What Is a Data Licensing Agreement? The Terms for AI Training

A data licensing agreement is a contract in which the holder of a dataset lets a buyer use it for defined purposes, for a defined time, at a defined price, while the holder keeps ownership. In AI training deals the licence decides what the price actually buys, and published terms vary enormously: one large buyer's standard terms take a worldwide, royalty-free, fully paid-up, nonexclusive, irrevocable, perpetual license, with the right to sublicense through multiple tiers for a single fee, while another restricts itself by contract to using the data for AI model training and evaluation only, with the company keeping ownership. Two offers with the same headline number can be worth very different amounts once the clauses are read.

This guide is for a holder (a founder, CFO or counsel of a company with data a buyer wants) reading a licence for the first time, and for a buyer's team that wants to know where holders push back. It explains how a licence differs from a sale or an assignment, then goes clause by clause through the terms that decide value in AI deals, using the 15 licence terms in our licence terms register and the 65 clauses we have read verbatim in buyers' published terms. It ends with how each term moves price, a worked example, and the questions holders ask most. It is not a template: every licence should be drafted for its data, and the offer check compares a real offer's terms with what buyers publish.

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

Contents

1. What a data licensing agreement is
2. Licence, sale or assignment
3. The clauses at a glance
4. Licensed purpose and field of use
5. Exclusivity
6. Term, and the life of trained models
7. Territory
8. Sublicensing and affiliates
9. Who owns models and derivatives
10. Deletion or return at the end
11. Naming the holder, and California AB 2013
12. Warranties of rights and consent
13. Indemnity and liability caps
14. Audit and certification
15. Payment structures
16. Delivery, acceptance and refresh
17. How terms move price
18. Worked example: two offers for the same records
19. Questions people ask
20. What this means for you

1. What a data licensing agreement is

A licence is permission. In a data licensing agreement the holder (the licensor) permits the buyer (the licensee) to do specified things with specified data, and the buyer pays for that permission and accepts limits on it. Everything the licence does not grant, the holder keeps. That is the basic difference from a sale: after a licence, the holder still owns the data, can usually keep using it, and, unless it promised exclusivity, can license it again to someone else.

For AI training, a data licence typically has five parts, and the rest of this guide follows them:

• The grant: what the buyer may do (train, fine-tune, evaluate, retrieve), with which data, where, for how long, whether others can have the same rights, and whether the buyer may pass them on.
• The limits: what the buyer may not do (resell the raw data, re-identify people, reproduce text verbatim in outputs).
• The money: a fixed fee, a minimum guarantee, per-unit pricing, a revenue share or refresh fees, and when each is paid.
• The promises: the holder's warranties about its rights and the people in the data, the indemnities behind them, and the caps on liability.
• The ending: what happens to the data, and to models trained on it, when the licence ends or is breached.

Data licences differ from the software or content licences most companies have signed in one important way. Much of what a company holds, such as transaction records, logs and factual fields, is not protected by copyright at all, because copyright protects expression, not facts. Other parts are (documents, code, written messages). Where copyright is thin, the contract does most of the work: the buyer's obligations exist because it promised them, not because the law would otherwise forbid the use. That is why the clauses matter more here than in a typical content licence, and why reading them is the holder's job before signing, not after.

2. Licence, sale or assignment

Three ways to hand over data

US copyright law draws the line clearly for the parts of a dataset it covers. A "transfer of copyright ownership" is an assignment, mortgage, exclusive license, or any other conveyance ... but not including a nonexclusive license. In other words, an assignment moves ownership outright, an exclusive licence moves ownership of the rights it grants (no one else, the holder included, may exercise them), and a non-exclusive licence moves nothing: the holder still owns everything and has only promised not to sue the buyer for the licensed use. An assignment or exclusive licence must also be in writing and signed by the owner of the rights conveyed to be valid, so an exclusivity promise made on a call or in an email thread may not deliver what the buyer thinks it bought.

