Copyright Licensing in the AI Era

Legal developments surrounding AI continue to move rapidly as we embark upon 2026, commanding front row attention from across diverse industries and practices, including within intellectual property. As AI continues to develop at breakneck speed, transforming industries and how they do business, so do surrounding legal frameworks and recommended industry standards and best practices. Of particular note are developments in licensing practices for generative AI models, as industries from music to film to photography work to keep pace with the demand for high-quality audiovisual and textual data.

Although there is ongoing litigation nationwide targeting the unauthorised scraping of content from the internet for use as training data in AI models, industries including publishing, film, music and image licensing are already rising to meet the moment by congruously developing and implementing licensing practices for their creative content. The implementation of such practices to conscientiously convey better-quality, clean and compensable datasets signals a growing trend of licensing frameworks facilitating AI. Far from putting the brakes on the emerging technology, licensing serves to augment and implement important oversight over data ingestion – actually fostering AI’s advancement – all within the bounds of copyright law that maintains incentives for human authorship and invention. 

AI’s effect upon existing licensing practices in creative industries

What role can licensing assume in ensuring clean and authorised datasets for use in training AI models? The answer requires understanding the key function that licensing already plays within creative industries. For the visual images industry, for example, licensing is crucial to disseminating a wide range of visual and audio-visual content, including photography, illustrations, audio-visual works, 3D models and their associated digital files and metadata to the media and commercial content creation companies. Licensing ensures that millions of still and video images including their data – often gathered by image aggregator platforms into easily searchable, readily downloadable digital collections – can reach licensees within the downstream industries that require this content. Industries such as editorial news, advertising, marketing, entertainment and media, all rely on images to both relay news and contribute to creative content that reflects our culture and world. Image aggregator platforms either own the visual content or have image partners and contributors whose contractual arrangements permit the licensing and distribution of their content for a share of the revenue, incentivising visual creators to continue to create new works. Licensing is the underpinning of the visual images industry and accounts for billions of dollars’ worth of transactions.

However, the advent of generative AI platforms, and particularly text-to-image AI models such as DALL-E, Midjourney and Stable Diffusion risk disruption to this industry. Not only does unauthorised and uncompensated online scraping of content-for-data bypass a crucial emerging licensing market for artists’ works, for which they may wish to ensure the implementation of guardrails and licence terms, but it undercuts sources of additional revenue, which, for many creators, can mean the difference between continuing to pursue their careers or being pushed out. Additionally, because AI models offer users the ability to generate realistic images from text prompts, this technology further risks disrupting and disincentivising human industry.

AI models in other industries present analogous concerns. Text-to-text and text-to-audio models trained on textual and musical works require the same copious amounts of content-as-data for training purposes, often scraped unauthorisedly, and ultimately poised to supplant markets for human authorship with the advent of AI-generated competing and/or infringing works. Accordingly, there is a push among many creative industries for AI platforms to afford creators the right to consent to the use of their work for training purposes (ie, to “opt in” rather than “opt out”); to attribution where their work develops or contributes to the output of AI systems; to  remuneration for the use of their work in training datasets and where AI-generated outputs create subsequent commercial value; and to  guardrails against digital replication of ingested works.

Creative industries have responded to many AI platforms’ disregard for such rights, where such platforms trained on millions of copyrighted works amassed online without authorisation, with a wave of litigation. Cases include class action lawsuits (and individual actions) brought by visual artists against generative AI platforms, as well as by print and music publishers and authors against large language model platforms, alleging infringement resulting from this unauthorised use of their original works, and which platforms often also permit end-users to create works in an identifiable artist’s style, and sometimes regurgitate works wholesale. Examples include:

