A data licensing agreement is the contract under which a company lets an AI buyer copy and train on its data for a stated purpose, term and price, usually without transferring ownership. Five clauses move the most risk: the grant, deletion, trained models, the indemnity and the liability cap.
The contract is where AI training data licensing becomes binding. micro1's Data License Master Terms, updated November 8, 2026, show the pattern: a perpetual nonexclusive grant, raw data deleted within 120 days, no duty to retrain models, a seller indemnity and a 3x-fees cap. Declining to sign is a valid outcome.
Examples carry their check or filing date; the DataPacts Editorial Team explains how we source, date and check every figure.
What a Data Licensing Agreement Is and What It Grants an AI Buyer
A data license grants permission to use data on stated conditions while the licensor keeps ownership, and a data sale transfers the data itself. Google's announcement of 2024-02-22 gives it Reddit content "to display, train on, and otherwise use", and Turing states "Ownership does not change hands" (checked 2026-11-09). A sale uses a bill of sale, as in the Spirit Airlines bankruptcy (Bankr. S.D.N.Y. 25-11897 (SHL), Dkt. 1463), not approved as of November 9, 2026.
Reddit's S-1 of 2024-02-22 describes arrangements "allowing third parties to access, search, and analyze data on our platform" for two to three years, with unnamed counterparties. Open licenses such as Creative Commons are a different instrument, with fixed public terms and no price.
How the Documents Fit Together: Master Terms, Order Form and Exhibits
Buyers usually pair standard master terms with a deal-specific order form, and the order form can change what the master terms say. The three layers are listed below.
- Master terms: micro1's (updated 2026-11-08) or Defined.ai's agreement with its own customers (effective 2022-11-30).
- Order form: price, delivery, Data Manifest (micro1 s.1.1), audit and exclusivity (s.4.2).
- Exhibits: schedules such as Defined.ai's Exhibit A.
Earlier term sheets, order forms and LOIs can bind before signing.
Every Clause at a Glance: Seller Risk and Buyer Reason
Each clause shifts a risk between seller and buyer, and the table names the risk on each side with a published example. It compares 20 clauses as of 2026-11-09.
| Clause | What it does | Seller-side risk | Buyer-side reason | Example source (dated) |
|---|---|---|---|---|
| Grant | who else may receive the data | exclusivity blocks other buyers | protects investment in unique data | Labelbox Prism FAQ (checked 2026-11-09); micro1 s.1.2, s.5.1 (updated 2026-11-08) |
| Term | duration of use | perpetual use cannot be ended | models are used for years | Reddit S-1, two to three years (2024-02-22) |
| Field of use | permitted purposes | broad purpose, unforeseen uses | room for future products | Tempus 10-K "sole purpose" (filed 2026-02-24) |
| Territory | where or by whom | worldwide reach | global deployment | micro1 s.1.2 (2026-11-08); Protege s.1.1 (checked 2026-11-09) |
| Data scope and manifest | licensed data and exclusions | refund if inaccurate | knows what it pays for | micro1 s.4.3, s.5.7 (2026-11-08) |
| Delivery and security | transfer, encryption, breach notice | short notice duties | security of sensitive data | micro1 s.2.4, s.2.5 (2026-11-08) |
| Acceptance | period to reject | silence counts as acceptance | quality control | micro1 s.1.3 (2026-11-08) |
| Warranties on rights and consent | seller may license | indemnity exposure | shifts title and consent risk | micro1 s.5.1 (2026-11-08) |
| Indemnities | who defends claims | seller pays IP and privacy claims | protection from rights holders | micro1 s.8.1 (2026-11-08); Protege s.4.1 (checked 2026-11-09) |
| Liability cap | limit on damages | cap below exposure | limits own exposure | micro1 3x fees (2026-11-08); Protege 2x fees (checked 2026-11-09) |
| Audit | inspection of records | little means to check | verifies supplier records | Defined.ai Supplier Program terms (checked 2026-11-09) |
| Deletion | raw copies destroyed | backups and models excluded | operational retention | micro1 s.4.4 (2026-11-08); Labelbox Prism FAQ (checked 2026-11-09) |
