AI data license terms explained: permitted use, exclusivity, snapshots and deletion

This guide is general information, not legal advice. Have counsel review any license against your intended use and the laws that apply to you.

The terms at a glance

Permitted use

Permitted use is the list of activities the license allows; unlisted activities are usually not licensed. Name each one you need:

• Training: pretraining or continued pretraining.
• Fine-tuning: supervised fine-tuning, preference training and reinforcement learning.
• Evaluation: benchmarking and regression testing, internal or published.
• Retrieval: indexing records for use at inference time, which exposes them to users more directly than training does.
• Synthetic data generation: using records as seeds or references for generated data.
• Commercial deployment: offering models trained on the data in products, not only internal research.

Also settle who may touch the data: affiliates, contractors such as annotators, and cloud or model providers, including third-party fine-tuning APIs.

Field of use

Field of use limits where the permitted uses apply: a domain, product line or market, such as support automation for financial services. It can also exclude uses, such as products that compete with the licensor. A narrow field may cost less but can follow your model for life, so check it against products you may build later.

Exclusivity and dataset snapshots

Exclusivity limits whom else the licensor may license the same data to. A non-exclusive license leaves the licensor free to license the same records to other buyers; an exclusive license rules that out within the scope it defines. The licensor keeps using its records to run its business, so exclusivity concerns other licensees and, where the license says so, the licensor's own AI development.

Exclusivity is a scope, not a yes-or-no. It can be limited to:

• A dataset snapshot: a defined extract, such as records from named systems over a set date range, frozen at delivery.
• A permitted use or field: for example, training for one domain.
• A period: after which others may license the snapshot.

Unless the license says otherwise, exclusivity over a snapshot does not cover records created after it or uses outside the exclusive scope. Buyers seek it to differentiate their models and to keep evaluation snapshots out of competitors' training data. Some SourceX programs include exclusivity for an agreed dataset snapshot or permitted use over a defined period; it is negotiated in the license and usually affects price.

Duration and territory

Duration is how long you may use the data, whether a fixed, renewable or perpetual term. Separate two questions: how long you may use the raw data, and what survives the end of the term, such as models trained during it and the right to keep serving them. Check whether the license transfers if your company is acquired.

Territory is where the data may be stored, accessed and used. Personal data leaving the European Economic Area needs a GDPR transfer mechanism, such as an adequacy decision or standard contractual clauses. Export controls and the licensor's customer contracts can also restrict access and location.

Derivatives and model ownership

Derivatives are anything you create from the data: cleaned and relabeled versions, annotations, embeddings, evaluation sets, synthetic data, model weights and outputs. A common split is that the licensor owns the data and any derivative that contains or can reconstruct records, while you own your models, outputs and annotations. Do not assume it; get explicit answers on three points:

• Whether models trained during the term may be used and commercialized after it ends.
• Whether synthetic data and evaluation sets built from the records count as the data, with its restrictions and deletion duties, or as your derivatives.
• Whether you must take reasonable steps to stop models reproducing licensed records verbatim.

Retention and deletion

Retention and deletion clauses say what you must destroy, when and how you prove it. Expect deletion of raw data and copies when the license ends, with a written deletion certificate and a carve-out for backups that expire on normal rotation. Negotiate whether deletion reaches derived datasets that contain records, and whether it reaches models, since deleting weights after training is costly.

Personal data raises the stakes. The European Data Protection Board has said that AI models trained on personal data cannot in all cases be considered anonymous and must be assessed case by case (EDPB Opinion 28/2024), one reason to de-identify thoroughly before delivery. Also agree how mid-term removal requests work, for example when a licensor's customer withdraws authorization.

Audit

An audit clause lets the licensor check your compliance. Agree the scope (usage records, access logs and deletion records, not model weights or unrelated systems), notice, frequency, whether an independent auditor does the work, and who pays; often the licensor pays unless the audit finds a material breach.

Attribution and confidentiality

Attribution terms say whether you may, or must, name the data source; many licensors prefer not to be named. Confidentiality clauses treat the data, and often the deal, as confidential, so check that they still allow what you need to disclose: evaluation results or examples, training data descriptions in model documentation and, if you provide a general-purpose AI model in the EU, the public summary of training content that the EU AI Act requires.

Warranties and indemnities

Warranties are promises about the data; indemnities say who pays if a third party makes a claim. Ask the licensor to warrant that it may license the data for the permitted use, that collection followed applicable law and the notices given at the time, that de-identification followed the agreed spec, that no earlier grant conflicts with yours, and that the delivery matches the manifest. You will usually warrant that you stay within scope, do not re-identify anyone and keep the data secure.

Indemnities typically follow the warranties: the licensor covers claims that it lacked rights, and you cover claims arising from misuse. Check liability caps and whether rights and privacy indemnities sit outside them. With an intermediary, confirm who gives each warranty: the data owner, the intermediary or both. Licensors usually disclaim fitness for purpose, so quality protection comes from acceptance criteria and remedies such as re-delivery.

Refreshes and updates

A license can cover one snapshot or a series of deliveries. For refreshes, agree the cadence, a stable schema and pseudonym mapping so IDs join across deliveries, the same de-identification, pricing, whether exclusivity extends to new data, and how corrections and takedowns of earlier records work.

Payment structures

Payment terms decide when money moves and who carries the risk that the data disappoints. Common structures, often combined:

• One-time fee for a snapshot: simple, but quality risk sits with you after delivery.
• Milestones tied to delivery and acceptance of each tranche: shifts some quality risk back to the licensor.
• Subscription for ongoing refreshes: spreads cost, with renewal risk on both sides.
• Volume-based pricing per record, hour or file: needs a precise counting rule.
• Exclusivity premium: an added fee for exclusive scope.
• Royalties or revenue share: hard to administer, because model revenue is hard to attribute to one dataset.

Where SourceX fits

SourceX agrees scope, permitted use, pricing and obligations in writing for each dataset it sources. Delivery follows the partner's approval, an executed agreement and explicit authorization of buyer access. Supply is not guaranteed. Before signing, work through the due diligence checklist; the full process is in how to license proprietary data for AI training. To start, send a data request with your licensing requirements.