• In re Open AI, Inc., Copyright Infringement Litigation (Multidistrict Litigation No 3143 consolidating existing Northern District of California and Southern District of New York cases, date ordered 3 April 2025);
• Dow Jones & Company, Inc v Perplexity AI, Inc (No 1:24-cv-07984) (S.D.N.Y. 21 October 2024);
• Bartz v Anthropic (No 3:24-cv-05417) (N.D. Cal. 19 August 2024);
• Concord Music Group, Inc v Anthropic PBC (No 4:24-cv-03811) (N.D. Cal. 26 June 2024);
• Kadrey v Meta Platforms, Inc (No 3:23-cv-03417) (N.D. Cal. 7 July 2023);
• Getty Images v Stability AI (No 1:23-cv-00235) (D. Del. 3 February 2023);
• Anderson v Stability AI Ltd (No 3:23-cv-00201) (N.D. Cal. 13 January 2023);
• Disney Enterprises, Inc, et al v Midjourney, Inc (2:25-cv-05275) (C.D. Cal. 11 June 2025); and
• Disney v Minimax (2:25-cv-08768) (C.D. Cal. 16 September 2025),

the latter two representing the first significant AI cases brought by major film studios.

Although the many ongoing AI lawsuits are presently situated across different stages of litigation, recent summary judgment decisions and settlements have begun to shed light on prospective trajectories and trends. The parties to Bartz v Anthropic, for example, recently reached a USD1.5 billion class action settlement, following Anthropic’s facing potentially massive statutory damages for its downloading of millions of pirated copies of authors’ books. Other AI developers watching this case likely face renewed concerns over liability particularly considering that many of them may have drawn material from the same pirated datasets; this, in turn, may encourage settlement and licensing deals.

Following Bartz, various other high-profile music cases have been resolved in settlements, such as UMG’s suit against music generation platform Udio (1:24-cv-04777) (S.D.N.Y. 24 June 2024), whereby the parties ultimately agreed to collaborate, entering into licence agreements for UMG’s recorded music and publishing catalogues. The parties also announced they will launch a new subscription service in 2026 for generative AI models trained exclusively on authorised and licensed music, with artists given the opportunity to opt-in. See Universal Music Group and Udio Announce Udio’s First Strategic Agreements for New Licensed AI Music Creation Platform, UMG Press Release (29 October 2025). Warner Music Group developed a similar opt-in licensing deal with Udio, to be launched in 2026. See Warner Music Group and Udio Collaborate to Build a New Licensed Music Creation Service, WMG Press Release (19 November 2025).

As the resolution of high-profile AI lawsuits has begun to demonstrate, collaboration between rights-holders, content aggregators and AI developers is possible, and arguably represents the natural resolution to disputes over infringement liability and the ever-thorny fair use doctrine. As a result of the myriad infringement cases, as well as for independent market-driven reasons, content owners and AI developers have begun to establish a new industry-wide standard: licensing for AI training, development and uses.

Current and emerging AI licensing practices

From the wild west of the early AI landscape, licensing practices and deals have come a long way, with many AI companies and key creative industry players joining together in licensing arrangements that benefit both entities. As noted, various notable licensing deals have arisen from the aftermath of litigation, as parties have settled cases and embarked upon collaborative ventures such as AI musical subscription services, which are poised to create revenue streams for artists and songwriters through training AI models exclusively on authorised, licensed music deriving from major music publishers’ catalogues.

Outside of litigation, creative industries – recognising the emerging markets for aggregated training data – and AI developers cognisant of the legal risks in training their models off of pirated data, have also entered into licensing agreements. Image licensing companies have, for example, entered multi-year deals with AI developers for image, video and music model training, allowing access to licensed content. Other image companies have created synthetic content to provide custom datasets to build legally clean marketplaces for the AI industry, such as vAIsual, which in 2023 launched a dataset series comprised of videos with full-body biometrics and postures from models who signed releases for use of their biometric data. A plethora of other companies have also recently partnered with AI developers through licensing deals, including within publishing, news media and film. See – eg, Katie Duffy, A Complete List of Publishers and Their AI Licensing Deals, FutureWeek (4 September 2025). Agreements include multi-year deals whereby aggregated content such as news and editorial articles are supplied, even on a daily basis, to AI developers.