| Retraining and model retention | fate of trained models | models stay | unlearning not offered | micro1 s.4.4 (2026-11-08); Labelbox Prism FAQ (checked 2026-11-09) |
| Derivatives and outputs | ownership of what is built | all goes to the buyer | freedom to commercialize | micro1 s.2.2 (2026-11-08); Defined.ai s.1(c) (checked 2026-11-09) |
| Attribution | citation, links, output limits | no visibility without it | design freedom | Axel Springer release (2023-12-13); Authors Guild on HarperCollins (2024-11-19) |
| Confidentiality | secrecy of terms and data | survives termination | protects strategy | micro1 s.2.3 (2026-11-08) |
| Payment structure | price and trigger | paid only after acceptance | pays for accepted data | micro1 s.3 (2026-11-08); Labelbox Prism (checked 2026-11-09) |
| Sublicensing and assignment | onward transfer | unknown third parties | affiliates and successors | micro1 s.1.2, s.9.1 (2026-11-08); Defined.ai s.1(a) (checked 2026-11-09) |
| Termination | who ends it, why | buyer exit, data kept | flexibility | micro1 s.4.1 (2026-11-08); Labelbox Prism FAQ (checked 2026-11-09) |
| Governing law | law and forum | foreign forum costs | predictability | micro1 s.9.3 to 9.4 (2026-11-08); Defined.ai (checked 2026-11-09) |
Sources: published terms, SEC filings and publisher releases, dated as shown.
No public survey gives typical term lengths, cap multiples or indemnity scope across all buyers. The license clause matrix sets buyer contracts side by side.
What Is Granted to the Buyer
Five clauses define what the buyer receives: the grant, its duration, the permitted purpose, the territory and the data description.
Grant Clause: Exclusive, Non-Exclusive or a Selling Window
An exclusive grant shuts out every other buyer for the scope it covers, and a non-exclusive grant can still limit parallel sales through a warranty. Three grant models appear in published terms, listed below.
- Exclusive by default: Labelbox Prism, as Labelbox states (checked 2026-11-09).
- Non-exclusive with a warranty: micro1 (s.1.2), with a no-prior-grant warranty (s.5.1).
- Selling window: Turing states it alone sells a record for 90 days (checked 2026-11-09).
Mercor's medical-practice listing names data "already licensed exclusively through a broker" as not a fit (checked 2026-11-09). In drug-target discovery, 23andMe disclosed USD 50 million from GSK on 2022-01-18 to extend exclusivity; no public premium figure exists for company data.
Duration Clause: Fixed Years, Until Terminated or Forever
The term decides how long the buyer may use the data, and published AI data terms range from a few years to perpetual and irrevocable. Reddit's 10-Q for the third quarter of 2026 says its two main partners are "approaching the end of their respective current terms". micro1's grant is perpetual (s.1.2). Defined.ai's customer agreement runs "until terminated" (s.8(a)), while its FAQ states "The license agreement is perpetual" (both re-read 2026-11-10).
Permitted Purpose: What the Data May Train
The permitted-purpose or field-of-use clause lists what the buyer may do with the data, and anything outside it is a breach. Tempus AI's 10-K (filed 2026-02-24) grants Pathos AI its data "for the sole purpose of" developing and training a foundation model. Kopinor's agreement with the National Library of Norway (2025-12-19) excludes specialized news models.
Territory and Recipient Restrictions
Most published AI data grants are worldwide, so restrictions, where they exist, name who may receive the data rather than where. micro1 (s.1.2) and Protege (s.1.1) grant worldwide; micro1's "exclusively" wording limits recipients to the U.S. and allied countries.
Licensed Data Description, Exclusions and Manifest Accuracy
The data description fixes what is licensed and what is carved out, and some buyers can claw back the price if it proves materially wrong. micro1 claws back if the Data Manifest is materially inaccurate, even after acceptance (s.4.3), and excludes health, biometric, payment card, children's and export-controlled data unless allowed (s.5.7).
How the Data Moves to the Buyer
Two clauses govern the handover: how the data travels and how long the buyer has to reject it.
