terms bar using the Services to develop models that compete with the Services.

The table quotes the short clauses in alphabetical order by provider, as read on 2026-11-17.

Document (effective) Who owns output Competition clause Source
Anthropic Commercial Terms of Service (2025-06-17) s.B: Customer "owns its Outputs"; Anthropic assigns "its right, title and interest (if any)", subject to compliance s.D.4: Customer may not "access the Services to build a competing product or service, including to train competing AI models" anthropic.com/legal/commercial-terms
Anthropic Consumer Terms of Service (2025-10-08) s.4: "we assign to you all our right, title, and interest (if any) in Outputs", subject to compliance s.3: no use "To develop any products or services that compete with our Services, including to develop or train any artificial intelligence or machine learning algorithms or models" anthropic.com/legal/consumer-terms
Google Gemini API Additional Terms of Service (2026-03-23) "Use of Generated Content": "Google won't claim ownership over that content" "Use Restrictions": "You may not use the Services to develop models that compete with the Services (e.g., Gemini API or Google AI Studio)" aistudio.google.com/docs/terms
OpenAI Services Agreement (2026-01-01; archived copy of 2026-11-10) s.4.1: Customer "owns all Output"; OpenAI assigns its rights "if any"; s.4.4: other users' responses are "not considered Customer's Output" s.3.3(e): "except for a Permitted Exception, use Output to develop artificial intelligence models that compete with OpenAI's products and services" openai.com/policies/services-agreement (archive)
OpenAI Terms of Use (2026-01-01; archived copy of 2026-11-16) "Ownership of content": the user owns the Output; assignment "if any"; the assignment "does not extend to other users' output or any Third Party Output" "What you cannot do": "Use Output to develop models that compete with OpenAI" openai.com/policies/terms-of-use (archive)

The Services Agreement defines a Permitted Exception as using Output to develop models "primarily intended to categorize, classify, or organize data (e.g., embeddings or classifiers), if these models are not distributed or made commercially available to third parties", and to fine-tune or customize models provided through OpenAI's fine-tuning or other Services listed on its Pricing Page. The individual Terms of Use read contain no such exception.

Which document governs depends on the account type, consumer, API or enterprise customer, and the account used must be checked, because the dates above differ and a provider can change its terms. Google's consumer Gemini terms, Anthropic's usage policy, OpenAI's service-specific terms and open-weight model licenses were not read.

What the Provider Clauses Leave Open for a Seller

The clauses read do not settle whether a dataset of outputs can be licensed to an AI developer, because they assign rights to the account holder "if any", define no competing model in the sentences quoted, and say output is not unique.

A text search of the saved pages found "resell" in connection with the Services and, in Google's terms, with Grounded Results, which may not be resold, syndicated or trained on; the page draws no wider conclusion.

  1. The assignment is of "right, title, and interest (if any)", so what the account holder receives depends on what the provider held.
  2. Output is non-exclusive: OpenAI says output "may not be unique", and Google says it "may generate the same or similar content for others", so no promise of exclusivity can rest on the output itself.
  3. "Competing" is not defined beyond the product references (OpenAI's "products and services", Google's "the Services"), and only OpenAI's Services Agreement lists exceptions.
  4. Each restriction binds the account holder, who cannot pass on a permission it does not have.
  5. Effective dates differ from 2025-06-17 to 2026-03-23, and a clause read today may not be the clause in force when a license is signed.

Buyer warranties and seller disclosures both turn on which model made the data and under which account, which belongs in the field of use in the license.

Distillation: What a Provider Has Stated Itself

Anthropic stated on 2026-02-23 that distillation, "training a less capable model on the outputs of a stronger one", is "a widely used and legitimate training method" that can also be used for illicit purposes, and that three labs generated over 16 million exchanges with Claude through approximately 24,000 fraudulent accounts in violation of its terms of service and regional access restrictions.

The table records Anthropic's post "Detecting and preventing distillation attacks", published 2026-02-23 and read 2026-11-17; every value is Anthropic's own statement.

Topic Anthropic states Source
Definition distillation "involves training a less capable model on the outputs of a stronger one"; "a widely used and legitimate training method"; labs "routinely distill their own models" anthropic.com/news
Scale "over 16 million exchanges with Claude through approximately 24,000 fraudulent accounts"; DeepSeek over 150,000, Moonshot AI over 3.4 million, MiniMax over 13 million exchanges anthropic.com/news
Basis "in violation of our terms of service and regional access restrictions"; Anthropic does not currently offer commercial access to Claude in China anthropic.com/news
Response classifiers and behavioral fingerprinting, sharing of indicators, stronger verification, and safeguards "designed to reduce the efficacy of model outputs for illicit distillation" anthropic.com/news

This is one provider's statement about access through fraudulent accounts. No court filing or statement by another provider on distillation was read for this page, and the page does not say how any court would treat the conduct described.

Rights in Model Outputs: The Copyright Office's January 2025 Conclusions

The US Copyright Office released Part 2 of its AI report on 2025-01-29 and concluded that outputs of generative AI can be protected by copyright "only where a human author has determined sufficient expressive elements", which "can include" a perceptible human-authored work, creative arrangement or modification of the output, "but not the mere provision of prompts".

