Data licensing for AI training
Data license vs data use agreement vs data sharing agreement for AI
Quick answer
A data license grants rights to use data, and for a commercial AI deal it is usually the core contract: it names the records, the allowed uses (training, fine-tuning, evaluation, retrieval), the term and the delivery. A data use agreement (DUA) mainly restricts how restricted data may be handled, often for research or HIPAA limited data sets. A data sharing agreement sets out an exchange between organizations. A DPA, an NDA and an evaluation agreement sit beside these and do not grant training rights on their own.
By SourceX Editorial · Updated
This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.
What each agreement actually does
Each agreement answers a different question, and the right one depends on whether you need rights, restrictions, a processing mandate or confidentiality. Leigh Dodds sorts data terms into standard licenses that apply without a signature and negotiated agreements used when data is restricted [1]. University guidance describes DUAs as the instrument for non-public data, often carrying duties such as GDPR compliance and a ban on re-identification [2]. Even naming matters: Microsoft called its open data terms an "agreement" rather than a "license" because intellectual property rights in data differ from country to country [3].
| Agreement | Main question it answers | Typical grantor | Grants AI training rights? | Common AI-deal failure mode |
|---|---|---|---|---|
| Data license (DLA) | What may the buyer do with these records, for how long? | Data holder | Yes, if the grant says so | Grant says "internal business purposes" and never mentions model training or model outputs |
| Data use agreement (DUA) | Under what restrictions may the recipient handle restricted data? | Data custodian, often a research or health institution | Rarely; usually limited to a named project | Research-only purpose clause blocks a commercial model |
| Data sharing agreement (DSA) | How do two or more parties exchange data, and who is responsible? | Each party, often reciprocally | Only if a use clause covers it | Roles and ownership of derived data left vague |
| Data processing agreement (DPA) | How does a processor handle personal data on a controller's instructions? | Controller instructs processor | No; it limits processing to instructions | Vendor trains its own model on client data and becomes a controller |
| NDA or evaluation agreement | Can the buyer inspect a sample confidentially before buying? | Data holder | No, or evaluation only | Team fine-tunes on the sample before a license is signed |
| Business associate agreement (BAA) | May a vendor handle PHI for a covered entity's functions? | HIPAA covered entity | No; services for the covered entity only | Assumed to permit product training |
For definitions, see the glossary entries on data licensing and the data processing agreement.
When a data license is the right primary contract
A data license is the right primary contract whenever you want to own the outcome of training: weights, embeddings, evaluation results and the products built on them. Law-firm guidance on data licensing stresses that the grant should define the data, the scope of use and who holds rights in anything derived from it [8]. For AI, that means the grant must spell out training, fine-tuning, evaluation and retrieval, and must address the model rights that survive license termination.
Standard public licenses such as the Community Data License Agreements are licenses too, written so data can be shared without negotiation [9]. Their limits differ: compare them in the open data license compatibility matrix and the CDLA-Permissive and CDLA-Sharing explainer. For proprietary operational data, the license is negotiated, and the AI training rights grant clause is where most of the value sits. For the wider picture, start with the AI training data licensing guide.
Why a data use agreement rarely carries commercial AI rights
A DUA is built to restrict, so it rarely gives a commercial AI team enough rights to ship a model. Research DUAs typically name a project, a principal investigator, approved users and a destruction date, and bar re-identification and onward sharing [2]. Those terms fit a study, not a model that will be retrained, versioned and sold.
Research corpora show the pattern. ImageNet's terms of access limit use to non-commercial research and educational purposes and disclaim warranties [6]. The Linguistic Data Consortium handles commercial access through a separate for-profit membership agreement rather than the research terms [7]. If you hold research-only data and need commercial rights, see moving from research-only to commercial licenses.
HIPAA limited data sets: the one DUA that the law requires
Under HIPAA, a DUA is mandatory when a covered entity discloses a limited data set, and that DUA cannot be swapped for an ordinary license. A limited data set is still protected health information: it removes direct identifiers listed in 45 CFR 164.514(e)(2) but may keep dates and some geography, and it may be disclosed only for research, public health or health care operations under a DUA meeting 164.514(e)(4) [4]. The DUA must set permitted uses, name who may use or receive the data, and bar re-identification and contacting individuals [4].
Fully de-identified data under 164.514(a)-(b), through Safe Harbor or Expert Determination, is no longer PHI, so a commercial license can govern it [4]. Read HIPAA limited data sets and DUAs for AI development and BAA or data license for health data before you pick a route.
The DPA trap: processing mandates are not training rights
A DPA cannot authorize a vendor or buyer to train its own model, because a processor may act only on the controller's documented instructions [5]. GDPR Article 28(3)(a) states that duty, and Article 28(10) says a processor that determines the purposes and means of processing is treated as a controller for that processing [5]. A team that receives client data under a DPA and then trains a general model has stepped outside the processor role.
In a licensed-data deal, the buyer usually receives data as an independent recipient, not as the supplier's processor. If any personal data remains, counsel will usually want a controller-to-controller data sharing agreement or license terms covering lawful basis, transfer mechanisms and data subject requests. See de-identified, anonymized and pseudonymized definitions for when those duties apply.
