Most image-data agreements are variations on three shapes. Knowing which one you are drafting settles half the clauses before negotiation starts.
An exclusive licence of copyright in the United Kingdom has a statutory definition. Section 92 of the Copyright, Designs and Patents Act 1988 describes it as a licence in writing, signed by or on behalf of the copyright owner, authorising the licensee to exercise a right to the exclusion of all other persons, including the person granting it. That last phrase is the one clients forget. An owner who grants a true exclusive licence for AI training is, within the scope of that licence, shut out from training a model on the same images itself.
Most agreements for image data fall into one of three shapes: non-exclusive, exclusive and evaluation-only. A fourth, the first-look arrangement, has appeared where archives are still being digitised. This article compares them for counsel advising buyers and rights holders.
This article is general information, not legal advice. Checked on 18 November 2026.
Why a licence at all? The position on 18 November 2026
In the UK the statutory starting point is clear even though much else is not. The only text-and-data-mining exception, section 29A of the 1988 Act, has been in force since 1 June 2014 and is confined to non-commercial research. The government's report of 18 March 2026 dropped the proposed opt-out exception and proposed no legislation; its May 2026 response to a House of Lords committee said AI developers must seek permission and acquire licences unless an existing exception applies, as summarised by Reed Smith on 30 September 2026.
What is pending is narrower than often reported. In Getty Images v Stability AI, decided on 4 November 2025, the High Court did not rule on whether training infringes UK copyright, because that claim was abandoned at trial for want of evidence that the copying happened in the UK. It held that the trained model was not an "infringing copy". The Court of Appeal heard the appeal on that point on 11 and 12 November 2026, and no judgment could be confirmed on the date above. In the EU, Article 53 of the AI Act has required providers of general-purpose models to maintain a copyright policy and publish a training-content summary since 2 August 2025.
For present purposes the consequence is practical. A developer that wants documented, attributable image data for a commercial model obtains it by contract, and the contract's shape determines what it can say about that data afterwards.
The three shapes compared
| Non-exclusive | Exclusive | Evaluation-only | |
|---|---|---|---|
| Purpose | Training or fine-tuning alongside other licensees | Training or fine-tuning with a period of sole access | Inspecting and testing before a decision |
| Who else may use the data | Other licensees and the owner | No one, within the defined field and term | Unaffected; owner remains free to license |
| Typical term | Fixed term, often renewable | Shorter and time-limited | Weeks to a few months |
| Use permitted | As defined: training, fine-tuning, evaluation | As defined, within a stated field of use | Internal benchmarking; no incorporation into model weights |
| Formalities (UK) | Writing strongly advisable | Must be in writing, signed by or for the owner | Writing strongly advisable; often paired with an NDA |
| Main drafting risk | Vague definition of permitted use | Scope of exclusivity wider than intended | Leakage of test data into training |
What is a non-exclusive licence, and when does it suit?
This is the default across the market. The large libraries license the same content to several developers. Getty Images' data-licensing page describes curated datasets for model training alongside display and retrieval uses; the point of a library at that scale is that many customers draw on it.
For a buyer, non-exclusive terms are cheaper and quicker. The trade-off is that the data confers no advantage a competitor cannot also buy. For a licensor, the drafting burden is in the definition of use. The Adobe arbitration reported by PetaPixel on 11 March 2026, in which a 2018 stock contributor agreement was read as covering Firefly training, is a reminder that broad grant language will be read broadly. Our article on how AI image licensing deals are structured sets out the four uses a modern grant clause distinguishes: training, fine-tuning, evaluation, and display or retrieval.
What does exclusivity actually buy?
Exclusivity gives one developer a period in which no rival has the same data. It is most coherent where the collection is distinctive and small enough for sole access to mean something, which is the opposite of the stock-library case.
Counsel should pin down five things:
- Field. Exclusive for what: image generation, all machine learning, a named product line?
- Subset. The whole collection, or defined categories?
- Term. A fixed period, with reversion to non-exclusive or expiry.
- The owner's own position. A statutory exclusive licence shuts out the grantor. If the owner needs to keep licensing the same images for editorial reproduction, the exclusivity must be confined to AI uses in terms.
- Standing. Under the 1988 Act an exclusive licensee has rights of action concurrent with the owner's. Both sides should understand who may sue and who controls proceedings.
A sole licence, under which the owner may use the material but will appoint no other licensee, is sometimes the better fit and should be labelled as such.
What are evaluation-only and first-look licences?
An evaluation licence is the smallest of the three and often the most useful. It lets a technical team run real checks on real files: resolution, duplication, caption quality, overlap with data the model has already seen. It grants no right to train.