The release (NewsNet Issue 1060) adds that using AI to assist in creation, or including material made by AI in a larger human work, "does not bar copyrightability". Only the release was read for this page, not the report body.

A provider's assignment of "right, title, and interest (if any)" and the Office's conclusion are different things: the first moves whatever rights the provider had, and the second concerns whether an output is a protected work at all.

A seller should therefore not describe unedited outputs as protected works in a warranty. The Office's separate report on training, covered under Part 3 on training and fair use, is a different question.

What Buyers Say About Model-Generated Content

Buyers that state a position on model-generated content mostly turn it away: five of the 26 listings in Mercor's feed on 2026-11-17 name synthetic, scripted, role-played or AI-written material as not a fit (a sixth excludes human "Scripted read speech"), and Shutterstock's contributor FAQ says it will not allow contributors to submit content generated by AI for licensing.

The table lists the entries alphabetical by buyer; each value is the buyer's own statement, and Mercor states no reason for any entry.

Buyer and listing or page What it states about generated material Source
Mercor, AI Product Company not a fit: "Synthetic or scripted traffic"; scoping question "Which models power your product?" api.data.mercor.com
Mercor, Design & Web Agency not a fit: template sites and sites generated by AI (paraphrased from the listing) api.data.mercor.com
Mercor, Digital Health Company not a fit: "Role-played or scripted visits" api.data.mercor.com
Mercor, Medical Practice or Health System not a fit: "Synthetic or sample charts" api.data.mercor.com
Mercor, Software Company not a fit: "Student, hackathon or side projects, or code mostly written by AI tools" api.data.mercor.com
Shutterstock, Contributor FAQ on content made with its AI tool and other AI generators (undated) states it "will not allow" such content to be submitted by contributors for licensing on its platform (paraphrased from the FAQ) submit.shutterstock.com

Mercor's listings were published 2026-10-17 and re-read 2026-11-17. Shutterstock gives as reasons that contributors must "prove IP ownership" and that ownership of content generated by AI "cannot be assigned to an individual".

The research on why such material is less valuable, and Mercor's other human-made entries, belong to the page on model collapse and the value of human data.

One listing asks for the opposite of scripted material. Mercor's AI Product listing asks for complete session logs with prompts, responses and every tool call from real users, and for user terms that allow licensing de-identified usage data or a way to get consent.

Mercor states USD 20,000 to 480,000 for the whole deal, "Depends on session volume and completeness", checked 2026-11-17; it is a claim and not a price per record. Session logs are covered under agent trajectory data, and the listing and program terms sit in Mercor's data buyer profile.

Offering Model-Generated Data: Benefits, Risks, Downsides and Alternatives, Including Not Offering

Offering model-generated data can monetize outputs and logs the holder already has, but the provider's terms, the holder's rights in the inputs and buyers' stated preference for human-made records limit what can be offered, and a buyer cannot be given a permission the holder lacks.

Benefits

  • The holder does not depend on a scarce human-made record.
  • A company's own AI-product logs match a named ask in Mercor's listing.
  • The quoted clauses assign output ownership to the account holder, "if any" of the provider's rights.

Risks

  • Breach of a competition clause if the outputs are used to build a competing model.
  • Inputs or logs that contain personal or third-party data, which a rights audit sorts out before any offer.
  • Warranting copyright in unedited outputs, given the Office's human-authorship conclusion.
  • Rejection by buyers that list synthetic material as not a fit.

Downsides

  • No exclusivity can rest on output that providers say may not be unique.
  • No public price for such datasets is on record as of 2026-11-17.
  • Logs cannot be shared where user terms do not allow it, and Mercor asks for terms that allow de-identified licensing or a way to get consent.
  • Terms change, and effective dates differ.

Alternatives

  • Not offering at all: keep the data and grant no license, as set out in the guide on whether to sell or license your data.
  • License human-made records instead.
  • Use the outputs internally only.
  • Ask the buyer for its position in writing before any sample moves.

Each alternative leaves the provider's terms as the first document to read.

Neighboring Data Types and the Seller Steps for Model-Generated Data

This page covers model outputs, AI chat logs and synthetic text; human-written text corpora, images, code and design files each have their own page.

The steps for model-generated data follow the general process:

  1. Run a data sale readiness check, noting which model made which data and under which account.
  2. Record the model, account type and terms version in training data provenance files.
  3. Read the clauses of AI data licensing agreements before agreeing to any warranty on the data.

Frequently Asked Questions

This page is information, not legal advice.

Can I sell outputs I generated with ChatGPT, Claude or Gemini?

The quoted terms assign or disclaim ownership of output and bar competing models, and none read addresses licensing a dataset of ordinary outputs. The account's own terms and the buyer's position decide, and both should be read before any offer.

Do the providers' terms let me train another model on their outputs?

Each document bars competing models in its own words. OpenAI's Services Agreement lists two exceptions, and Gemini's bar covers "the Services (e.g., Gemini API or Google AI Studio)".

The Copyright Office concluded on 2025-01-29 that human authorship of expressive elements is required. Prompts alone are not enough, in the Office's words.

How much do AI companies pay for model-generated data?

No price for such datasets is on record. Mercor's range of USD 20,000 to 480,000 (checked 2026-11-17) is its claim for whole-deal session logs. Methods for other data are set out in what company data is worth to AI companies.