# How to buy AI training data

URL: https://sellmydatatoai.net/guides/how-to-buy-ai-training-data
Reviewed: 2026-11-08

To buy AI training data safely, define the exact training use first, then source from sellers who can prove they hold the rights, test a randomly drawn sample in your own pipeline and sign a license covering training, derived assets and deletion. Rights evidence matters more than volume, because data cannot practically be removed from a model once it has been trained on.

## Key takeaways

- You are almost always buying a license with defined rights, not the files outright.
- The Data Provenance Initiative found license omission rates above 70% and error rates above 50% on popular dataset hosting sites, so a license tag proves little on its own.
- Ask for a seeded random sample, an evidence pack with file hashes and a paid pilot before committing to a full purchase.
- Your license should cover derived assets such as embeddings, evaluation sets and fine-tuned weights, plus what happens if rights fail after training.
- Providers of general-purpose AI models selling into the EU face fines of up to 3% of worldwide turnover from August 2, 2026 if they cannot document their training content.

Most AI teams start buying data when public sources run out or stop being safe to rely on. The hard part is rarely finding data. It is finding data whose rights will stand up to scrutiny, and checking its quality before you pay.

This guide covers the full buying process, from writing a requirement to recording provenance after delivery. It draws on Nixon Peabody's October 2026 diligence alert, the Commission's guidance under the EU AI Act and published pricing from buyers and marketplaces.

## What does buying AI training data involve?

Buying AI training data almost always means buying a license, not the files themselves. The seller keeps ownership and grants you defined rights: which models may be trained, for how long, in which territories and what happens to derived assets such as annotations, embeddings and fine-tuned weights.

The distinction matters because training is hard to reverse. Daniel Goldberg of Frankfurt Kurnit Klein and Selz notes that once data has been used for training it cannot practically be pulled back out. Every check you skip before training becomes a risk you carry for the life of the model.

A typical purchase runs through six stages: a written requirement, a shortlist of sellers, a sample under NDA, rights and quality review, a pilot, then a signed license with staged delivery. The sections below cover each one.

## What should you define before you approach sellers?

Your requirement should define the training use, the shape of the data and the rights you need before you speak to any seller. Sellers package and price differently for pre-training, fine-tuning, evaluation and reinforcement learning, and a vague brief produces samples you cannot compare.

| Training use | What to specify | Rights you usually need |
| --- | --- | --- |
| Pre-training | Modality, volume in tokens, hours or images, languages, date range, deduplication standard | Broad training rights across model versions |
| Supervised fine-tuning | Task types, input and output format, expertise of authors, labeling guidelines | Training rights for named models and rights to outputs |
| Evaluation | Number of tasks, expected answers, scoring rules, contamination controls | Evaluation use, often exclusive, with no onward publication |
| Reinforcement learning | Environment, task set, verifier or reward rules, replay requirements | Rights to run, modify and generate new trajectories |
| Domain adaptation | Industry, record types, source systems, de-identification standard | Training rights plus confidentiality and output safeguards |

Add the commercial terms you can accept: exclusive or non-exclusive, term, territory, budget range and delivery date. If you can name the models the data will train and whether outputs will be commercialized, write that down as well. Sellers and their lawyers will ask.

Our [training data catalog](https://sellmydatatoai.net/buy-ai-training-data) shows what each data type usually includes, which helps you write a requirement sellers can quote against.

## Where can you buy AI training data?

AI training data is sold through five main routes: direct deals with data owners, training-data companies, data marketplaces, content licensing platforms and brokers. Each route trades price and speed against how much rights evidence you receive.

| Route | How it works | Rights evidence you usually get | Best for |
| --- | --- | --- | --- |
| Direct from the data owner | You negotiate a license with the company that created the records | Whatever you request in diligence | Unique company data and exclusive deals |
| Training-data company | The vendor collects, cleans or licenses data and delivers datasets | Varies by vendor, often summarized | Speed and managed delivery |
| Data marketplace | Sellers list datasets with previews and standard terms | Listing terms and seller statements | Off-the-shelf tabular and media datasets |
| Content licensing platform | Rights holders license catalogs through a platform or agent | Catalog-level or per-asset rights | Media, publishing and archives |
| Broker | An intermediary finds and vets sellers, then you contract directly | A vetting file for each seller | Proprietary data with documented rights |

Our [directory of data buyers and platforms](https://sellmydatatoai.net/data-buyers) lists the training-data companies, marketplaces and licensing platforms active in this market, with their published terms.