A sale of data, in the everyday sense, usually means the holder hands over copies and promises not to sell the same copies to anyone else, sometimes with an assignment of whatever intellectual property exists. The Spirit Airlines bankruptcy sale of corporate records to Google, for example, was structured as a sale agreement that, among other things, bars the seller from selling the same records to anyone else while letting it keep copies for its wind-down and legal compliance.

What buyers' pages say

Most company-data buyers now say plainly that the holder keeps ownership. Turing says You retain ownership of your underlying data. We acquire defined rights to the prepared record; Protege says partners are licensing limited rights for defined AI training purposes and are not transferring ownership; Troveo licenses on your behalf under terms that you sign off on; ownership never transfers. Ownership on paper, though, is only half of it. A licence that is perpetual, irrevocable, worldwide and sublicensable through multiple tiers leaves the holder owning the data while the buyer can do almost anything the holder could, including resell access. Our register treats that as what it is: a perpetual, sublicensable licence is a resale right, and it should be priced like one.

For a closing company, a sale can be the right answer: the records have no future use to the holder, and a clean sale ends its obligations. For an operating company, a licence almost always is, because the company keeps using its own data and may license it again. Our guides on selling company data to AI labs and selling customer data in bankruptcy cover each case.

3. The clauses at a glance

The table summarises each clause the rest of this guide explains: what buyers typically ask for, where holders commonly stand, and the record behind each. "Where holders stand" is the position our register records from public agreements, templates and disclosed deals; it is not legal advice for any particular deal. Our register notes that for 2 of these terms the public record is thin.

4. Licensed purpose and field of use

The grant begins with what the buyer may do. In AI deals the uses that matter are pretraining (building a model's general abilities from large volumes of data), fine-tuning (adapting a model to a task), evaluation (testing a model against the data), and retrieval (looking content up at answer time and showing or summarising it). These are different uses with different value, and a licence that says only "AI purposes" may grant all of them. Our register's position is to name the permitted uses and exclude what is not paid for, such as redistribution of the raw data, verbatim display in answers, or use outside a defined field (a buyer licensing legal records for legal models, say, not for every model it builds).

Buyers' published terms show the range. Handshake's says it is contractually restricted to using the data for AI model training and evaluation only; Ooak's contract licenses them for training only, with your exclusions annexed. At the other end, submitting assets to SimpleClosure's marketplace grants it permission to de-identify the data and use for any legal purpose. A field-of-use limit also has a price: our valuation shows an evaluation-only licence at 0.2 to 0.5 times the value of a full commercial one, so a holder who narrows the purpose should expect a lower fee, and one who grants everything should be paid for everything.

Two limits belong in almost every licence of company records whatever the purpose. The buyer will not attempt re-identification of the people in de-identified data, and for text, code or media where a model could reproduce confidential content verbatim, there should be output filters and a cap on verbatim reproduction. HarperCollins' licence to Microsoft is the public example: it limited outputs to no more than 200 consecutive words and/or 5 percent of a book's text.

5. Exclusivity

What it is worth

Exclusivity means the holder promises not to license the same data, in the same field, to anyone else for a period. It is among the most valuable concessions a holder can make, and it should never be given by default. Founders who sell training environments told Epoch AI that exclusive deals cost roughly 4 to 5 times what non-exclusive ones do, and our valuation shows an exclusive licence at 4 to 5 times the non-exclusive figure. The other side of the trade is that non-exclusive data can be sold again: Troveo advertises several sales per dataset, and Kled signed a non-exclusive licence worth $12M over two years.

Many disclosed AI content deals are non-exclusive. The Financial Times, Axel Springer and Stack Overflow deals all are, and Google's content-offer pilot tells partners This is a non-exclusive opportunity. Where exclusivity appears, it is usually a window: Disney's three-year licence to OpenAI carried about one year of exclusivity, and Troveo says its exclusivity is time-limited and it has to be earned.

