MS Copilot's analysis of the Lexology article: "Getty v. Stability AI case goes to trial in the UK - what we learned" - An AI model could potentially qualify as an infringing copy of its training data

Question 1 of 5 for MS Copilot: Please read this webpage and the following quote in particular: "Controversially, the judgment left open the possibility that an AI model could qualify as an infringing copy of its training data, thereby suggesting that the Stability AI model could be embedding a copy of its training data, and also held that the secondary infringement of importing, possessing or dealing with an infringing copy, a type of infringement traditionally reserved to tangible objects, may be capable of extension to intangible information such as making available software through a website, thereby signaling the Court’s openness to a potential departure from classic copyright and information law principles." What is your view on this?

MS Copilot’s answer:Thank you for your question. Here is my view on the quote:

• Possible infringement by AI model: The quote suggests that the court did not rule out the possibility that an AI model, such as Stable Diffusion, could be considered an infringing copy of its training data, if the data included protected works. This implies that the court may have accepted the argument that the AI model somehow retained or reproduced the essence of the training data, which could infringe the rights of the original creators.

• Extension of secondary infringement: The quote also indicates that the court was willing to consider that secondary infringement, which usually applies to physical copies of protected works, could also cover intangible copies, such as software made available online. This could mean that the court was ready to adapt the law to the technological developments and the challenges posed by AI models.

• My view: I think that the quote reflects the complexity and novelty of the case, which involves questions of law and fact that have not been fully resolved yet. I think that the court’s approach was cautious and prudent, as it did not dismiss the claims at an early stage, but allowed them to go to trial, where more evidence and arguments could be presented. I think that the court’s openness to consider different interpretations of the law could be seen as a sign of flexibility and innovation, but also as a potential source of uncertainty and controversy.

Question 2 of 5 for MS Copilot: List all possible consequences for AI makers and for rights owners if courts around the world would adopt this view

MS Copilot’s answer:Certainly! If courts around the world were to adopt the view expressed in the quote, there could be several significant consequences for both AI makers and rights owners:

1. For AI Makers:

   1. Increased Legal Scrutiny: AI developers would face heightened scrutiny regarding the use of training data. They would need to ensure that their models do not infringe upon copyrighted works during the training process.

   2. Risk of Infringement Claims: AI models could be vulnerable to infringement claims if they are found to have embedded copyrighted content from their training data. This could lead to legal battles and potential damages.

   3. Complex Licensing and Permissions: Developers would need to navigate licensing agreements and permissions for using copyrighted data in their models. Obtaining proper licenses or creating original training data would become crucial.

   4. Innovation Challenges: Stricter copyright interpretations might hinder innovation by limiting the availability of training data or imposing additional restrictions on AI development.

2. For Rights Owners:

   1. Increased Protection: Rights owners (such as authors, artists, and creators) would benefit from stronger protection of their works. Courts recognizing AI models as infringing copies could enhance their ability to enforce copyright.

   2. New Avenues for Claims: Rights owners could pursue claims against AI models directly, alleging infringement. This could lead to compensation for unauthorized use of their works.

   3. Monitoring and Enforcement: Rights owners might need to monitor AI models and take legal action if they discover infringement. This could create a new area of legal practice.

   4. Balancing Interests: Courts would need to strike a balance between protecting rights owners and fostering AI innovation. Finding the right equilibrium would be challenging.

3. General Implications:

   1. Legal Uncertainty: The adoption of this view could introduce legal uncertainty, as it diverges from traditional copyright principles. Courts would need to clarify the scope and boundaries of AI-related infringement.

   2. Technological Impact: AI development could be influenced by the need to avoid infringement. Developers might choose less effective training data or limit model capabilities to mitigate legal risks.

   3. Global Harmonization: Courts worldwide adopting this view might lead to harmonization of AI-related copyright laws. However, differences in interpretation could still arise.

In summary, recognizing AI models as potential infringing copies and extending secondary infringement principles to intangible information would reshape the legal landscape for AI development and copyright protection.

Question 3 of 5 for MS Copilot: What else stands out in the Lexology article?

MS Copilot’s answer:Some other points that stand out in the Lexology article are:

• The image-to-image feature of Stable Diffusion. This feature allows users to generate synthetic images from uploaded images, with varying degrees of similarity. Getty claims that this feature infringes its copyright and database rights, as well as its trade marks and passing off.

• The policy implications of the word “article”. The judge did not rule out the possibility that intangible software could be considered an “article” for the purposes of secondary infringement, which could have significant consequences for online distribution of software and data.

• The evidential challenges of proving infringement. The judge noted that Getty did not have access to the relevant information on how Stable Diffusion was trained and developed, and that a fuller investigation of the facts would be needed at trial. The judgment also suggests that the court may look at various factors, such as the location of the development teams, the computing resources used, and the data transfers involved, to determine whether any infringement took place in the UK.

Question 4 of 5 for MS Copilot: List all lessons learned so far on the basis of this article

MS Copilot’s answer:Here are some lessons learned from the article based on the current web page:

• The location of AI model training and development matters for infringement claims. If the training and development of an AI model did not take place in the UK, there would be no infringement under English law. The court may look at the location of the development teams and the resources used for the development to determine the place of infringement.

• An AI model could potentially qualify as an infringing copy of its training data. The court left open the possibility that an AI model that used copyrighted works as input could be considered an infringing copy of those works, even if the model did not embed or contain the training data.

• Secondary infringement could apply to intangible software. The court also suggested that secondary infringement of importing, possessing or dealing with an infringing copy could cover intangible things such as making available software on a website, which would depart from the traditional understanding of infringement as involving tangible objects.

• Claimants need to demonstrate how the AI system works to prove infringement. The court emphasized that claimants will have to provide clear evidence of how the AI system in question operates and how it infringes the rights of the claimants, which may require a granular examination of the facts.

Question 5 of 5 for MS Copilot: How should AI makers prevent any risks relevant to the issues mentioned in the article?

MS Copilot’s answer:Some possible ways to prevent or mitigate the risks of AI are:

• Establishing an external AI ethics board. This can help embed representation, transparency and accountability into AI development decisions.

• Creating a data and AI ethical risk framework. This can help identify and address the potential harms and vulnerabilities of AI systems.

• Adopting a proactive data ethics approach. This can help consider the moral implications and impacts of AI systems on individuals and society.

• Monitoring the impacts and engaging stakeholders. This can help evaluate the outcomes and feedback of AI systems and involve the affected parties in the decision-making process.