AI state of play for media businesses

The use of AI to generate creative assets can provide significant operational and cost efficiencies for media businesses, but also gives rise to both legal and reputational risks. 

Chief among these is the uncertainty surrounding the subsistence and ownership of IP in AI-generated content. If there are no IP rights to own, license and enforce, the AI-generated content in question may have little commercial value. 

• The prevailing view among IP practitioners in the EU remains that a human author is required for a creative work to be considered original and therefore protected by copyright.
• In December 2023, the Court of Appeal of England and Wales in THJ v Sheridan reaffirmed that the EU's "author's own intellectual creation" test for originality continues to apply in the UK – thereby implying a human author is also required for copyright to subsist in the UK. However, the requirement for a human author is at odds with s9(3) of the UK's Copyright, Designs and Patents Act 1988, which provides for authorship (and therefore ownership) of computer-generated works, defined as being works "generated by computer in circumstances such that there is no human author". The UK position therefore remains a conundrum that requires resolution by lawmakers.
• In Germany, obstacles to protectability start with the author concept of Article 7 of the Copyright Act, which is understood to require a natural person. A growing minority argues that there is no issue as the prompting natural person uses the AI-like a tool to create the output. Most lawyers reject this argument, pointing to the lack of control of the human over the creative process. At the same time, most lawyers agree that human transformation of raw AI-generated output could open the door to protection. Changes to copyright law with regard to AI – which could include a new neighbouring right – are primarily a matter for the EU rather than individual EU Member States. In March 2024, the Federal Ministry of Justice supported the carrying out of an open-ended review in the near future to determine whether AI-generated products require an adapted copyright legal framework under EU law.

In the meantime, media businesses should ensure that, where AI is used in the production of any material creative assets, sufficient human review and revision, including human creative choices, takes place to meet the "author's own intellectual creation" test for originality.  

Media businesses also face a risk that creative assets generated by third party AI that has been trained on third party creative works without a licence could infringe those works.  The High Court of England and Wales will consider in the ongoing Getty v Stability AI case what it means for AI-generated outputs to infringe copyright and other IP rights. This includes what it means to "reproduce" a work in the AI context and the extent to which the defence of fair dealing for purposes of pastiche can apply – and whether the user or the AI provider (or both) should be liable. The case will come to trial in June 2025. 

In the meantime, some AI providers are now offering contractual protections in their enterprise agreements, such as indemnifying users for third party IP claims in connection with generated outputs (more here). Others follow the opposite path, requesting their users to indemnify them should they be drawn into an infringement dispute stemming from output generated with their AI. Some media businesses are taking steps to mitigate this risk by either training AI tools themselves on their own proprietary assets or ensuring that third party AI tools they use principally draw from a specific pool of proprietary assets when generating responses to prompts. 

Personality rights are also likely to gain in importance. AI tools allow voice cloning, opening new licensing markets, in particular for celebrities with distinctive voices, as discussions between Scarlett Johansson and Open AI on the use of her voice for its virtual assistants in May of this year have shown. Where there is no such agreement, courts may soon have to determine the extent to which relevant IP rights exist and are infringed by these AI use cases. In the UK, for example, where no 'personality right' is recognised as such, claimants may need to attempt to rely on the law of passing off to establish infringement in relation to cloned voices (more here).

Regardless of the answer to outstanding legal questions and whether media business have any contractual recourse against an AI provider, there may be wider relationship, ethical or reputational considerations around AI-generated content that relevant stakeholders believe is too similar to pre-existing works. Ensuring a level of human review of AI-generated assets will provide a safeguard against obvious and potentially embarrassing instances of over-similarity, whether or not this amounts to infringement.

Finally, quite apart from any legal ownership or infringement risk, media businesses can also face reputationally damaging backlash from creators and the public at large in connection with the use of AI-generated assets. For example, in June 2024 the film "The Last Screenwriter", written entirely using AI, was withdrawn following public backlash over the use of AI in place of a writer.  Earlier in the year, the BBC ended its trial of using AI to create marketing materials for Doctor Who following user complaints. Media businesses therefore need to remain sensitive to wider perceptions around the use of AI in creative industries among their customers, creators and other key stakeholders.