Exclusivity hidden in other words

Exclusivity sometimes arrives under another name. micro1's terms ask the holder to warrant that it has not previously entered into any transaction granting a third party rights to license, access, use or otherwise exploit the same Data on the same or substantially similar terms: a backward-looking promise that rules out holders who already licensed the data elsewhere. Kled's contributor terms go further and make the licence exclusive to us. Intermediaries also seek exclusivity over a holder's catalogue so that buyers deal only with them; Protege's own case study says holding exclusive training rights let it maximize the price because AI buyers were not able to negotiate with multiple parties. That can help a holder, or it can lock the holder into one channel.

Our register's position: offer exclusivity only for a defined field and a defined window, charge a multiple for it, and disclose any platform vendor that has already trained on the content. Robot data made to order is the exception that proves the rule: there exclusivity is the starting point and non-exclusive is the discount, because the data was collected for that buyer.

6. Term, and the life of trained models

How long the licence runs

The term is how long the buyer may use the data. Disclosed AI licences are finite and short: Reddit's January 2024 licences ran terms ranging from two to three years, Meta's deal with News Corp at least three years, and HarperCollins' with Microsoft three years ( lever). Buyers' standard terms, by contrast, often ask for perpetual rights. Our register's position is that for a living company, a two to three year term with renewal and refresh rights keeps the holder's leverage, while a one-off archive (a closed company's records, say) can reasonably be priced as a perpetual sale.

What happens to models already trained

The harder question is what happens at the end of the term to models already trained on the data. A model cannot easily "unlearn" a dataset, so buyers ask for the right to keep and use trained models forever, and that is the default ask, phrased in one template as a perpetual, worldwide license for training and use of models trained. micro1's terms are explicit that deleting the data does not require Licensee to delete or retrain any model, weights, or derivative works.

Model deletion is hard to win, and in most deals a holder should not spend leverage chasing it. What they can do is price it and limit it. Licensing counsel quoted in our register recommends a perpetual right to models already trained plus a finite licence to the dataset, limited to a single model with the right to extend on additional fee. And because a licence grants only what it says, any survival must be written in; silence grants none. In health and personal-data contexts the register's position is to limit post-term rights to de-identified data.

7. Territory

Territory is where the licence may be exercised. In data licences for AI it is usually worldwide, as in micro1's grant ( clause), because models are trained in one country and served in many. Holders gain little by restricting territory in the ordinary sense, and buyers will resist it.

What does matter is where the data goes and who may receive it. The US Justice Department's Data Security Program forbids certain transactions that give countries of concern, or people tied to them, access to bulk US sensitive personal data above set thresholds, such as biometric identifiers over 1,000 US persons, health or financial data over 10,000 and covered identifiers above a higher threshold. A licence with a foreign buyer should therefore carry an onward-transfer clause; the Department offers model language that prohibits the foreign party from selling, licensing of access to, or other similar commercial transactions with countries of concern. Holders in privacy-sensitive sales also offer to keep data in the US: one bidder in the Spirit records sale offered to keep data in the US among other protections. Territory, in practice, is a compliance clause rather than a commercial one.

8. Sublicensing and affiliates

Who else gets the rights

Sublicensing is the buyer's right to grant the licensed rights on to others. It is where a licence quietly becomes a resale right. micro1's master terms grant the licence to "Licensee and its affiliates", with the right to sublicense through multiple tiers: micro1 may grant rights to a lab, which may grant them on again. Kled's contributor licence is likewise fully sub-licensable (through multiple tiers). At the other end, Defined.ai's licence to its own buyers is non-exclusive, non-sublicensable, and non-transferable, so they cannot pass the data on.

Even buyers that take sublicensing rights often limit the most damaging use. micro1's terms bar it from reselling the data as received from Licensor and as a standalone data product, and Grepped says the data it handles is never resold to competitors or published. The difference between "may sublicense the right to train" and "may resell the dataset" is the clause to read twice.

Affiliates and flow-down

"Affiliates" means companies under common control with the buyer. For a large lab or a vendor inside a group, an affiliate grant can extend the licence to businesses the holder never vetted. Our register's position is that the buyer will not resell or sublicense the raw data, and will bind any permitted recipient to the same terms. The Spirit sale agreement shows the model: the buyer may pass the data on only to recipients bound by the same no-re-identification commitment. That flow-down mirrors California's privacy law, under which a business sharing de-identified consumer data contractually obligates any recipients to keep it de-identified.