Public and scraped data is the cheapest route and the least certain. The Data Provenance Initiative audited more than 1,800 text datasets and found license omission rates above 70% and error rates above 50% on popular hosting sites, according to its 2024 paper in Nature Machine Intelligence.

## How do you check that a seller has the right to license the data?

A seller has the right to license data only if it can show ownership, permission for AI use and no conflicting obligations. Holding the files proves none of these.

> "Possession does not simply mean permission." (Troy Lieberman, Nixon Peabody, October 2026)

Nixon Peabody's alert frames the rights test as three questions: can the seller transfer the data to you, can you use it for the intended AI purpose and can you keep and commercialize what training produces. Restrictions can sit in privacy notices, employee policies, client contracts, IP assignments, inbound licenses, open-source terms and NDAs.

- Ask for the privacy notice in force when the data was collected, not the current one. The same alert notes that regulators treat quietly broadening privacy terms to allow AI training as potentially unfair or deceptive.
- Ask for IP assignments from the staff and contractors who created the content, especially for code and design files.
- Ask which client contracts and NDAs cover the records, and whether any of them prohibit disclosure or reuse.
- Ask how third-party material such as licensed images, open-source code or music in video was identified and then cleared or removed.
- Ask who will sign and whether that person can bind the company. For a closed company it may be a liquidator or trustee.

Our [due diligence checklist](https://sellmydatatoai.net/guides/ai-training-data-due-diligence-checklist) lists every document to request, and [data provenance explained](https://sellmydatatoai.net/guides/data-provenance-explained) covers how to trace a dataset back to its origin.

## How should you test a sample before you buy?

A sample is useful only if it was drawn at random across the whole dataset and labeled with how it was drawn. A hand-picked sample shows you the seller's best rows, not what you will receive.

Sellers face a real constraint here. In a Hugging Face forum thread that ran from July to September 2026, data sellers described the deadlock: they cannot prove quality without exposing the data, and buyers will not pay without seeing enough of it. Contributors proposed fixes you can ask for directly.

- A seeded, stratified random sample that is labeled as representative
- A locked preview where you run fixed queries without downloading the full set
- An evidence pack with a file manifest and hashes tied to an exact version
- A paid pilot on an agreed experiment, with only aggregate results returned if the seller prefers
- Acceptance tests at delivery, with replacement or refund for objective specification failures

Run the sample through your own pipeline rather than reviewing it by eye. Check the schema against the manifest, then test for duplicates, label accuracy, residual personal data and fitness for your task. [How to evaluate a dataset sample](https://sellmydatatoai.net/guides/how-to-evaluate-dataset-samples) sets out the full test plan.

## How much does AI training data cost?

AI training data costs anything from a few thousand dollars for a research license to tens of millions of dollars a year for a large platform deal. Price depends on exclusivity, uniqueness, rights clarity and how much preparation the seller has already done.

| Deal type | Figure | Source |
| --- | --- | --- |
| Research license | $5,000 to $15,000 | Phyllo, reported market ranges, September 2026 |
| Commercial training license | $25,000 to $500,000 and above | Phyllo, reported market ranges, September 2026 |
| Company data, buyers' published offers | Median of $280,000 per company | AIData.Broker, via ddiy.co, October 2026 |
| Company data, public sales | Median of $31,600 per company | AIData.Broker, via ddiy.co, October 2026 |
| Video footage | About $3 per minute | Bloomberg on Adobe, April 2024 |
| Platform content | About $60 million a year | Reuters on Reddit and Google, February 2024 |

A research license is cheap for a reason. Phyllo notes that research licenses usually do not permit training a production model, and that evaluating on research terms and then shipping without re-licensing is a common and expensive mistake.

