Can You Sell AI-Generated or AI-Assisted Data for AI Training?

Data a model generated rarely sells to AI companies. The buyers of company records and code screen it out, and anything a model can make, a lab can make for itself at a cost of cents. Work that people made with AI tools helping is a different matter: it still sells, as long as you can show which parts people made and which parts a model wrote.

The buyers say so on their own pages. Turing takes only pull requests by a human author, and no LLM-written code. Collinear lists human-written code among what raises its price, and AI-generated code among what it does not need, and Abaka AI names handwritten code from 2023 or earlier as its best fit. Scale AI wants years of records from systems in daily use, not demos or synthetic data. Meanwhile a research pipeline generated coding tasks for about $0.05 a task with no person writing any of them, while tasks made to a buyer's order sell for a median of $2,000 per task.

This guide is for any company whose records now carry AI's work: code written with a coding assistant, documents drafted in a chat tool, support replies sent by an AI agent, calls placed by an AI voice, a task set built with a model's help, or a dataset a model produced outright. It covers why buyers pay for what people made, what each buyer says, the three kinds of AI-touched data and how each sells, the AI vendors' own terms that come with their output, who owns what a model wrote, and how to find and mark the AI-made parts before you offer anything. It sits under what AI labs buy, beside the guides to selling a codebase and to RL environments.

Contents

• Why AI companies pay for what people made
• What each buyer says about AI-made data
• Generated, assisted or mined from real work
• The AI vendor's terms travel with its output
• Who owns what a model wrote
• Finding the AI-made parts of your records
• A worked example
• Questions people ask
• What this means for you

Why AI companies pay for what people made

An AI company buys data to give its models something they cannot get elsewhere. The labs describe their inputs in their own reports: Anthropic, for one, lists public web text, non-public data from third parties, work by data-labelling services and paid contractors, data from users who opted in, and data it generates itself. Only one of those inputs is bought from a holder. The rest the lab collects, commissions or makes.

That last input is the reason generated data struggles to sell. Generation is the buyer's own tool. micro1 says it invests in data generation models to make datasets and RL environments dramatically less expensive and more abundant, and the firms that buy company records describe what they do with them in the same terms: Collinear says the environments it builds from a company's data are expanded synthetically from what you send. A holder who sells generated data is selling the buyer something its own pipeline already makes, and makes more cheaply every quarter.

Anything a model can make, a lab makes for cents

The published costs of generating training data are tiny next to the prices paid for people's work. The figures below are the costs that research teams and vendors reported for producing data with models, each from its own paper or page.

Across every cost-to-produce figure on record for tasks, the median is $37 per task a task. Against that, the people who write tasks are paid a median of $260 per task a task, and tasks made to a buyer's order sell at a median of $2,000 per task. Each figure links to the prices behind it on the market rates page. The gap is the buyer's whole reason to pay a person: a task that only a person could write, about a problem that really happened, is worth what a generated one is not.

Training on generated data wears models down

There is a second reason, and it comes from research rather than price lists. A paper in Nature in July 2024 found that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models: the rarer parts of what the original data held disappear, generation after generation. The authors called the effect model collapse, and concluded that data about genuine human interactions will become more valuable as generated text fills the web.

The finding has a limit worth knowing. A follow-up study found that keeping generated data alongside the original real data, rather than replacing the real data with it, avoids the collapse (the same record holds both papers). Generated data is therefore usable in a mix. But the mix needs a base of real data underneath, and that base is what a lab cannot make for itself. That is what a holder of real records is selling.

The supply of that base is shrinking relative to demand. Epoch AI estimates the stock of public text written by people at around 300 trillion tokens, and expects models to use it up between 2026 and 2032 if trends continue. Lab leaders say the same in plainer words: OpenAI's former chief scientist told a conference in December 2024 that "we've achieved peak data". What has not been published is what people have not yet shared, and much of it sits inside companies.

What a buyer pays for instead

The buyers of company records put their reason in one sentence, and two of them use nearly the same words. Mercor and Appen both say that public and synthetic data cannot replicate how work is really done inside a company. Labelbox's Prism programme says synthetic data still depends on real-world grounding.

Vendors that collect data for labs now sell the absence of generated data as a feature. Liva AI, a voice and video collector, says it records real, consented human voice and video in house, with no synthetic data. Kled licenses video, music, transcriptions and other human-made data to labs. When the suppliers advertise "human-made" as the product, a holder whose records are human-made holds the thing the market is short of.