9. Who owns models and derivatives

Buyers expect to own the models they train, and holders who license training rights generally accept that: the buyer's model is the buyer's product. The licence should say so plainly, because the opposite ambiguity (a holder claiming a share of a model's weights) is a deal-breaker for buyers and rarely worth anything to a holder.

The more important question is derivatives of the data itself. A buyer may produce cleaned versions, labelled versions, embeddings, synthetic rewrites or benchmarks from the holder's records. micro1, for example, says that certain datasets may undergo synthetic rewrites that preserve workflow structure. Those derived datasets can be licensed to others, so the licence should decide whether they are covered by the same limits as the original (no resale as a standalone product, flow-down, deletion at the end) or belong to the buyer outright. A holder who licenses records "for training" and lets the buyer own every derived dataset has, in effect, sold a product line.

The practical rule in our register's terms: accept buyer ownership of models and weights; treat any derived dataset that could stand in for the original as subject to the same restrictions as the original; and decide in writing whether the buyer may keep derived datasets after the licence ends.

10. Deletion or return at the end

When a licence ends, or when the buyer breaches it, the holder wants the data back or destroyed. Our register's position is to require deletion of the dataset (not the models, unless agreed) within a fixed number of days, with a written certificate, and deletion on breach. NVIDIA's own model-training data agreement is a published example: on termination the company will stop using the Datasets and delete and destroy copies of Datasets within fourteen days.

Buyers' pages describe different retention practices. Ooak keeps only the anonymized dataset ... for the licence term; Grepped says original datasets are deleted once processing is done; Nyne sets retention and deletion terms per engagement. Raw copies are the sensitive part. SimpleClosure warned founders in October 2026 that a buyer who takes raw files and promises to scrub them later already holds what was meant to be removed, which is why who de-identifies, and where, belongs in the same clause as deletion. Our deletion clock shows how long software vendors keep a closed account's data, which matters for the holder's own copy too.

11. Naming the holder, and California AB 2013

Confidentiality and publicity

Most AI data deals keep their terms private: of the 111 licensing deals on our deal ledger, only 19 state a dollar amount. A holder should decide early whether it wants to be named. Some want the publicity of a deal with a known lab; many company holders do not want their customers, employees or competitors to learn their records were licensed. The default in a licence should be confidentiality of the holder's identity and the terms, with any announcement agreed in writing.

There is also a disclosure the holder itself may owe. A company that is later sold may have to disclose its data licences in the acquisition agreement: in one 2025 software acquisition the company being sold represented that, except as scheduled, it had not let any company or customer data, de-identified data included, be used as training data by another party. Keep a record of every licence for that reason.

What AB 2013 requires the buyer to say

California's Generative AI Training Data Transparency Act (AB 2013, now Civil Code section 3110 and the section that follows it) requires developers of generative AI systems offered to the public in California to post documentation about their training data, including a high-level summary that covers, among other things, the sources or owners of the datasets and whether the datasets were purchased or licensed by the developer. A federal court refused to block the law in March 2026, describing the required summary as touching sources and owners of datasets; the size of datasets; the period of collection and more; xAI's appeal is pending in the Ninth Circuit, and the law stays in force meanwhile. The duty falls on the developer that builds or substantially modifies the public system, not on a data vendor that trains nothing.

Does that mean a holder will be named? Not necessarily. The statute does not say how specific "sources or owners" must be, and early disclosures by large developers stayed at the level of categories, such as nonpublic data obtained from third-party partners, according to a law firm's review. The EU's template for the equivalent summary under the AI Act asks whether a provider concluded licences with rightsholders and the modality(ies) of the content covered, not the licensor's name. Still, a holder that does not want to be identified should say so in the licence: allow the buyer to make the disclosures the law requires, at the level of generality the law allows, and nothing more without consent.