Clauses to look for:
- a closed list of permitted activities, with training and fine-tuning expressly excluded;
- a ban on incorporating the data, or anything derived from it, into model weights;
- named individuals or a limited team, and a secure environment;
- a short term and certified deletion at its end;
- an NDA covering both the sample and the documentation supplied with it.
There is a point here that is specific to test data. An evaluation set is valuable because the model has not been trained on it. A licence that lets held-out images drift into a training pipeline destroys the thing being paid for, so the separation is in both parties' interests.
A fourth shape: first look as material is digitised
Where an archive is partly on film, the data does not all exist yet. A first-look arrangement gives a partner a time-limited option on defined material as it is scanned, sometimes in return for funding or carrying out the digitisation. It is an option with a delivery schedule attached. The drafting questions are what counts as a delivered batch, what technical standard applies, how long the option on each batch stays open, and which of the three shapes above the option converts into.
Clauses that matter in every shape
- Title warranty, and its limits. The licensor warrants what it owns. Third-party rights in what the images show are a separate matter, addressed by a documented review: see what a rights review covers before images are licensed for AI.
- Exclusions. Categories removed from the licensed set, listed in a schedule.
- Outputs. Whether reproducing licensed images, or prompting by the photographer's name, is prohibited.
- Retention. What happens to copies and to trained models at expiry.
- Transparency co-operation. The licensee may need facts about the data for regulatory summaries; the licence can say what the licensor will supply.
- Metadata. Whether rights fields must be preserved in delivered files. The IPTC "Data Mining" property can record a prohibition in published copies; the AI training opt-out metadata tool shows how the value is written.
How this applies at Archive.London
The Robert Harper Photograph Library is 95,487 unique photographs by inventory count, 42,803 digital and 52,684 analogue, with collections ranging from still life to portraits. It is owned by 825 Ventures Ltd, the owner and rights holder, so each of the shapes above can be discussed with one counterparty. No clearance claim is made for any image, for AI training or otherwise. Licences are negotiated directly with the single rights holder, subject to rights review; the rights and stewardship page says more about how the archive is looked after.
Counsel and procurement leads who want the dataset card, or to discuss an evaluation-only licence as a first step, can get in touch through the AI training data licensing page.
Frequently asked questions
What is the difference between an exclusive and a non-exclusive AI training data licence?
A non-exclusive licence lets the owner grant the same rights to other licensees. An exclusive licence gives one licensee the stated rights to the exclusion of everyone else, including the owner, usually for a limited term and field of use. Under UK law an exclusive licence must be in writing and signed by or on behalf of the copyright owner. Exclusivity costs more and needs a tightly drafted scope.
What is an evaluation-only licence for a dataset?
It is a short, limited licence allowing a prospective buyer to inspect and test data without using it to train or fine-tune a production model. Typical terms restrict use to named personnel and internal benchmarking, forbid incorporating the data into model weights, set a period of weeks or months, and require deletion at the end. It often precedes a full training licence.
Does an AI training licence need to be in writing?
In the UK an exclusive licence must be in writing signed by or on behalf of the copyright owner to have statutory effect. A non-exclusive licence need not be, but an unwritten licence for AI training would be unwise for both sides: the scope of permitted use, retention and outputs are exactly the points on which disputes arise. Take advice on your own agreement.
Can an AI training licence cover images that have not been digitised yet?
Yes, by agreement. A licence or option can be framed around future deliveries, for example a first look at defined subject categories as film is scanned, sometimes linked to the licensee funding or supporting digitisation. The contract should define the material, the delivery standard, the time the option stays open, and what rights review applies to each batch before delivery.
Sources
- Copyright, Designs and Patents Act 1988, section 92 (exclusive licences)
- Copyright, Designs and Patents Act 1988, section 29A (text and data analysis for non-commercial research)
- Reed Smith: UK copyright and AI training, where the law stands (30 September 2026)
- Getty Images v Stability AI 2025 EWHC 2863 (Ch)
- Burges Salmon: Getty Images v Stability AI, Court of Appeal hearing (12 November 2026)
- Regulation (EU) 2024/1689 (AI Act), Article 53
- Getty Images Q4 2025 earnings call transcript (16 March 2026)
- Getty Images: data licensing
- PetaPixel: he tried to stop Adobe from training its AI on his photo library (11 March 2026)
This article is general information, not legal advice. It reflects the position as we understood it on the date shown and may be out of date. Published by Archive.London. Corrections are welcome under our editorial policy.

