The AI licensing economy

This week, the UK government will receive submissions to their consultation seeking views on the legal framework for AI and copyright. At the forefront of issues being consulted upon is whether AI developers should be compelled to seek authorisation (a licence) when they use copyrighted works to train their AI models.

Ahead of the consultation, we collected information on all ‘known commercial agreements’ between content providers (licensors) and AI developers (licensees). The full spreadsheet – which details all 83 agreements – can be explored here.

+ + + UPDATE 31 January 2026 + + +

Our database now contains 274 agreements (Oct 2022 – Jan 2026). The public spreadsheet has been updated and is available here.

+ + + UPDATE 4 June 2026 + + +

The database can now be interrogated as a Sankey diagram, visualising agreement flows by firms and type: https://www.copyrightevidence.org/ai-licensing/

If you use this data, please cite in this form: Kretschmer, M. and Thomas, A. (2026), The AI Licensing economy: Empirical evidence on agreements between creative content firms and AI developers, University of Glasgow: CREATe Centre (AHRC infrastructure grant AH/Y007662/1)

Known: We only looked at agreements available or announced to the public by the parties on official sites or news reports. We consulted a number of public sources, including trackers on Scholarly Kitchen, Press Gazette, and CB Insights. We also checked for completeness with Boolean search strings. Whilst we have taken every measure to ensure the list is comprehensive at the time of writing, we encourage you to let us know if something is missing (contact@create.ac.uk).

Commercial: We only considered agreements which are concerned with making profit. This information doesn’t include speculated or anticipated deals, ‘non-commercial’ agreements between AI developers and their users (e.g., terms of service), nor openly licensed resources like Wikipedia.

Agreements: The scope of agreements collected is broad and in principle can apply to any agreement between a content provider and AI developer. Of the 83 agreements, 78 detailed an explicit licence to use the works of content providers – the remainder either implied that such a licence existed (2) or were part of a fellowship (3) which implied information sharing more broadly.

Public knowledge of commercially sensitive information relating to these deals is inevitably limited, but we can make some observations based on the data gathered.

AI licensing is increasing. Agreements followed the increase in public awareness after the release of ChatGPT in November 2022. Since mid-2024 the number of known commercial agreements is accelerating. This likely reflects the risks inherent in US litigation and in the EU (after the adoption of the AI Act in May 2024). The EU now links rights reservation provisions for text and data mining under copyright law in the Copyright in the Digital Single Market Directive to transparency provisions in the AI Act. Offering generative AI models in major markets comes with risks, regardless of the UK’s policy choices.

Big Tech is not the biggest licensee. OpenAI and Perplexity make for the largest share of known commercial agreements, whereas ‘Big Tech’ companies, like Amazon, Apple, Google, Meta, and Microsoft, have comparatively fewer agreements. This may be because these companies have already been given access to training data which has been ‘licensed’ through existing services, such that more licensing agreements are unnecessary. Users of Facebook, for example, agree to let Meta use their public posts and available data for AI training purposes; in addition to data obtained from search services, maps, and email, users of YouTube likewise agree to let Google use their public videos and available data when they use this service.

The News/Media sector makes for the largest share of licences agreed. 68% of known commercial agreements are with News or Media companies. This is in line with expectations that AI models need regular, up-to-date information to offer the most accurate responses to user queries. However, the prevalence of this sector in these agreements may also raise other interesting questions: for example policy awareness following lobbying by the News/Media sector in the context of a press publishers’ right.

… but image banks seem to have a comparatively larger share of declared revenues (31%). Whilst it’s tricky to calculate exact revenues generated by these agreements, those which have publicly declared figures suggest that image banks (and Scholarly/Publishing at 16%) may be more profitable, despite the prevalence of News/Media in the overall sector-by-sector assessment (noting of course that News/Media content may also include audiovisual content). This may also be attributed to the fact that one particular licensee – Shutterstock – seems to favour the licensing model, with known commercial agreements with seven AI developers (the most of any licensors surveyed).

While agreements may be reported as generative AI training licences, we don’t know if copyright is involved at all. It is likely that many agreements are concluded as data access contracts, and in the case of news, include post-training clauses, such as link-backs or AI tools in the news room. We also don’t know the governing law. This matters for how royalties are treated with respect to authors.

It’s unclear how money from these licences will reach primary creators. While many of these agreements promise that revenues from AI usage of their work will be used to support primary creators, only two licensees have specific methods to ensure this: Shutterstock, through their Contributor Fund, and HarperCollins (which will give authors approx. £2,000 if they consent to the use of their book for AI training). As most revenue sharing methods are assessed per usage of the individual content (65%) these specifics would seem important if they are to fulfil this promise.

The UK government’s favoured option for reform is an ‘opt-out’ model; essentially, that AI training based on copyrighted works should only be permitted so long as the rightsholder has not reserved their rights (i.e., opted-out). This would be bolstered by transparency measures that would allow rightsholders to manage the use of their works by AI developers. Opt-outs are designed to increase the data available for training, while transparency is supposed to facilitate licensing.

However – this data indicates – that the AI licensing economy is already happening, and at scale. Opting-out may offer an illusion of control, while the rewards flow in predictable ways.

With this in mind, how do we re-frame the priorities for AI legislation? Among these, we might point to mechanisms that would ensure any distribution of monies already secured through these licensing agreements should reach creators, especially amid evidence of low and declining incomes.

In the coming days, we will publish our response to the UK government’s consultation which outlines the evidence on AI and copyright to date, and our suggestions for legislative change.

Read our other policy responses here, and see our recent post with the Opinion by the European Copyright Society on Generative AI.

We have published widely on AI law, with recent substantive highlights including (1) Private Ordering and Generative AI: What Can we Learn From Model Terms and Conditions?, and (2) Is AI replacing creative work? New study finds complexity and continuities.