Per-image fees, lump sums, revenue shares and display-only arrangements: the working parts of an AI image deal, and what the public record does and does not tell rights holders.
When Shutterstock's chief financial officer told Reuters in April 2024 that the company's first data deals with large technology companies were worth $25 million to $50 million each, he gave rights holders one of the very few hard numbers this market has produced. Most AI image licensing is done under non-disclosure agreements. What can be pieced together from earnings calls, court papers and trade reporting is less a price list than a set of recurring structures.
This article describes those structures for rights holders weighing an approach. This article is general information, not legal advice. Checked on 18 November 2026.
What does an AI image licensing deal actually grant?
Every deal starts with a definition of use, and the definitions have become more precise as the market has matured. Four distinct grants now appear.
- Training. Copying images to pretrain a model.
- Fine-tuning. Copying images to adapt an existing model to a domain or style.
- Evaluation. Using images to test a model, with no right to train on them.
- Display and retrieval. Fetching images at query time and showing them to a user, usually with credit. The model does not learn from them.
The last of these is easily mistaken for the first. Getty Images announced a multi-year agreement with OpenAI on 21 June 2026 that puts its editorial photographs into ChatGPT's search features; it was reported as display-only, with no training rights, following a similar arrangement with Perplexity in October 2025. On Getty's fourth-quarter 2025 earnings call on 16 March 2026, chief executive Craig Peters said: "We do not license out our editorial content in any way, shape, or form with respect to AI training."
The grant clause is where old contracts cause trouble. PetaPixel reported on 11 March 2026 that the founder of a stock archive had failed in an arbitration against Adobe over the use of nearly 12,000 images to train Firefly; a contributor agreement signed in 2018 was read as covering it. PetaPixel reported that he was asking a California court to vacate the award. Libraries have responded by drawing the line explicitly. Bridgeman Images' published restrictions bar use of its content, captions, keywords and metadata for machine learning or AI unless Bridgeman has specifically authorised it in the licence. AI use becomes a separately negotiated right.
How are AI image licences priced?
Four payment models can be seen in public reporting.
| Model | How it works | Reported example | Source and date |
|---|---|---|---|
| Lump sum for a library | One fee for access to a large corpus over a term | Shutterstock: $25m to $50m per initial deal | Reuters, 5 April 2024 |
| Per-asset fee | A rate per image delivered | Freepik: 2 to 4 cents per image for most of a 200 million image archive | Reuters, 5 April 2024 |
| Revenue share | Data owner paid when its data is used or sold on | Protege: partners receive revenue-share payouts on use | SiliconANGLE, 8 January 2026 |
| Attribution-based | Payment linked to measured influence on outputs | Bria: ongoing compensation for works exceeding relevance thresholds | Bria's website, read 18 November 2026 |
Further per-image figures from the same Reuters report give a sense of the spread. Photobucket's chief executive said rates under discussion ran from 5 cents to $1 per photo. Defined.ai's chief executive said companies were generally willing to pay $1 to $2 per image. Writing for Kaptur on 12 June 2025, Paul Melcher cited bulk datasets on offer at roughly $0.01 to $0.25 per image depending on resolution, subject matter and exclusivity, and noted that much of the market operates under NDA, without pricing transparency.
Three cautions apply to all of these. They are two or more years old in a fast-moving market. They describe bulk, non-exclusive stock and user-generated content. And none is a rate for a curated, single-owner photographic archive, for which no public figure exists. Large composite deals blur the picture further: Getty told investors in March 2026 that two multi-year agreements signed in 2025 were worth about $65 million combined, but its management said they were not pure data-licensing deals.
Exclusive or non-exclusive?
Bulk library deals are non-exclusive by nature: the same images are licensed to several buyers. Smaller, distinctive collections can offer something a stock library cannot, which is a period during which only one developer has the data. Exclusivity is paid for, limited in time, and usually confined to a field of use. A third option, the evaluation-only licence, lets a buyer test before committing. The trade-offs are set out in exclusive, non-exclusive, evaluation-only: three licence shapes for image data.
What terms do buyers and sellers negotiate hardest?
Beyond price and scope, six clauses take most of the drafting time.
- Warranties. The buyer wants the licensor to warrant that it owns the rights it is granting. A licensor is generally in a position to warrant its own title; licensors are usually slow to warrant matters outside their knowledge, such as releases from every person pictured. What to warrant in a given agreement is a matter for professional advice.
- Exclusions. Which categories are left out: identifiable people, branded goods, third-party artwork. These follow from the rights review described in what a rights review covers before images are licensed for AI.
- Outputs. Whether the licensee may generate images that reproduce licensed works, or that are prompted by the photographer's name.
- Retention and deletion. Whether source files must be deleted at the end of the term, and what happens to models already trained. A trained model cannot simply have a dataset removed from it, so the clause needs to say what is expected.