12. Warranties of rights and consent

What buyers ask the holder to promise

A warranty is a promise that a fact is true, with liability if it is not. Buyers ask holders to warrant title (the holder owns the data or has the right to license it), authority (it may sign), lawful collection, non-infringement and privacy compliance. Their intake forms ask the same questions up front: Mercor asks Do you have the legal right to license this data for commercial use and AI / model training?, and Defined.ai requires that All data must be ethically sourced, including proper consent, documentation, and the right to commercialize the data for AI training. SimpleClosure's terms have the holder warrant every notice and consent from any individuals whose personal information may be contained in the Digital Assets, including employees, contractors, customers, investors, and vendors.

Some buyers attach sharp remedies to these promises. micro1's terms provide that if the data manifest is materially inaccurate, no amounts are payable to Licensor, Licensor will refund any amounts already paid, a right that survives acceptance. Kled requires a contributor to immediately refund what it was paid for content whose consent is withdrawn.

Where holders stand

Our register's position, drawn from public templates and agreements, is that holders commonly give title and authority without qualification and qualify the rest by knowledge, tied to a disclosed rights inventory. The warranty then says the holder's disclosed limits are accurate, not that the data is perfect. A typical lawful-collection warranty reads "That the data was collected in compliance with applicable law". On consent, the register's position is to warrant to knowledge, keep a consent ledger per person or per work, and agree a process (not a guarantee) for later opt-outs, and to exclude people who did not opt in rather than promise consent you cannot show.

This is where the rights file earns its keep: a written inventory of what the holder owns, what its customer contracts and software vendors' terms allow, what its privacy policies promised, and what it has excluded. With it, a holder can give the warranties buyers need without guessing; without it, every warranty is a bet. Our readiness check and guides to whether it is legal to sell customer data and employee emails and chats cover what goes in it.

13. Indemnity and liability caps

An indemnity is a promise to cover the other side's losses from defined claims. Buyers ask holders to indemnify them for third-party claims that the data infringes someone's rights or breaches privacy law; a template form reads the licensor indemnifies the licensee for any third-party claim alleging that the licensed data infringes or violates the law. For a holder, an uncapped indemnity on a licence can expose the company to more than the deal is worth.

Our register's position is to limit indemnity to precisely defined claim types, cap it (a multiple of fees), and carve out claims caused by the buyer's own use, combination with other data, or outputs. The carve-outs matter most in AI deals: if a buyer's model reproduces someone's text, or the buyer combines the holder's records with other data to re-identify people, that is the buyer's risk, not the holder's. The reverse arrangement also exists. NVIDIA's own data agreement makes the recipient indemnify the provider ( term), which shows that indemnity is negotiable in both directions, not a fixed feature of being the licensor.

The liability cap usually sits beside the indemnity: a ceiling on what either side can recover under the licence, typically tied to the fees paid. Holders should check that the cap applies to the indemnity, that it is not swallowed by broad exceptions, and that the buyer's own liability (for misuse, re-identification or breach of confidentiality) is not capped so low that the holder has no real remedy.

14. Audit and certification

Audit clauses let one side check the other's compliance. Holders want to verify that buyers use the data only as licensed; buyers want to verify that the data was de-identified before it arrived. Our register's position is to ask for records sufficient to show compliance with the licensed uses, an annual compliance certificate, and audit only on reasonable suspicion of breach, and to expect the mirror request: buyers want certification of de-identification before each transfer.

The Spirit Airlines sale shows how far certification can go in a privacy-sensitive deal. The $10M price was payable only on delivery of de-identified data, and delivery required an agent's certification to the California privacy standard; the debtors hired Tonic.ai as an independent agent, which must file an affidavit with the court before the first and every later transfer. Who controls de-identification is itself priced: in the same auction one bidder offered $5.2M with third-party de-identification or $7.0M if it could de-identify with its own tools, that is, 35% more for control of the scrub. Our register's position is that the holder selects or jointly selects the de-identification provider, approves the protocol and requires certification before anything leaves. Our de-identification standard and de-identification tool set out what that protocol should cover.