When you need several agreements together
Most commercial AI data deals use a stack of agreements, signed in sequence. A typical order is NDA or evaluation agreement first, then the data license (often a master agreement plus order forms), with a DUA, DPA or BAA added only where the data type or law requires it.
- Before purchase: an NDA or evaluation agreement covers sample review; see evaluation licenses and NDAs for dataset samples.
- Core deal: the data license, or a master data license agreement with order forms for recurring purchases.
- Health data: a HIPAA limited data set DUA, or a de-identification evidence package so a license can govern; see the de-identification evidence package checklist.
- Residual personal data: controller-to-controller terms, not a processor DPA.
- Ongoing feeds: refresh and delivery terms; see ongoing data supply agreements.
Make sure the documents agree. A precedence clause should say which controls if the DUA's purpose limit conflicts with the license's grant; otherwise the narrower term will usually be argued to apply.
Checklist: what each agreement must contain for AI use
Use this checklist to test drafts before they reach counsel. A missing row is a gap to raise, not proof the deal fails.
Illustrative example: invented to show structure; it does not describe an available dataset.
| Clause | Data license | DUA | DSA | DPA | NDA / evaluation |
|---|---|---|---|---|---|
| Record definition (tables, fields, date range, file formats such as Parquet or JSONL) | Required | Required | Required | Categories of data | Sample only |
| Use grant naming training, fine-tuning, evaluation, RAG | Required | Usually absent | If AI use intended | Not applicable | Evaluation only, no training |
| Rights in models, weights, embeddings and outputs | Required | Often silent | Required | Not applicable | Prohibit or exclude |
| Purpose limit and approved users | Optional | Required | Required | Instructions | Named reviewers |
| Re-identification and contact bans | If de-identified | Required | If personal data | Security terms | If personal data |
| Term, termination and post-term model retention | Required | Destruction date | Required | Return or deletion | Destroy sample |
| Audit and usage reporting | Recommended | Common | Common | Required (Art. 28) | Rare |
| Warranties on ownership and consents | Required | Rare | Mutual | Not applicable | Disclaimed |
| Precedence over other agreements | Required | Required | Required | Required | Required |
Go deeper on audit and usage-reporting rights, data warranties and the negotiation checklist.
How SourceX structures commercial AI data deals
SourceX sources operational datasets from US companies on request and manages the commercial process, from licensing agreements to ongoing purchases. Its process runs Find, Assess (data and licensing permissions), Agree (pricing and allowed uses in a license), Transact and Manage, and nothing is contracted until a supplier agrees. Every dataset is rights-reviewed for ownership and consents and delivered under a license defining records, uses, term and delivery. Health records require HIPAA de-identification by Safe Harbor or Expert Determination. Buyers can describe the data they need on the SourceX buyers page.
Choosing the right agreement for your AI data deal
For commercial AI training, put a data license at the center and add DUAs, DPAs or BAAs only where the data or law demands them. SourceX manages licensing agreements and ongoing purchases for operational datasets sourced from US companies, and every release is approved by the supplying company. Start on the SourceX buyers page.
Frequently asked questions
Is "DLA" a standard term?
No single standard defines "DLA". Teams use it loosely for data license agreement; what matters is whether the grant clause names AI training and model rights, not the title on the document.
Can a DUA be amended to allow commercial training?
Sometimes, if the custodian has the rights and the underlying consents or law permit it. A HIPAA limited data set DUA cannot extend beyond research, public health or health care operations [4], so commercial use usually requires de-identified data under a license.
Does a data sharing agreement replace a license?
Only if its use clause grants the rights you need. Many DSAs focus on governance and security and stay silent on models and outputs, so add explicit AI terms or sign a separate license.
Sources
- Leigh Dodds, "Three types of agreement that shape your use of data" (2020). https://blog.ldodds.com/2020/02/21/three-types-of-agreement-that-shape-your-use-of-data/
- Radboud University, "Licenses and data use agreements". https://ru.nl/rdm/vm/licenses-data-use-agreements
- Microsoft, "Open Data Campaign backgrounder and FAQ" (2021). https://news.microsoft.com/wp-content/uploads/prod/sites/560/2021/03/Backgrounder-FAQ-Sheet_FINAL.pdf
- eCFR, Office of the Federal Register / HHS, "45 CFR 164.514 - Other requirements relating to uses and disclosures of protected health information" (2026). https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-E/section-164.514
- gdpr-text.com, "GDPR Article 28: Processor". https://gdpr-text.com/en/read/article-28/
- ImageNet (Princeton University and Stanford University), "ImageNet download and terms of access". https://image-net.org/download.php
- Linguistic Data Consortium, University of Pennsylvania, "LDC for-profit membership agreement". https://Catalog.Ldc.Upenn.Edu/license/ldc-for-profit-membership.pdf
- Mayer Brown, "Data licensing tips and tactics" (2018). https://www.mayerbrown.com/ja/insights/publications/2018/05/data-licensing-tips-and-tactics
- The Linux Foundation (cdla.dev), "CDLA FAQ". https://cdla.dev/faq-resources/faq/
Tell us what your models need
Share scope, volume, language, format, timing and licensing requirements.