- Security and onward transfer. Who may hold the files, where, and whether subcontractors or affiliates are included.
- Reporting and audit. How usage is evidenced, which matters most where payment depends on it.
Metadata is part of the negotiation too. Licensors generally want creator and copyright fields kept intact in delivered files, and it is good practice to embed them before delivery; the add copyright metadata tool writes creator, copyright notice and licensor details into a JPEG in the browser.
Who is on the other side of the table?
Direct deals with the largest developers have been reported at the scale of hundreds of millions of assets. Smaller rights holders more often meet intermediaries. Wirestock, which raised a $23 million Series A announced on 14 May 2026, told TechCrunch it supplies six large foundation-model makers from a contributor base of more than 700,000 and has paid contributors $15 million. Protege, which raised $30 million from Andreessen Horowitz in January 2026, shares revenue with data partners, though its reported holdings are video, audio and medical data. A map of the field is in after Getty and Shutterstock: who sells licensed image data in 2027.
An intermediary brings buyers and takes a share. A direct licence keeps the relationship and the terms in the rights holder's hands, at the cost of doing the selling.
Why licensing is the route in the UK
For UK rights holders the statutory backdrop has not changed, though the courts have not finished with it. The only text-and-data-mining exception in UK copyright law covers non-commercial research. The government's report of 18 March 2026 dropped the proposed opt-out exception and proposed no legislation, and its response to a House of Lords committee in May 2026 said that AI developers must seek permission and acquire licences unless an existing exception applies, as summarised by Reed Smith on 30 September 2026. What remains pending is the Court of Appeal's judgment in Getty Images v Stability AI, heard on 11 and 12 November 2026, on whether a trained model can be an "infringing copy". That case does not decide whether training itself infringes; the claim on that point was abandoned at trial.
How Archive.London approaches a deal
The Robert Harper Photograph Library is a single photographer's archive of 95,487 unique photographs by inventory count, owned by 825 Ventures Ltd, the owner and rights holder. It has depth in studio subjects such as fragrance and jewellery, and more than half of it is still on film. Because there is one owner, a licence can be shaped around the use: non-exclusive, time-limited exclusive, evaluation-only, or a first-look arrangement tied to scanning. The general framework is on the image licensing terms page. No price is published, and no clearance claim is made for any image; licences are negotiated directly with the single rights holder, subject to rights review.
Rights holders comparing notes, and developers who want to open a conversation, can start at the AI training data licensing page.
Frequently asked questions
How much do AI companies pay per image for training data?
Public figures are few and mostly from one Reuters report of 5 April 2024: Freepik's CEO cited 2 to 4 cents per image for bulk deals, Photobucket's CEO discussed 5 cents to $1 per photo, and Defined.ai's CEO said buyers generally pay $1 to $2 per image. These describe bulk, non-exclusive stock or user content. Most deals are confidential, and no public rate exists for curated single-owner archives.
What rights does an AI training licence usually grant?
It defines permitted uses, which may include training, fine-tuning, evaluation, or retrieval and display, and the models or products covered. It sets term, territory and exclusivity, and addresses whether the licensee may keep copies after expiry. A well-drafted licence also states what is not granted, such as reproducing the images in outputs, sub-licensing the data, or using it to imitate a named photographer.
Is a display licence the same as a training licence?
No. A display or retrieval licence lets an AI product fetch and show images, usually with credit, in response to queries; the model does not learn from them. A training licence permits copying images to adjust model weights. Getty Images' agreements with Perplexity in October 2025 and OpenAI in June 2026 were reported as display arrangements without training rights.
Should a rights holder accept a revenue share or a flat fee?
Each has a cost. A flat fee is certain but fixes the price before anyone knows the data's value. A revenue share or attribution model follows usage, but depends on the buyer's reporting and on measurement methods the licensor cannot easily verify. Many agreements combine a minimum guarantee with a usage element. The right answer depends on the use, and is a matter for professional advice.
Sources
- Reuters, Inside Big Tech's underground race to buy AI training data (5 April 2024), archived copy
- Paul Melcher, Kaptur: The hidden economy behind AI data licensing takes center stage (12 June 2025)
- Getty Images Q4 2025 earnings call transcript (16 March 2026)
- TechTimes: Getty Images and OpenAI deal (24 June 2026)
- Bria: licensed training catalogue
- SiliconANGLE: Protege raises $30M (8 January 2026)
- TechCrunch: Wirestock raises $23M (14 May 2026)
- PetaPixel: he tried to stop Adobe from training its AI on his photo library (11 March 2026)
- Bridgeman Images: restrictions of use
- Reed Smith: UK copyright and AI training, where the law stands (30 September 2026)
- Burges Salmon: Getty Images v Stability AI, Court of Appeal hearing (12 November 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.