15. Payment structures

The five shapes

How a licence is paid is as important as how much. Five structures appear in the public record, often combined:

Each shape puts risk in a different place. A fixed fee is certain but caps the upside. A minimum guarantee keeps a floor while sharing upside; Tempus AI's extended licence with Recursion replaced discretionary fees with committed payments of $42 million over the next three years, trading possible upside for certainty. Per-unit pricing makes payouts traceable to individual contributors, which matters where consent is per person or per work. A revenue share depends on the intermediary's sales and its reporting, so the percentage and the audit right both matter; several platforms do not publish their share. Refresh fees reward a living company for new data and keep the relationship going.

When the money arrives

Timing and triggers matter as much as the amount. Buyers pay on different events: Grepped within 7 days of your data passing review, Mercor by wire upon delivery of the anonymized dataset to the buyer, Nyne once the dataset is delivered and accepted. micro1's terms let the holder invoice Licensee only after Licensee has given written notice that the Data has been accepted. Much of a multi-year licence can be paid at the start: Getty recognised about $40M of two multiyear deals worth $65M in the quarter they were signed. Our register's position is to anchor on a fixed minimum plus refresh or usage fees, with a tranche on signature where the buyer will agree to one. Processing costs are the buyer's: in the Spirit sale Google and micro1 each bore all de-identification costs with no price reduction.

16. Delivery, acceptance and refresh

The delivery clause says what the holder hands over, in what format, on what schedule, and how the buyer accepts it. Acceptance is where many one-off deals go wrong, because payment usually waits for it. micro1's terms give it thirty (30) business days after delivery of the Data to reject it by written notice; if it says nothing, the data is deemed accepted. Holders should define acceptance criteria in writing (format, fields, volume, de-identification standard) so that "acceptance" cannot become a renegotiation.

For recurring licences, performance obligations appear. Reddit warned investors that its data licences carry terms including API performance requirements, that we may be unable to meet. Our register's position is to specify format, schedule and acceptance; avoid service levels the holder cannot meet; tie refresh fees to new data. A company that is not in the business of running a data feed should not promise uptime like one.

Refresh is the commercial reason to license rather than sell. A living company produces new records every month, and buyers pay for continuing access: our register's lever is to sell access to refreshing data, not a one-off file, as Reddit's licences carried $203.0 million aggregate contract value over two to three years. Archive deals, by contrast, often do not repeat: Getty attributed a forecast revenue decline to one-time recognition of earlier licensing deals.

17. How terms move price

Every clause above is a price term in disguise. A holder who accepts a buyer's standard terms has accepted a price for each of them, usually without seeing it. The register of 69 negotiation levers, 36 of them backed by a primary record or several sources, shows which terms move value and by how much where the public record allows a number:

The last two rows are not clauses, but they decide what every clause is worth. A holder with one offer negotiates terms one at a time against a buyer who knows the market; a holder with several offers on comparable terms can see what each concession costs. That is why our negotiation page and the guide to how much company data is worth treat terms and price together, and why the valuation shows exclusive and narrower licences beside the standard figure.

18. Worked example: two offers for the same records

This is an illustrative example, not a real company, and the offers are composites of published terms, not quotes from any buyer. A 60-person B2B software company, still operating, has a decade of support tickets, engineering tickets linked to code changes, and internal chat. It has done its rights work: customer contracts allow de-identified aggregate use, its chat vendor's export route is clear, and it has a datasheet. Two buyers make offers with similar headline amounts.

Offer A is a single fee, invoiced after written acceptance, for a worldwide, perpetual, irrevocable, non-exclusive licence with the right to sublicense through multiple tiers. The buyer de-identifies with its own tools, may rewrite some records synthetically, has thirty business days to reject the data, and keeps every model and derived work forever. The holder warrants it has not licensed the data before, and refunds everything if the data manifest proves materially inaccurate.