Transfer Method, Encryption and Breach Notice
Delivery and security clauses set the transfer method, the encryption standard and how fast each side reports a breach. micro1 requires encryption in transit and at rest (s.2.4) and incident notice within 2 business days (s.2.5). GDPR Art. 33(1) (applicable since 2018-05-25) sets 72 hours for personal data.
Review Window After Delivery
An acceptance clause gives the buyer a period to reject delivered data, and under deemed acceptance silence counts as approval. micro1 has 30 business days to reject, then the data is deemed accepted, and 60 more days for latent defects (s.1.3).
Who Carries the Risk if Something Goes Wrong
Representations, indemnities and the cap together decide who pays when a rights holder, a customer or a regulator objects.
Representations on Rights and Consent
In the representations the seller promises that it may license the data and that consents exist, and these promises are what an indemnity enforces. micro1's licensor keeps consent records for 7 years after the agreement ends. Defined.ai licenses to its customers AS IS, and in Spirit's bill of sale representations do not survive.
Who Defends Third-Party Claims
An indemnity makes one party defend and pay the other's third-party claims, and in published seller terms the seller indemnifies for the data. micro1's licensor covers IP, privacy, publicity and data protection claims (s.8.1), Protege's providers indemnify (s.4.1), and Grepped's sellers cover regulator claims too (2026-11-08).
Cap on Damages and Its Carve-outs
A liability cap limits what one party can recover, usually as a multiple of fees, and carve-outs remove the cap for the worst conduct. micro1 caps at 3x fees, except fraud, gross negligence and willful misconduct (s.7.2); Protege at 2x fees of 12 months (s.6.2). Data Act Art. 13(4) (Regulation (EU) 2023/2854) makes imposed limits for intent or gross negligence always unfair.
What Happens to the Data and the Models Trained on It
The clauses on audit, deletion, trained models, outputs and attribution decide what the seller still controls after delivery.
Checking Use and Deletion After Delivery
An audit clause lets one party inspect the other's records, and in published terms it is rare, limited or left to the order form. micro1 leaves audit to the order form (s.4.2); Defined.ai's Supplier Program allows one audit a year on 15 business days' notice.
Raw Data After Delivery or Termination
A deletion clause sets when the buyer must destroy its copies of the raw data, and it rarely reaches the models already trained. Erasure duties under GDPR and AI training data rules (Regulation (EU) 2016/679, Arts. 17, 19, 28(3)(g)) apply whatever a license says. micro1 deletes raw data within 120 days of delivery or at termination, excluding backups (s.4.4).
Labelbox states its kill clause forces deletion within 30 days (checked 2026-11-09).
Models Already Trained: Keep, Retrain or Delete
Most published terms let the buyer keep models trained before deletion, because removing data from a trained model is not something buyers promise. micro1 has no duty to delete or retrain Licensee Materials (s.4.4). Labelbox states "Models already trained on the data cannot be unlearned" (checked 2026-11-09).
Regulators have ordered model deletion. The FTC's Everalbum release (File No. 192-3172, 2021-05-07) requires deleting "the models and algorithms it developed by using the photos and videos uploaded by its users", a remedy called algorithmic disgorgement.
Who Keeps Weights, Outputs and Synthetic Data
Derivative and output clauses decide who owns what the buyer builds from the data, and buyer terms assign all of it to the buyer. micro1 owns derivatives, models, weights, outputs and synthetic data (s.2.2). Tempus' filing shows the reverse: Tempus receives a license to the model Pathos trains.
Attribution, Citation and Output Limits
Attribution and output clauses appear mainly in publisher deals, where the buyer must cite and link the source or cap how much of a work a model reproduces. Axel Springer (2023-12-13) gets attribution and links, Hearst (2024-10-08) "appropriate citations and direct links". The Authors Guild states HarperCollins' license caps outputs at 200 consecutive words and/or 5 % of a book (2024-11-19). No company-data example is published.
Money, Confidentiality and Exit Clauses
The last group sets the price mechanics, what stays secret, who else may receive rights, how the deal ends and where disputes go.