What each buyer says about AI-made data

Most buyers' pages say nothing about AI-made data at all. The ones that do are mostly code buyers, because code is where AI assistants have spread fastest and where a buyer can test what it receives. The table gathers every statement on record, each from the buyer's own page. Fuller terms for each buyer are on the licence terms page.

Some buyers make the synthetic part themselves, from your real data, and that does not make your data "generated". Appen removes personal data with a synthetic rewrite where the workflow has to survive the redaction, Centific rewrites synthetically to keep a record's structure, and micro1 says some datasets may undergo synthetic rewrites. The rewrite is the buyer's de-identification step. What you supply is still the human record, and the price is set on it. The guide to de-identification covers how those rewrites work and what to check in the sample.

No buyer on record publishes a discount for a share of AI-assisted work. The published rules are a screen, not a sliding scale: code a model wrote is out, code people wrote is in, and the date of the history is the quickest proxy a buyer has. Expect a buyer that learns of AI-written material after the fact to treat it as a misdescription. micro1's terms say that if it decides in good faith that the data manifest was materially inaccurate or misleading, nothing is payable and you refund what was paid, even after it accepted the data.

Generated, assisted or mined from real work

"AI-generated" covers three different things, and buyers treat them differently. What decides which kind a set is, is where its substance came from: who had the problem, who solved it, who said the words, who took the picture. Whether a model helped tidy, format or describe it matters much less.

Generated: sold as a product, priced as one

A set is generated when a model invented its content: synthetic conversations, generated images, tasks a model made up together with their answers, trajectories an agent produced on its own. Such sets do sell, but through a different door. They sell as catalogue products, priced per unit and sold to many, on the data catalogues and listing sites described in the guide to AI training data marketplaces.

The prices show where they sit. One listing offers 551 verified coding-agent trajectories for $100; another sells about 3.5 hours of synthetic talking-head video for $130. Generated data earns real money only where generation is the point, such as test images that must not show real people: one vendor lists 100,000 AI-generated passport images from $30,000 for face-recognition work. A buyer of company records or code does not buy any of these, and a valuation built on the prices paid for people's work has nothing to price them against.

Assisted: people's work, with the AI parts marked

Most of what companies hold now is assisted: people did the work and an AI tool helped. A developer wrote a feature with a coding assistant suggesting lines; a manager drafted a policy in a chat tool and rewrote it; a support agent sent a reply an AI drafted and edited it first. The problem, the decision and the context are human. Some of the words or lines are a model's.

Assisted records still sell, because what a buyer of company records pays for is the record of how work was done, and that record is human. But the AI-written parts are discounted or cut, most strictly in code, where Turing and Collinear screen them out. The practical answer is to know which parts are which, say so in the datasheet and the inventory, and offer the human parts first. US copyright law draws the same line, as the section on ownership below shows.

Mined from real work: tasks built from what people did

A third kind sits between the two, and it matters most to anyone selling coding or agent tasks. A task is mined when its problem and its solution come from real work, and a model only helped package it: the bug really happened, the fix is the commit that fixed it, the failed attempts are the real ones, and a model drafted the task statement or the grader from those materials.

Researchers have built benchmark task sets this way for years, SWE-bench among them, from real issues and the pull requests that resolved them, and buyers pay people to check such sets. Mercor has offered $70 to $90 an hour to auditors of tasks of this kind, AfterQuery pays $75 for each task that is fully approved, and Prime Intellect offered $3,500 for one environment built on SWE-bench Verified. Review cuts deep: Terminal-Bench 2.0 drew 229 tasks from 93 contributors and kept 89 after review.

A mined set counts as people's work, because what a model cannot make (the real problem, the real fix, the real failures) is people's. No buyer on record states a rule for a task whose statement or grader a model drafted, so the safe course is to give a buyer what review would check anyway:

1. Say what a model drafted. List, per task, whether the statement, the grader or the tests were drafted by a model.
2. Have a person check every grader. It should fail on the starting state and on the real failed attempts, and pass on the real fix.
3. Keep the provenance. Link each task to the commit, ticket or session it came from, so a buyer can trace it. The guide to training data provenance covers what to record.
4. Add what buyers measure. A container that runs each task, and how current models score on it, are what task buyers ask for; the RL environments guide lists them.

