Artificial Intelligence (AI) is an increasingly present force and one that requires attention. This is especially true in the copyright space, where the policies, legislation and judgements that will guide how copyright and AI interact are being formed now. There are three main issues that require focus.  

Does the AI training process infringe copyrights in other works?

The first issue is fundamental to how AI platforms work and how they source their data. Generative AI – programs that create new output – are ‘trained’ on very large data pools. This data corresponds to the output the AI seeks to generate, such as images and text. Patterns and relationships between datapoints are established and used to create rules, from which the generative AI is able to create responses to prompts from a user. Two popular generative AI tools, Stable Diffusion and Midjourney, source the data for their models from the LAION-5B dataset which contains close to 6 billion images obtained from the web. This dataset is known to contain copyrighted creations. Several lawsuits have emerged from this fact, with a group of writers filing a lawsuit in New York last week accusing several technology companies of using their work to train artificial intelligence systems. Likewise, Getty Images, the stock photo provider, has sued Stability AI for misusing over 12 million Getty photos to train its AI Image generation algorithm.

Do AI outputs infringe copyrights in other works?

Generative AI platforms are also being sued by the owners of original works for using their work to produce outputs that are insufficiently transformative, and therefore derivative. Clarification is therefore being sought through the courts as to what the limits of “derivative work” are and how the concept of fair use is to be interpreted within the context of AI. Naturally, the risk of litigation may act to limit the uptake of generative AI tools, and technology companies are responding. Microsoft has already made a copyright commitment for its Copilot services in an effort to mitigate this. Copilot is an AI assistant feature and the commitment extends IP indemnification coverage to this tool, such that if a third party sues a commercial customer for copyright infringement whilst using Copilot or the output it generates, Microsoft will defend the customer and pay the costs of any adverse judgements or settlements. Google, Adobe and others have since made similar commitments, but this is unlikely to be a long-term solution to the question.

Do AI generated outputs enjoy copyright protection, and who owns the copyright to generative outputs?

A third issue is that copyright offices are reluctant to afford protection to AI generated works. The US Copyright Office (USCO) maintains that most AI generated works are not copyrightable and require “creative contribution from a human actor.” This position impacted the recent Hollywood strikes by members of the Screen Actors Guild-American Federation of Television and Radio Artists (Sag-Aftra) and the Writers Guild of America (WGA), in which the use of AI was a prominent (but not exclusive) issue. Members of the WGA were given the greenlight to return to work after a deal was made that included provisions around the use of generative AI. This deal does not ban the use of AI, and the agreement signals that studios foresee writers working with AI tools to generate an output that can then be protected.

Yet there is precedent for courts considering extending authorship to non-human entities; in a previous article we referenced the famous case of Naruto, a macaque, who took a selfie in 2011 which became the focus of a long-running lawsuit. Ultimately, the Ninth Circuit Court of Appeals determined in April 2018 that animals have no legal authority to hold copyright claims. That said, administrative bodies and policymakers are also conscious of the need for clarity on the issue of AI and copyright. The US Copyright Office is currently conducting a study regarding the copyright issues raised by generative AI and is presently collecting views. The office will use this information to “analyze the current state of the law, identify unresolved issues, and evaluate potential areas for congressional action.” The deadline for submissions is the 30^th October 2023. Likewise, the EU’s digital strategy is partly focussed on the regulation of AI and includes a proposed requirement that generative AI comply with certain transparency requirements, such as disclosing the content was generated by AI and publishing summaries of copyrighted data used for training the AI. The stakes are high too: over 7 million jobs in the EU are considered copyright-intensive.

Conclusion

The landscape surrounding AI and copyright is an uncertain one. It is evident that people stand to benefit from the presence of AI tools, but it is also apparent that there are risks – both real and perceived – and it is these as well as the questions raised in this article that are the focus of policy makers, legislators and judges now and in the immediate future. OxFirst has extensive experience in performing economic analysis in the copyright space to support policy formulation and advocacy.