However, attitudes to AI use can vary and AI is also being used by media businesses to create efficiencies and detect and prevent potential reputational pitfalls. For example, the same technologies are already used to increase the effectiveness of content creation in the press – AI systems can be used to summarise complex data, making research faster and easier. As competition on the online advertising and paid content market increases, AI-generated content can be used to further personalise content to better address specific audience groups media businesses are targeting. AI systems can also be used for detecting mistakes, deepfakes or biases in publications. Keeping in mind the fact that AI may hallucinate, human control is likely to remain key to maintaining the level of trust in the media that is essential for its role in democratic societies.

High quality media assets are an extremely valuable resource for AI training (as compared to the content of the internet at large, for example). The past year has seen an increase in licensing deals in which media rightsholders permit AI developers to use their IP-protected assets to train AI tools. During this time, Axel Springer, News Corp and the Financial Times have all entered into high-profile deals with OpenAI, reportedly for large sums of money.  Image database Shutterstock has entered into a number of licensing deals with OpenAI, Meta, Apple and others and is reported to have generated over $100 million in revenue from AI licensing in the year to June 2024. 

Media businesses are increasingly considering partnerships with AI developers for a number of reasons other than simply earning significant licensing revenue. Importantly, licensing allows media businesses a measure of control in the way that an AI tool may be trained and may react to certain user queries. For example, a licence agreement could have built into it provisions that the rightsholder's content will only form a certain percentage of the material the model is trained on. It may allow the rightsholder to withdraw content and oblige the AI developer to stop using that content if certain circumstances occur, eg litigation against the AI developer. It could also provide obligations for the AI developer to hard code the AI model to return specific responses to user attempts to prompt engineer in order to access the rightsholder's content.

Media businesses are also increasingly deploying AI in their own businesses, for example, to improve user experience or content discovery, to offer tools for user creativity, or to generate creative assets themselves. Partnering with an AI developer can provide favourable terms on which to access the developer's services. Before entering a licensing deal of this nature, the rightsholder will need first to be sure of its own rights position from creators, which can be a complicated assessment where rights are held multi-nationally.

Despite an uptick in licensing activity, disagreements and disputes regarding the use of copyright-protected works for AI training remain rife. Where no specific licensing deal has been agreed in relation to AI training, rightsholders are now increasingly amending their existing and template distribution agreements to expressly exclude certain AI use cases.  Rightsholders are also taking steps to avail themselves of the ability to "opt-out" of the EU's text and data mining exception to copyright and database rights infringement. For example, Sony Music and Warner Music Group have issued letters to the world at large stating their position, and other rightsholders are taking steps to amend website terms and conditions and copyright notices. 

Questions remain around the methods rightsholders can use to effectively opt-out, due to the requirement in the EU Digital Single Market Directive for the opt-out to be "machine-readable".  It is currently untested whether methods such as sending letters can be effective. This question around the form of the opt-out, and specifically whether an opt-out in website terms and conditions can be sufficient, is the subject of litigation in Germany in the LAION case, which came to an oral hearing on 11 July 2024. 

In the UK, there is no general text and data mining exception to copyright infringement, but an exception does apply to text and data mining for non-commercial research and there is also a temporary copying exception (that also exists in the EU). Both are untested in this context. The Getty v Stability AI case may not consider the application of these exceptions.  Stability AI - in its defence to Getty's claims regarding the training and development of the Stable Diffusion model - is largely relying on the fact that training took place outside of the UK and therefore no copyright-relevant acts took place in the UK. This will need to be argued and decided as a matter of fact rather than law. However, Stability AI does argue, in the context of Getty's secondary infringement claim, that, if the Stable Diffusion model as a whole is an infringing copy of Getty's works used for training, then the making of the model would not have infringed copyright in the UK. If these arguments are expanded then the case may yet address questions of whether AI training infringes copyright.

The EU AI Act will potentially make it easier for media rightsholders to identify where their works are being used for AI training and either enter into licences or seek to enforce their rights.  It will impose obligations on general purpose AI (GPAI) models, including generative AI models, to publish a sufficiently detailed summary of the content used for training. This will need to be in the form of a template to be prepared by the EU AI Office.

It is unclear how granular the disclosures will need to be and to what extent licensed sources will need to be disclosed. The Recitals to the Act state that the summary should be generally comprehensive instead of technically detailed, for example, by listing the large private or public databases or data archives used and providing a narrative explanation about other data sources. There is concern among both AI developers and media rightsholders who have entered into confidential AI training deals as to whether the AI Act will require these to be disclosed, though the Recitals do acknowledge the need to protect trade secrets and confidential business information. 