Price, Payment Trigger and Clawback
The payment clause fixes the price and the event that triggers it, and in micro1's terms and on Mercor's data page that event is acceptance (micro1 s.3) or delivery, not signing. Mercor states payment upon delivery, typically 2 to 4 weeks after signing (checked 2026-11-09). Four structures appear, with examples below.
- Fixed fee: Endless states one sum agreed upfront (checked 2026-11-09).
- Upfront plus revenue share: Labelbox states both (checked 2026-11-09).
- Multi-year fee: Tempus' 2025 license to Pathos sets USD 200 million over three years, USD 50 million upfront.
- Per record or year: Tempus' Neuro order form of 2026-10-22 sets USD 100 per record; Kopinor NOK 45,000,000 for 2026.
Each price records how much company data is worth to AI companies in one deal.
Confidentiality Duties and How Long They Last
Confidentiality clauses protect the deal terms and the data, and they survive termination for a fixed period. micro1's last 3 years, and indefinitely for trade secrets (s.2.3).
Passing Rights to Affiliates, Successors and Other Licensees
Sublicense and assignment clauses decide whether the buyer can pass the data on, and published grants range from non-sublicensable to sublicensable through multiple tiers. Defined.ai's are non-sublicensable (s.1(a)); micro1's is multi-tier (s.1.2) and assignable in a merger, not to a competitor (s.9.1).
Termination Triggers: Convenience, Breach and Kill Clauses
Termination clauses list who may end the agreement and why, and what ending it does to the data depends on the deletion and model clauses. micro1 allows termination for convenience on 30 days' notice and for breach after a 10-day cure (s.4.1).
Which Law Applies and Where Disputes Are Heard
The governing-law clause picks the law and the forum, and published buyer terms choose the buyer's home state. micro1 picks California law and JAMS arbitration (s.9.3 to 9.4), Defined.ai Washington, Protege and Grepped Delaware, and the Spirit bill of sale New York.
Where Buyers' Published Terms Contradict Their Own Pages
Read the contract, not the program page, because buyers' marketing pages and their terms sometimes say different things. A dated comparison, AI data partnership payouts and program terms compared, tracks them; four gaps checked on 2026-11-09 follow.
- micro1 acceptance: 30 business days in the terms (s.1.3), "30 days" on the data-partnerships page.
- micro1 payment: "We pay in full within 30 days of acceptance" vs "payment will be made within 30 days of invoicing".
- micro1 exclusivity: "exclusively" on marketing pages, a nonexclusive grant in s.1.2.
- Handshake AI: no exclusivity, term, deletion, indemnity or audit stated.
Labelbox's kill clause appears only on its FAQ.
Signing, Negotiating or Walking Away: Benefits, Risks, Downsides and Alternatives
A license brings money for data the company already holds, and the clauses decide how much control it gives up in return. Four lists weigh it below.
Benefits
- Income while keeping ownership, as Turing states (checked 2026-11-09).
- Room for other buyers under a non-exclusive grant.
Risks
- Seller indemnity for claims from the data (micro1 s.8.1, Protege s.4.1).
- Perpetual, irrevocable use (micro1 s.1.2).
- Trained models stay (micro1 s.4.4, Labelbox).
Downsides
- Audit and exclusivity exist only if the order form adds them (micro1 s.4.2).
- Silence for 30 business days counts as acceptance (micro1 s.1.3).
Alternatives
- Negotiate the order form.
- Limit the grant by time or purpose, as Reddit's terms and Tempus' sole-purpose clause do.
- Do not license at all and keep the data.
Options are weighed under alternatives to selling data to AI labs.
Where the Contract Sits in the Process to Sell or License Company Data
The contract comes after valuation and choosing a buyer and before delivery and payment, and the order form is signed last. It is one step to sell or license company data to AI companies, after asking should you sell or license your data. The steps follow.
- Describe the data in the data manifest.
- Pass due diligence in an AI data deal.
- Agree terms by negotiating an AI data licensing deal.
- Sign the agreement and order form.
- Complete delivering licensed data to an AI buyer and security during an AI data sale.
- Invoice per when and how AI data buyers pay.
- Monitor deletion after the sale.