Many existing AI licensing agreements pertain to text-and-data mining (TDM) licensing, allowing AI developers to use rights-holders’ content for training large language models. These agreements routinely include provisions accounting for display rights for AI models to summarise works (in the case of publishing agreements for news articles, for example), as well as imposing rules on data retention and deletion, usage reporting, attribution and indemnification for legal issues such as AI hallucinations. They also often grant the licensor access to the AI company’s tools. The specific terms of many AI licensing agreements, however, are strictly confidential. See AI Licensing for Creative Works, Copyright Alliance(organising ongoing list of major licensing agreements by “Copyright Owners,” AI Companies,” and “Organization”) (last visited 26 January 2026).

Still, various high-profile agreements signal the direction in which other licensing arrangements between key industry players and AI developers may trend. Notably, in December 2025, Disney signed a licensing agreement with OpenAI granting the platform and its Sora AI video app access to over 200 characters. The Sora video tool and ChatGPT Images will, in turn, allow consumers to create short videos starting in early 2026, while Disney will use OpenAI technology on its own platforms, including Disney+, integrating AI into storytelling. Disney concurrently invested USD1 billion into OpenAI, representing one of the first Hollywood studios to strike such an arrangement in support of an AI platform.

In the music publishing arena, Warner Music Group, Universal Music Group, and Sony Music Entertainment entered into separate licensing agreements with AI music start-up Klay to train its models, in addition to the UMG and WMG arrangements with Udio. Other highlighted agreements include Microsoft and HarperCollins’ deal for licensing book content for training, Shutterstock’s deals with various AI platforms to licence image and video datasets, and a deal between Reddit and Google for access to Reddit’s API for AI training. An agreement between the New York Times (NYT) and Amazon, meanwhile, will allow Amazon products, including Alexa speakers, to excerpt from NYT stories or use stories and recipes from the publication, employing this content, additionally, for training. AP and Google’s content licensing deal allows AP’s real-time news information to appear in Google’s Gemini chatbot.

Such undertakings speak to the value in licensing from trusted and curated datasets. AI models trained from datasets without such oversight pose greater risk of AI platform failure. Such systems may often, for example, propagate harmful biases in their training data as to race, gender, socioeconomic status and other sensitive cultural factors, as well as lead to poor model performance, unreliable outputs, and privacy violations including unauthorised data and personal information exposure. Contractually licensing for curated content mitigates such risks, as companies with aggregated original content may serve as trusted sources, correcting issues in current commercially available models. Additionally, licensing ensures fair compensation for rights-holders that avoids their resorting to bringing infringement claims in court, mitigating risk and leaving developers free to focus on their product.

Licensing aggregated content for AI contexts is a natural evolution in content licensing, as companies have been licensing mass quantities of content well prior to the advent and proliferation of generative AI. The Copyright Clearance Center (CCC), for example, follows a voluntary “collective licensing” model to facilitate large-scale licensing for corporate and academic users of copyrighted materials, including books, newspapers, magazines, television shows, images and blogs. Collective licensing refers to a system whereby a Collective Management Organisation (CMO) represents rights-holders to license their works to third parties, via a collective, or blanket licence. This is an effective way to manage the reuse of small portions of published copyrighted works, balancing the needs of rights-holders with those who wish to use their content.

In addition to CMOs, the music, fine art, photography, news media, book publishing and motion picture industries all also have a history of facilitating aggregated licensing deals. In the United States, most aggregate licensing frameworks are voluntary, with these industries independently setting licensing norms and standard practices in relation to economic considerations. Though the government has imposed compulsory licensing systems in rare instances, such as for regulating musical works, these are “exceptional cases”, as noted by former Register of Copyrights Marybeth Peters, reserved for “when the marketplace is incapable of working”; of course, it is “difficult to say that the marketplace is incapable of working... when the marketplace has” not been given enough chance to succeed, so compulsory licences may not be an immediately necessary solution in the case of generative AI platforms. See Regan Smith, Licensing of Text or Generative AI: Learnings from Non-AI Licensing Practices, 48 Colum. J.L. & Arts 450, 454 (2025) (quoting Music Licensing Reform: Hearing Before the Subcomm. on Intellectual Property of the Comm. on the Judiciary, 109th Cong. 13 (2005) (statement of Marybeth Peters, Register of Copyrights, U.S. Copyright Office)). Indeed, compulsory licences may also risk depressing technological and market development by imposing mandatory, less flexible licensing terms on an industry.