The AI vendor's terms travel with its output

When a person writes with an AI tool, the output comes with the tool's terms, and those terms say something about building AI models. Every major AI vendor's terms that we read bar using its service, or its output, to build models that compete with it. A licence to an AI company that trains its own models is exactly the case those clauses were written for.

The same vendors also hand the output to the user. OpenAI assigns its rights in Output to you, "if any", and Anthropic assigns its rights in Outputs to the customer, if any, subject to its terms. Owning the output and being free to train someone else's model on it are two different questions, and the vendors answer the second one no whenever the model competes with theirs.

We know of no court ruling on how far these clauses reach. Whether licensing your stored output to a buyer is your own "use" of it to develop a competing model, or the buyer's, is a question the terms do not settle, though Anthropic's wording expressly covers supporting a third party's attempt. What is clear is the direction: the AI company buying your records may be one of the vendor's competitors, and the clause is the vendor's protection against exactly that. The cautious course is to keep a vendor's output out of a licence to an AI company, or to ask the vendor first. This is the same kind of check the CAD guide describes for design software and the game data guide for game engines.

The second OpenAI clause cuts closer still. Under OpenAI's terms for individual users, describing ChatGPT's text as human-made, to a buyer or anyone, breaks your own terms with OpenAI, on top of being a misdescription to the buyer. That makes knowing which parts a model wrote a matter of your own contracts, not only of price.

Who owns what a model wrote

A licence grants rights the holder has, and for content a model generated on its own, there may be few to grant. In the United States copyright needs a human author. The D.C. Circuit held in March 2025 that the Copyright Act requires all work to be authored in the first instance by a human being, upholding the Copyright Office's refusal to register a work listed with a machine as its author.

The same court said what that does not mean: the human-authorship rule does not prohibit copyrighting work made by or with the assistance of AI. The Copyright Office's January 2025 report draws the line in two sentences: where AI merely assists an author, the output stays copyrightable, and content entirely generated by AI cannot be protected. Writing the prompt is not enough; the Office concluded that prompts alone do not make the user an author. Between the two ends, each work is judged on its facts.

For a holder, that has three consequences:

• Assisted work keeps its value as property. Code and documents people wrote with AI help keep copyright in what the people wrote, though not in passages a model generated on its own, and the licence works as it would for any company records.
• Purely generated content carries no copyright to license. What you can sell is access and a contractual promise, not an exclusive right. That is one reason an exclusive licence of generated data is worth little: nobody can stop the buyer, or anyone else, from generating the same thing.
• Buyers ask you to warrant ownership. SimpleClosure's terms, for example, have the company warrant that it owns the assets or holds all the rights to license them, and cover its losses from a breach. Shutterstock refuses AI-generated images from contributors partly for this reason: their ownership cannot be assigned to one person.

These are US rules; other countries decide authorship under their own law. The rights guides cover the other questions a licence raises, and the guide to data licensing agreements covers the warranty clauses themselves. This guide is research from public sources, not legal advice.

Finding the AI-made parts of your records

Marking AI-made material is easier than it sounds, because most of it carries traces: dates, authors, tool names and labels the systems themselves add. The aim is not to prove a negative about every line. It is to give a buyer an honest, checkable account of where AI tools were used, so the human parts can be sold at their full value.

Each kind of record has its own traces:

Two things make this account worth more. The first is a date: a buyer that cannot check every line can check when the work was done, which is why Abaka AI writes its rule as a year. The second is consistency: an account that says the same thing as the records, with the agent pull requests where it says they are, is what lets a buyer trust the rest. The guide to what AI buyers check shows how buyers test what a holder says, and the rights check asks about AI-made material among its questions on what you may license.

A worked example

This example is illustrative. A software company of about 40 people has built products since 2018. Its developers began using coding assistants in 2024, and from 2025 some pull requests were opened by a coding agent. Its wiki holds a few years of design documents, some drafted in ChatGPT and rewritten. Its help desk added an AI agent in 2025 that answers routine tickets. One engineer also built a set of coding tasks from the team's own bug fixes, with a model drafting each task's statement.

