AI data licensing glossary: what the terms mean
This glossary defines 38 terms you will meet when licensing company records to an AI buyer, each in one or two plain sentences. Terms that come from law, such as Safe Harbor and the EU training-content summary, name their source.
What could your company's records earn? Get a range in 30 seconds. Free, no obligation.
What do the terms in an AI data license mean?
Terms are in alphabetical order.
| Term | Plain meaning |
|---|---|
| Acceptance | The buyer's check that delivered records match the contract. Payment should follow acceptance within a set number of days. |
| Anonymization | Removing details so that no one can identify the people in the records by any reasonable means. It is stricter than pseudonymisation. |
| Audit right | Your right to check that the buyer uses the records only as the license allows and deletes them when required. |
| Chain of title | The record of who created the data and how your company came to hold the right to license it. |
| Claw-back | A right to take money or records back if something proves wrong, such as personal data found in delivered files or a rights problem. |
| Confidentiality agreement (NDA) | A contract that bars the other side from sharing what it learns. Buyers and sellers sign one before records are reviewed. |
| Data license | A contract that lets a buyer use a copy of your records in set ways while you keep ownership. See AI data licensing. |
| Data processing agreement | A contract that sets how a party handling personal data may use it and must protect it. |
| Data provenance | Where records came from, who made them, and under what permissions. Buyers ask for it before they pay. |
| Data room | A secure online folder or controlled space where a buyer reviews samples and documents before signing. |
| De-identification | Removing or replacing names and other identifiers so the people in records cannot be identified. See de-identification. |
| Deletion certificate | A signed statement from the buyer that it deleted its copy of your records, usually due when the license ends. |
| Evaluation data | Records used to test how well a model performs, not to train it. |
| Exclusivity | A promise not to license the same records to anyone else, for a set period or use. Exclusive deals should pay more. |
| Expert Determination | A US health-privacy method: a qualified expert finds the risk of identifying a person is "very small" and documents why (45 CFR 164.514(b)(1)). |
| Fair use | A US copyright defense that can allow some copying without permission. Whether AI training qualifies depends on the facts of each case. |
| Fine-tuning | Further training of an existing model on a smaller set of records so it does one job better. |
| General-purpose AI model | The EU AI Act's term for a broad model that can do many tasks. Its providers must publish a training-content summary. |
| Indemnity | A promise by one side to cover the other's losses from a named problem, such as a claim that the data was not yours to license. |
| Model weights | The numbers a model learns during training. A license should still say what happens to your records, which the weights do not replace. |
| One-time fee | A single payment for a license, often paid on delivery or acceptance. |
| Perpetual license | A license with no end date. The buyer may keep using the records you licensed, so it should usually pay more than a fixed-term license. |
| PHI (protected health information) | Health information that identifies a person and is held by a covered health provider or plan under US health-privacy law. |
| PII (personally identifiable information) | Details that can point to a specific person, such as a name, email address or account number. |
| Privacy ombudsman | In a US bankruptcy, a person a court appoints to review the privacy effects of selling personal information. Spirit Airlines' data sale had one. |
| Pseudonymisation | Replacing identifiers with codes, such as "Employee 14", while a key that can reverse the codes still exists somewhere. |
| Read-only connection | A link that lets a buyer or tool copy records but not change or delete them in your systems. |
| Referral fee | A fee a buyer pays someone for introducing a company that signs a license. See referral fee benchmarks. |
| Representations and warranties | Statements you promise are true, such as that you own the records and have the right to license them. |
| Revenue share | Payment as a percentage of what the buyer earns from a product built with your records, instead of or on top of a flat fee. |
| RL environment | A simulated workplace where an AI agent practices tasks by reinforcement learning. Builders may base it on real emails, chats and documents. |
| Safe Harbor | A US health-privacy method: remove 18 listed kinds of identifiers and have no actual knowledge the rest could identify someone (45 CFR 164.514(b)(2)). |
| Scope | The records, systems and date ranges a license covers. You decide what is in and out. |
| Sublicense | A right for the buyer to pass the license to another party. Resale to other labs should raise the price or be excluded. |
| Synthetic data | Records made by a program to look like real ones. Some AI builders use them to supplement or replace real records. |
| Term license | A license that runs for a set period, such as 2 years, after which the buyer must stop using and delete the records. |
| Training data | The records a model learns from during training. |
| Training-content summary | A public summary of training content that EU AI Act Article 53(1)(d) requires general-purpose model providers to publish. See training data rules. |
Where do these terms come from?
Most terms come from contract practice, not statute. Two come from laws we name in the table: the US health-privacy rule at 45 CFR 164.514 (Safe Harbor and Expert Determination) and Article 53 of the EU AI Act. HHS says data that meets either health-privacy method is no longer protected health information (HHS guidance on de-identification).
How do I use these terms when I read a contract?
Start with the six that set the price: scope, term, exclusivity, sublicense, payment and acceptance. Then check the protections: representations, indemnity, deletion and audit. The 10 contract clauses to check page walks through them, and what AI data buyers check explains the buyer's side. For a price range, use the calculator.
Questions
What is the difference between de-identification and anonymization?
De-identification removes or replaces identifiers to cut the risk that someone can be identified. Anonymization goes further: no one can identify the person by any reasonable means.
What is a claw-back in a data license?
A claw-back lets one side recover money or records if a named problem appears, such as hidden personal data. Ask who can use it, for how long and for which problems.
Is a perpetual license the same as a sale?
No. You keep ownership of the records. The buyer only gets a right to keep using its copy, with no end date.
Where can I see these terms in a real contract?
See the 10 contract clauses to check for a plain-language walk through each clause.
Sources
- EU AI Act, Article 53(1)(d), AI Act Service Desk, read 21 Nov 2026: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-53
- U.S. Department of Health and Human Services, "Guidance Regarding Methods for De-identification of Protected Health Information in Accordance with the HIPAA Privacy Rule": https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html
- Reuters, "US bankruptcy judge approves Spirit Airlines data sale to Google", 17 Nov 2026 (privacy ombudsman): https://www.investing.com/news/stock-market-news/us-bankruptcy-judge-approves-spirit-airlines-data-sale-to-google-5034262