Exclusivity is the largest single multiplier. AIData.Broker puts an exclusive deal at about 2.5 to 5 times a non-exclusive one. [AI training data pricing models](https://sellmydatatoai.net/guides/ai-training-data-pricing-models) explains flat fees, unit pricing, subscriptions and revenue shares.

## What should the license agreement cover?

The license agreement should cover what you may do with the data, what you own of everything training creates from it and who carries the risk if the data is not as described. Nixon Peabody recommends layer-by-layer terms, because deleting source files does not remove their influence from a trained model.

| Clause | What to agree |
| --- | --- |
| Permitted use | Training, evaluation, fine-tuning or product development, and which models and affiliates |
| Derived assets | Ownership of annotations, embeddings, evaluation sets, fine-tuned models, weights and outputs |
| Exclusivity, term and territory | Exclusive or non-exclusive, start and end dates, permitted regions |
| Onward transfer | Whether data may pass to cloud vendors, subprocessors, model partners or research partners |
| Representations and warranties | Authority, privacy notices, contract restrictions, IP, litigation holds, completeness of exclusions |
| Indemnities and caps | Privacy claims, IP disputes, regulatory inquiries, cost of removal or retraining |
| Acceptance and payment | Acceptance tests, inspection period, staged payments and holdbacks |
| Rights failure | Deletion, retraining or model retirement if rights lapse or prohibited material surfaces |
| Expiry | What happens to trained models and retained copies when the license ends |

## What legal obligations do you have as the buyer?

As the buyer, you carry the regulatory obligations that attach to training and deploying your model, and they do not pass to the seller. Which ones apply depends on where your model is placed on the market and what the data contains.

- EU AI Act: providers of general-purpose AI models must keep a copyright compliance policy and publish a summary of training content under Article 53(1)(c) and (d) of [Regulation (EU) 2024/1689](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng). From August 2, 2026 the Commission can fine providers up to 3% of worldwide annual turnover or €15 million, whichever is higher.
- Personal data: Nixon Peabody notes that under the CCPA, data counts as de-identified only if it cannot reasonably be linked to a person, the holder publicly commits not to re-identify it and recipients are contractually bound. UK GDPR and EU GDPR apply where records relate to people in the UK or EU.
- Health data: HIPAA allows de-identification by expert determination or by removing 18 types of identifier under the safe harbor method, according to the [US Department of Health and Human Services](https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html).
- Remedies: the same alert notes that the FTC has required companies to delete models and algorithms built on improperly obtained data.

This is a summary of what the law and regulators say, not advice on your situation. Take independent legal advice before you sign. Our guide to [the EU AI Act and training data transparency](https://sellmydatatoai.net/guides/eu-ai-act-training-data-transparency) covers the Article 53 summary in detail.

## What are the steps to buy a dataset?

1. **Write the requirement**: Define the training use, modality, volume, languages, date range, exclusivity, budget and delivery date.
2. **Build a shortlist**: Compare sellers on rights evidence and fit for your task, not on volume alone.
3. **Sign a mutual NDA and review samples**: Ask for a seeded random sample labeled with its sampling method, plus the data manifest.
4. **Run rights diligence**: Collect privacy notices, IP assignments, contract reviews and the de-identification method before any pilot.
5. **Pilot**: Test the data on an agreed experiment in a controlled environment and expand access only as conditions are met.
6. **Negotiate the license**: Agree permitted use, derived assets, warranties, indemnities, acceptance tests and staged payments.
7. **Accept delivery**: Check the delivered files against the manifest and hashes within the agreed inspection period.
8. **Record provenance**: Keep the license, manifest, hashes and diligence file so you can complete regulatory summaries and answer challenges later.

## How does Sell My Data To AI help buyers?

Sell My Data To AI helps buyers as an independent broker. We do not own an AI model and we do not buy data for our own use. Buyers pay nothing to submit a requirement or review a shortlist, and any fee is agreed in writing before an introduction.

Every seller passes our eight-point vetting standard before any buyer sees their data, covering authority to sign, ownership, contract restrictions, privacy notices, third-party content, de-identification, sample quality and manifest accuracy. You see redacted samples under a mutual NDA and sign the license directly with the seller. [How we vet](https://sellmydatatoai.net/how-we-vet) explains each check.