Code. The history from 2018 to 2023 is human-written and is the most wanted part: it meets Abaka's stated line and raises Collinear's price. The history from 2024 on is assisted. The company lists the agent pull requests and leaves them out, keeps the human-written pull requests from those years, and says plainly in the inventory when assistants came into use. Turing's script would refuse the agent pull requests anyway.

Documents. Design documents a person wrote stay in. Documents that a chat tool wrote whole come out of a licence to an AI company: they add little, and the company's own terms with OpenAI bar using that output to develop competing models, which a licence to an AI company may count as. Documents drafted with a chat tool and then substantially rewritten stay in and are described as such.

Support. Human agents' replies stay; the AI agent's replies come out, using the help desk's own record of which replies the agent sent. The customer's side of every conversation is human either way.

Tasks. The bug-fix tasks are mined from real work. The engineer lists which statements and graders a model drafted, checks that each grader fails on the starting commit and passes on the real fix, links each task to its commit, and adds a container and current models' pass rates. Described that way, the set is sold as people's work, which is what it is.

The result is a smaller lot than the company's raw total, and a better one: everything in it is what it says it is, and the part buyers want most is clearly labelled at the front.

Questions people ask

Do AI companies buy synthetic data? Rarely from outside, because they make it themselves: micro1 says it is building data generation models to make datasets cheaper and more abundant, and generated tasks cost about $0.05 each to produce. Synthetic sets sell mainly as catalogue products, at catalogue prices, through the data marketplaces.

Can I sell AI-generated images for AI training? Not to Shutterstock, which licenses its library to labs: it accepts no AI-generated content from contributors, in part because no one person can be assigned its ownership. Generated images sell only where generation is the point, such as test images that must not show real people. The photos guide covers what sells.

Is code written with GitHub Copilot or Claude worth anything to AI labs? The code people wrote with an assistant suggesting lines is still their code, but the buyers that state a rule screen out code a model wrote: Turing takes no LLM-written code, and Collinear does not need AI-generated code. History from before assistants were in daily use is worth most; mark agent and bot pull requests and lead with the rest.

Does AI-assisted data count as human-made? Where people supplied the substance and wrote the words that matter, yes. The US Copyright Office's view is that where AI merely assists an author, copyrightability does not change, though copyright covers only what people wrote. Say which parts a model wrote, because a buyer that finds out later may treat it as a misdescription.

Can I sell my ChatGPT conversations for AI training? Your own side of them is yours, but OpenAI's terms bar you from using its output to develop models that compete with OpenAI, the business agreement says the same, and a licence to an AI company may count as that use. Anthropic's and Google's terms draw the same line for their services. Keep the vendor's output out of a licence to an AI company unless the vendor agrees.

Can a model-built task set be sold if the problems are real? Yes, if the problems, solutions and failures come from real work and a model only drafted the packaging. Buyers pay people to audit such tasks ( $70 to $90 an hour at Mercor), so give them the audit's answers up front: what a model drafted, a person's check of each grader, and the commit each task came from.

Why do AI companies want human data at all? Because models trained indiscriminately on generated output degrade: a Nature study found that indiscriminate use of model-generated content causes irreversible defects. Generated data works in a mix only with real data underneath, and the stock of public human text is expected to run out between 2026 and 2032.

What this means for you

AI tools in a company's records are normal now, and they do not make the records unsellable. What lowers a lot's value is not knowing which parts a model made, or describing them wrongly. Five steps put you on the right side of that:

1. Date your records. Find when AI tools came into daily use in each system, and treat the history before that date as the core of what you offer.
2. Mark the AI-made parts. Agent and bot pull requests, documents a model wrote whole, bot replies, AI-voiced calls, generated media: list them, and leave them out of a licence to an AI company.
3. Keep vendors' output out, or ask first. OpenAI's terms bar using its output, and Anthropic's and Google's bar using their services, to build competing models.
4. Describe it honestly. Put the account in your datasheet; a buyer's terms can take back what it paid for a misdescribed lot.
5. Value the human part. Value your data on the part people made.

AIData.Broker acts as the holder's agent: the holder itself licenses its data to the buyer, and only the holder pays us, a fee fixed in the mandate and due only if a sale closes. We run sales for holders based in the United States, the United Kingdom and Canada. For the wider picture, see how to sell company data to AI companies, how to prepare data for sale and how to prove data quality without sharing it.

This guide is research from public sources, not legal, tax or investment advice.