Twelve clauses have their own pages, as of 2026-11-09.
| Clause | Page | Source |
|---|---|---|
| Acceptance | deemed acceptance of delivered data | micro1 s.1.3 |
| Audit | audit rights in AI data licenses | Defined.ai |
| Deletion | deletion, return and machine unlearning clauses | micro1 s.4.4 |
| Derivatives | ownership of derivatives and trained models | micro1 s.2.2 |
| Disputes | governing law, arbitration and disputes | micro1 s.9.3 |
| Exclusivity | exclusive vs non-exclusive data licenses | Labelbox, micro1 |
| Indemnity | indemnities and warranties | micro1 s.8.1 |
| Liability | liability caps | micro1 s.7.2 |
| Retraining | retraining and model retention rights | micro1 s.4.4 |
| Scope | field of use and scope limits | Tempus 10-K |
| Sublicensing | sublicensing and onward transfer | micro1 s.1.2 |
| Term | license term and perpetual licenses | Reddit S-1 |
Not Legal Advice, and the EU Rule on Unfair Data Terms
This page explains clauses for information and is not legal advice; a lawyer in the seller's jurisdiction should review any agreement before signing. Data Act Art. 13(1) (Regulation (EU) 2023/2854) makes unfair data terms that one enterprise imposes on another non-binding, for contracts after 2025-09-12; Recommendation (EU) 2026/2394 of 2026-10-29 adds non-binding model terms, covered under Data Act Article 13 unfair terms.
Contracts in a Court-Supervised Data Sale
When an estate sells data under court supervision, the contract is a bill of sale with few seller promises, and the court order replaces most negotiation. Spirit's form sells as is under 11 U.S.C. 363. Google LLC holds the successful bid of USD 10,000,000, not approved as of November 9, 2026; hearing November 17, 2026 (Dkt. 1779). The pattern applies to data assets in insolvency.
Published Terms and Deals Cited on This Page
The examples above come from these documents, dated as shown.
| Party | Document | Date |
|---|---|---|
| 23andMe and GSK | 8-K exhibit | 2022-01-18 |
| Axel Springer, Hearst | releases on OpenAI | 2023-12-13, 2024-10-08 |
| Defined.ai | Data License Agreement, FAQ, Supplier Program terms | checked 2026-11-09 and 2026-11-10 |
| Endless | home page and FAQ | checked 2026-11-09 |
| FTC | Everalbum release | 2021-05-07 |
| Grepped | marketplace terms | 2026-11-08 |
| Handshake AI | license page | checked 2026-11-09 |
| HarperCollins | Authors Guild statement | 2024-11-19 |
| Kopinor, National Library of Norway | agreement | 2025-12-19 |
| Labelbox | Prism home page and FAQ | checked 2026-11-09 |
| Mercor | data page and listings | checked 2026-11-09 |
| micro1 | Data License Master Terms | updated 2026-11-08 |
| Protege | Provider Terms | checked 2026-11-09 |
| Reddit, Google | Reddit S-1 and Google announcement; Reddit 10-Q | 2024-02-22; quarter to 2026-09-30 |
| Spirit Airlines estate | Dkt. 1463; Dkt. 1779 | 2026-08-14; 2026-11-04 |
| Tempus AI, Pathos AI | 10-K FY2025; 10-Q | 2026-02-24; quarter to 2026-09-30 |
| Turing | Project Lazarus page | checked 2026-11-09 |
- micro1 data buyer profile
- micro1 Data License Master Terms clause by clause
- Company data partnership programs
- Who buys AI training data
- AI data licensing deals tracker
Frequently Asked Questions
What is a data license?
A data license is permission from the data holder to use data on stated terms, such as purpose, duration and price, without transferring ownership.
What is an example of a licensing agreement for AI training data?
micro1's Data License Master Terms, updated November 8, 2026, are a published example: a nonexclusive, perpetual, irrevocable grant with a 3x-fees liability cap.
Is there a data licensing agreement template?
DataPacts does not publish a template, because terms depend on the data and the jurisdiction; buyers' published master terms show what a buyer's first draft contains.
Can a buyer keep a model trained on my data after the license ends?
Usually yes: published terms such as micro1 s.4.4 require deletion of raw data but no retraining of models.