We are a broker, not a law firm, so take your own legal advice before signing. To start, [book a buyer consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer) or [submit a data requirement](https://sellmydatatoai.net/request-data).

## Frequently asked questions

### Can I buy AI training data outright rather than license it?

Outright purchase of AI training data is rare. Most sellers grant a license because they want to keep the data or license it again elsewhere. An exclusive perpetual license comes closest to a purchase. Polyshares, for example, has published four company data deals between $110,000 and $450,000, all perpetual and exclusive, according to ddiy.co's October 2026 compilation of buyer terms.

### Is scraped web data safe to train a production model on?

Scraped web data carries more rights uncertainty than any other source. Website terms, copyright and personal data law all apply, and in the EU, providers of general-purpose AI models must respect text and data mining opt-outs under Article 4(3) of Directive (EU) 2019/790. Licensed data with documented provenance is the safer basis for a model you plan to commercialize.

### How long does it take to buy a dataset?

Buying a dataset usually takes weeks rather than days, and most of that time goes on rights diligence and legal review rather than delivery. Published seller-side timelines give a guide: Handshake AI states 7 to 14 business days for its process and Polyshares makes offers in 1 to 2 weeks, per ddiy.co. Allow longer for exclusive deals, personal data or a pilot.

### Do I need an exclusive license?

You need an exclusive license only if your advantage depends on competitors not training on the same data, as with a private evaluation set or a unique domain archive. Exclusivity costs more: AIData.Broker puts exclusive deals at about 2.5 to 5 times a non-exclusive one. A time-limited or field-limited exclusive can be a cheaper middle ground.

### What if the data contains material the seller did not disclose?

Your license should answer this before it happens. Agree warranties on the completeness of disclosed exclusions, a process for removing prohibited records, indemnities for resulting claims and whether deletion, retraining or model retirement applies once the data has been used. Nixon Peabody recommends settling these remedies at signing, because deleting source files does not remove their influence from a trained model.

### Does Sell My Data To AI hold the data I buy?

No. Sell My Data To AI does not take custody of raw data. Buyers see redacted samples under a mutual NDA, and full data moves only after you sign a license directly with the seller. That keeps the contract, the delivery and the rights evidence between you and the data owner, with us helping both sides prepare.

## Sources

- [Shopping for AI training data? Five diligence questions before you buy](https://www.nixonpeabody.com/insights/alerts/2026/10/26/shopping-for-ai-training-data-five-diligence-questions-before-you-buy), Nixon Peabody LLP, accessed November 2026
- [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026
- [Is selling datasets way harder than building them, or is it just me?](https://discuss.huggingface.co/t/is-selling-datasets-way-harder-than-building-them-or-is-it-just-me/177440), Hugging Face forums, accessed November 2026
- [A large-scale audit of dataset licensing and attribution in AI](https://www.media.mit.edu/publications/a-large-scale-audit-of-dataset-licensing-and-attribution-in-ai/), MIT Media Lab (Nature Machine Intelligence, 2024), accessed November 2026
- [Best public data providers for AI training and grounding](https://www.getphyllo.com/post/best-public-data-providers-ai-training-grounding), Phyllo, accessed November 2026
- [How much is company data worth to AI?](https://ddiy.co/how-much-is-company-data-worth-to-ai/), ddiy.co, accessed November 2026
- [Sell company data to AI](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026
- [Adobe is buying videos for $3 per minute to build AI model](https://www.bloomberg.com/news/articles/2024-04-10/adobe-is-buying-video-clips-for-3-per-minute-to-build-ai-model), Bloomberg, accessed November 2026
- [Exclusive: Reddit in AI content licensing deal with Google, sources say](https://tech.yahoo.com/ai/articles/exclusive-reddit-ai-content-licensing-041012503.html), Reuters (syndicated by Yahoo), accessed November 2026
- [Regulation (EU) 2024/1689 (Artificial Intelligence Act)](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng), EUR-Lex, Publications Office of the European Union, accessed November 2026
- [Guidance regarding methods for de-identification of protected health information](https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html), US Department of Health and Human Services, accessed November 2026

---

Book a free consultation: https://sellmydatatoai.net/book-a-consultation