Training data licence agreement

Data used to train a model is licensed on terms that have to answer questions ordinary content licences never faced: whether the licensee may train, what it may do with the model that results, whether the licence survives in the weights after it ends, what happens to the personal data in the dataset, and what the licensor can warrant about where the data came from. The agreement can be drafted for a licensor with data to sell or a licensee with a model to build. Drafting or reviewing the AI provisions for the agreement is £595 fixed, delivered in five working days.

Who this is for

Businesses in England and Wales licensing datasets (content, images, records, transcripts, labelled data) for AI training, and AI developers acquiring data to train or fine-tune models.

What matters in a training data licence

The rights in a dataset and the two owners it may have

A dataset may be protected as a collection of copyright works (each image, article or recording with its own copyright, owned by whoever created it or took an assignment), as a database under the Copyright and Rights in Databases Regulations 1997 (the database right in a collection that took substantial investment to obtain, verify or present, owned by the maker), or both, and the licensor may own the database right without owning the works in it; the agreement states what the licensor owns, what it has the right to license, and the chain of title or the licences under which it acquired third-party content, because a licensee training on data the licensor could not license is infringing with the licensor's warranty as its only protection.

The scope of the licence, training, derivatives and the model that results

The licence states the permitted uses (training, fine-tuning, evaluation, retrieval-augmented generation, each a different thing), the models and the versions it covers, whether the licensee may sublicense or share the data with its model providers and contractors, whether the licensee may use the data to create derivative datasets, and the position of the trained model: the licensor usually accepts that the model and its weights belong to the licensee and that the licence cannot be unwound from them, in exchange for restrictions (no reproduction of the data as outputs, no use of the model to compete with the licensor in stated ways, no distribution of the data itself) and a payment that reflects the permanence.

Personal data in the dataset and the basis the licensee needs

entering personal data into an AI tool is processing under the UK GDPR and brings with it the need for a lawful basis, a notice to the individuals, an Article 28 data processing agreement with the provider and, where people are affected, an Article 35 impact assessment, and a dataset containing personal data (names in records, faces in images, voices in recordings, writing that identifies its author) makes the licensee a controller for the training with its own lawful basis (legitimate interests, with the balance recorded, in most cases), its own notice under Article 14 (which the Information Commissioner's guidance on generative AI expects developers to address), its own impact assessment, and its own obligations on special category data; the agreement allocates the roles (the licensor as controller for the collection, the licensee for the training), warrants the licensor's own compliance, and prohibits re-identification.

The text and data mining exception and why it does not help a commercial licensee

Section 29A of the Copyright, Designs and Patents Act 1988 permits copying for text and data analysis without a licence only for non-commercial research and only where the person has lawful access to the work, so a commercial developer cannot rely on it, and a government proposal to widen the exception was not enacted; commercial training on copyright works therefore needs a licence, and the agreement states that the licensee relies on the licence and not on any exception, so that the licensor's price is paid for the rights the licensee needs.

Warranties, provenance and the dataset of uncertain origin

The licensor warrants that it owns or has the right to license the data for the permitted uses, that the data was collected lawfully (not scraped in breach of terms, not obtained in breach of confidence, not in breach of data protection law), that it has the consents or notices the data required, that the data does not contain content the licensee has excluded (child sexual abuse material, content under licences incompatible with the use), and that it has disclosed the provenance records it holds; the licensor's indemnity for breach is capped at a level connected to the fee; section 9(3) of the Copyright, Designs and Patents Act 1988 assigns authorship of a computer-generated work to whoever made the arrangements, untested on modern tools, and the Act's protection extends only to original works, which matters where the dataset itself contains generated content.

Payment, audit and the model that outlives the licence

Payment is a lump sum, a fee per record or per use, a revenue share on the model's commercialisation, or a combination, with the audit right that a revenue share needs; the licensor may audit the licensee's compliance with the scope (what was trained, what was shared) on notice and under confidentiality; on termination the licensee deletes the data and certifies it, the trained model survives as the agreement provides, and the restrictions on the model (no reproduction of the data, no competing use) survive with it; confidential information, clients' data and trade secrets are kept out of tools whose terms permit retention of inputs or training on them, because both the Trade Secrets (Enforcement, etc.) Regulations 2018 and the business's confidentiality obligations ask what steps were taken to protect them, and a licensor's dataset is its trade secret until it is licensed, so the agreement also contains the confidentiality and security terms for the data in the licensee's hands.

What it costs

AI provisions in a contract, £595. Review or draft AI-specific terms in a supply or customer agreement. Five working days.

AI use policy, £995. Internal policy plus the accompanying contract provisions. Five working days.

Email me at geoffrey@caesar.co.uk with what you need and any documents, and I will confirm the scope and the fee in writing, usually the same working day.

What you get

• An internal AI use policy your staff will follow, rather than a document that bans everything and is quietly ignored
• Contract provisions covering output ownership, training data, confidentiality and liability
• A defensible position to give customers who ask what your AI governance looks like
• Review of your AI suppliers' terms, including what they reserve the right to change
• Practical guidance on what to permit, what to restrict and what to prohibit outright
• One round of amendments

What is not included

• Technical AI assurance, model testing or bias auditing
• Sector-specific regulatory approval
• Advice on non-UK regulatory regimes beyond flagging where they apply
• Ongoing regulatory monitoring

Questions I am often asked

Can we train on content we have a licence to use for other purposes?

Only if the licence covers training, which most content licences written before generative models do not. The text and data mining exception covers non-commercial research only. A training licence states the uses, the models and the position of the resulting model.

Does the licensor get the model back when the licence ends?

Usually not: the model and its weights belong to the licensee and cannot be unwound, in exchange for restrictions on the model that survive and a price that reflects the permanence. The agreement should say so rather than leave it to argument.

The dataset contains personal data. Whose problem is that?

The licensee's for the training (its own lawful basis, notice, impact assessment) and the licensor's for the collection. The agreement allocates the roles, warrants the licensor's compliance and prohibits re-identification.

Related guidance and services

• AI contracts and AI use policies, £595, the service this page describes
• SaaS and technology contracts, £995
• Data protection agreements and privacy terms, £795
• AI development agreement for a bespoke model
• AI-generated images and copyright in client work

This page is general guidance for businesses in England and Wales, not advice on your own circumstances. Last reviewed: October 2026. Email geoffrey@caesar.co.uk.