The AI Act will also impose obligations on media rightsholders insofar as they use AI tools in their own businesses. These obligations can arise in several ways. 

• First, the AI Act imposes various transparency requirements on providers and deployers of "limited risk" AI systems and GPAI models, which will apply from 2 August 2025. Media businesses should be aware that, where they have AI tools developed by a third party but provide these to customers under their own brand, they will be considered the "provider" and not merely the "deployer" and will therefore need to comply with additional obligations.
• Secondly, there are a variety of transparency obligations. Providers and/or deployers of “limited risk" and GPAI systems must inform users from the beginning that they are being exposed to AI output or interacting with an AI system. Hence, providers of these AI systems must ensure that, where the system is intended to interact directly with human beings (eg a chatbot), those individuals are informed that they are interacting with AI unless this is already obvious. Providers of “limited risk” and generative AI systems need to ensure that the outputs generated are marked in a machine-readable format and detectable as being artificially generated or manipulated. This is likely to pose technical challenges for certain content types and formats but is already being implemented voluntarily in some sectors of the industry, eg in relation to the "news over audio" service various publishers are offering. Deployers of AI systems that generate or manipulate (a) content constituting a deepfake (as broadly understood, meaning something that appears to be real-life but is not), or (b) text published to inform the public on matters of public interest, need to disclose that the content in question has been artificially generated or manipulated. The obligation in relation to text of public interest is likely to affect news media in particular.

More on the AI Act here.

While there is no centralised regulation of AI in the UK similar to the EU AI Act, UK sectoral regulators and industry participants have published position statements regarding uses of AI in media industries. For example, Ofcom has published a Note to Broadcasters pointing to existing provisions of the Broadcast Code that will protect audiences from misinformation, erosion of trust, and infringements of privacy posed by synthetic, including AI-generated, media. The Advertising Standards Authority has published a variety of guidance including guidance on the use of AI as a marketing term, which applies existing CAP Code principles to 'AI-washing' where advertisers overstate the AI capability of their products or services. The BBC has also published its own guidance on the use of AI within the BBC itself and by anyone commissioned to make content for the BBC. It emphasises the importance of trust, transparency, human oversight, accuracy and impartiality.

Germany’s State Media Authorities are following a similar approach in their joint position paper published in March 2024 in which they request that the German Interstate Media Treaty be amended to better reflect the use of AI in the media. To strengthen the plurality of perspectives, they call for certificates and test scenarios for AI applications being used for the generation and distribution of media content including criteria and provisions to ensure diversity of content. Emphasising their responsibility for AI generated content and processes, the authorities encourage media businesses to specify their internal control mechanisms pre and post publication. To maintain public trust, they request the inclusion of AI-related transparency and disclosure obligations on data origin, training methods and distribution mechanisms in the Interstate Media Treaty and promote mandatory labelling of synthetic voices, AI presenters, AI-generated images, texts and AI programs.

Regulatory challenges 

Notwithstanding the attempts made by lawmakers and regulators, AI regulation on transparency and deepfakes is unlikely to be able to solve all issues that may arise in connection with the use of AI technologies in the media industry. As well as tackling misinformation driven by deepfakes, the law will also have to come up with solutions to tackle the 'liar’s dividend' problem, whereby proliferation of deepfakes allows public figures to deny true reports. Any laws and regulations in this area must also maintain freedom of expression and of the press not only on the political but also the commercial level.

In the absence of effective regulation on this front, it will be interesting to see who will step up to create or enforce standards of responsible AI practice, for example, through industry alignment or private action in the courts. Recent regulatory attempts in the EU – such as the Digital Services Act – are aiming to centralise monitoring and supervision with one authority in each Member State with a view to better streamlining compliance and enforcement, and also provide a right for individuals to bring private damages actions for loss suffered as a result of businesses' non-compliance. However, it is questionable whether previous attempts to outsource enforcement on market participants and/or related associations through standing to bring private actions have been effective.

The impact of AI on the media sector is substantial. Embracing change but doing so in a considered manner is therefore essential for seizing its opportunities and minimising its risks – and all of that at a speed that matches the rapid development of this technology.