Offer B is a three-year licence for training and evaluation of the buyer's own models only, non-exclusive, no sublicensing of the data, affiliates only if bound by the same terms. A jointly chosen provider de-identifies inside the company's systems and certifies before each transfer. There is a fixed minimum each year, paid in part on signature, plus a fee for each quarterly refresh of new tickets. At the end the buyer deletes the dataset within set days and certifies it, keeping one model family trained during the term. Warranties are to knowledge against the rights schedule; the indemnity is capped at a multiple of fees and excludes the buyer's outputs.

Reading them side by side. Offer A's grant is close to a sale: with perpetual sublicensable rights the buyer can resell access, and the company's own future licences are worth less because the same data is already in circulation. The company also carries the acceptance and refund risk and gives up control of the scrub, a term one Spirit bidder valued at 35% more. Offer B pays less up front but keeps the data's future value with the company, pays again for every refresh, and leaves the risk where the company can manage it. If the company wanted a clean exit (it was closing), Offer A's structure would be reasonable at a higher price. As an operating company with growing data, Offer B is likely worth more over the term, and its minimum guarantee gives a floor. The useful question is not "which number is bigger" but "what would Offer A have to pay for me to accept its terms", and the levers in section 17 give that question an answer. The offer check runs this comparison on real offers.

Questions people ask

What is a data licensing agreement? A contract in which the holder of a dataset permits a buyer to use it for defined purposes, for a defined term and price, while keeping ownership. In AI deals it also sets what happens to models trained on the data, who de-identifies it, and what each side promises about rights and consent ( register).

What is the difference between licensing and selling data? After a licence the holder still owns the data and, unless it granted exclusivity, can license it again. A sale or assignment moves ownership; under US copyright law an exclusive licence also transfers the rights it covers, but a nonexclusive license does not.

How long do AI data licences last? Disclosed licences are short: Reddit's ran two to three years, and HarperCollins' with Microsoft three. Buyers' standard terms often ask for perpetual rights, which should be priced as a sale, and the right to keep trained models after the term is usually negotiated separately.

Can the buyer keep using its AI model after the licence ends? Usually, if the licence says so: buyers ask for perpetual rights to trained models, and micro1's terms say deleting data does not require Licensee to delete or retrain any model. Nothing survives unless written in, so the holder should limit and price model survival.

How much more is an exclusive data licence worth? Founders selling training environments told Epoch AI exclusive deals cost roughly 4-5x non-exclusive ones, and our valuation uses 4 to 5 times. Exclusivity should be limited to a field and a window.

Will an AI company say it licensed my data? It may have to say something. California requires developers of public generative AI systems to summarise their training data's sources or owners and whether it was licensed, but early disclosures stayed at category level. Put confidentiality and the limits of any required disclosure in the licence.

What warranties will a buyer ask for? Title, authority, lawful collection, non-infringement and consent. Holders commonly give title unqualified and the rest to knowledge, tied to a disclosed rights inventory, so the promise is that the disclosed limits are accurate.

Do I need a lawyer for a data licence? For anything beyond a small, simple deal, yes: the warranties, indemnity and model-survival clauses carry real liability. Reading the terms against the licence terms register first will make that conversation shorter and cheaper.

What this means for you

Read the grant before the price. Four clauses decide most of a licence's value: purpose (what the buyer may do), exclusivity (whether you can license again), term and model survival (how long the rights last and what outlives them) and sublicensing (whether the licence is really a resale right). Then check who de-identifies, when you are paid and on what trigger, and what you are warranting. A buyer's standard terms are a starting position, not the market: the 65 clauses on record show how widely buyers' own terms differ.

Before an offer arrives, build the rights file and a datasheet, and get a valuation that shows what exclusive and narrower licences are worth beside the standard figure. When offers arrive, put each through the offer check, compare them clause by clause with the licence terms register, and see who else buys on the buyer board. Our pages on negotiating and on selling direct or through an agent explain the trade-offs of each route.

AIData.Broker acts as the holder's agent: the holder itself licenses or sells its data to the buyer, and we do not take ownership of it. We put the data before every buyer that fits at the same time, and only the holder pays us: a fee fixed in the mandate, only if a sale closes. The legal background is in our guides to fair use and AI training, scraping for AI and training data provenance, and on the rights hub.

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