Recognising the need for voluntary collective AI licensing, the CCC currently facilitates various systems, including AI re-use rights within its Annual Copyright Licences (ACL), a content licensing system that offers rights for millions of works to subscribing companies. Pioneered by CCC in mid-2024, the ACL re-use rights marked the first collective licensing solution for internal use of copyrighted materials in AI systems. Noting the need for “[r]esponsible AI”, CCC President and CEO Tracey Armstrong advised that “[i]t is possible to be pro-AI and pro-copyright, and to couple AI with respect for creators”. (CCC Pioneers Collective Licensing Solution for Content Usage in Internal AI Systems, CCC (16 July 2024)). The CCC also maintains an AI Systems Training Licence, a voluntary non-exclusive collective licence to aid organisations wishing to comply with copyright laws to use third-party content to train AI systems. A growing number of other voluntary collective rights organisations and similar services offering content licensing solutions include Protégé and Created by Humans.

From collective licensing to individual licence arrangements between industry leaders and AI developers, it is clear that AI licensing will be a major part of the AI/content ecosystem. As the marketplace rises to meet the moment, AI developers and start-ups have less recourse to claim that licensing content for training is impossible due to prohibitive cost and the large-scale amount of material required. Ultimately, whether driven by litigation or market considerations, the players within this space are increasingly recognising the possibility of fostering AI technology in a way that fairly compensates rights-holders and accommodates their legal rights.

Recommendations for best practices in AI-licensing

Considering the many sectors increasingly affected by AI technology, AI licensing is far from a one-size-fits-all approach; however, various trends and best practices are emerging across industries.

Fundamentally, rights-holders’ interests in maintaining an intentional say in how, and whether, their works are used in AI model training and development drives any consideration of best practices for licensing agreements. To this end, many licensing agreements establish an opt-in approach for rights-holders – rather than the opt-out approach more likely to be preferred by AI developers – whereby rights-holders must affirmatively agree to include their work in a training dataset, rather than proactively opt out of inclusion. It is also important to ensure that licensing revenue, whether lump sum, or otherwise, will be appropriately divided and shared with each group of rights-holders. Ensuring transparency as to the copyrighted content contained within a dataset and crediting contributors thereof is also paramount. Considering that many works, such as musical or visual works, are globally administered through international agreements and licensing networks, consistency among AI licensing provisions, and accounting for international laws is also important to avoid disruption. Finally, rights-holders should ensure that any agreement protects them from the unauthorised use of their name, image, likeness, voice and performance via digital replicas and imposes guardrails against infringing output, where applicable.

Conclusion

While 2025 was clearly another fast-paced year for generative AI development, it was also an encouraging year for content owners with the emergence of strong AI licensing standards, practices and deals between major media companies and AI developers. Throughout 2026, expect the market for AI licensing to continue to expand and strengthen. The slate of ongoing litigation surrounding the use of pirated content for AI training only throws the need for this market into starker relief, with settling cases establishing a precedent for collaboration and licensing between parties formerly on opposite sides of the “v”. Expect any determinations as to fair use to inform the progress of AI licensing; however, considering the increasing number of key industry players and AI developers entering into voluntary licensing agreements outside of or notwithstanding litigation, it certainly appears that AI licences are quickly becoming an entrenched cost of doing business for AI model developers, and that licensing will continue to play a crucial role in the continued advancement of generative AI technologies.