A Taxonomy of Legal Risks for Generative AI
Intersecting AI #6: Here's a summary of a paper we've written to identify the various legal claims people may making about Generative AI
If you’ve heard about legal issues around generative artificial intelligence (GenAI) like ChatGPT, Stable Diffusion, Gemini, DALL-E, and others, you’ve most likely heard about copyright and privacy concerns. But these are just two of the many claims that have been filed against Google, OpenAI, Microsoft, Meta, and others. If you are at a GenAI company, and especially if you’re at a startup, you may not know what other legal risks may apply, when they may apply, or, perhaps most importantly, how to weigh those risks.
To tackle this issue of unclear information and to help structure how one might think about legal risks, we’ve created the first legal risk taxonomy for GenAI. Taxonomies are a way to classify and organize information. In GenAI, there are taxonomies for ethical, sociotechnical, code generating, privacy, and security risks. There are also specific taxonomies of risks for text-to-image, audio generation, and other modality-specific use cases.
Something Was Missing…
Importantly–glaringly, if you happen to be a lawyer–there is no taxonomy or comprehensive organization for how to think about the legal risks associated with GenAI. It can sometimes seem as if lawsuits and claims pop in and out of the ether based on who’s talking. That’s where our paper comes in.
For the first time, our paper presents a taxonomy of legal risks currently associated with GenAI by breaking down complex legal concepts to provide a common understanding of potential legal challenges when developing and deploying GenAI models.
What We Did
The methodology for the paper was based on:
Examining the legal claims that have been filed in existing lawsuits and
Evaluating the reasonably foreseeable legal claims that might be filed in future lawsuits
For starters, we identified 29 lawsuits against prominent GenAI entities, such as Google, Meta, OpenAI, and Microsoft. We then tallied the claims in each lawsuit and identified the claims made in at least four lawsuits and considered those to be the most likely claims for future GenAI lawsuits. For those most common claims, we described the elements of the claim and provided an example of how it may apply to GenAI.
What Else?
Next, we identified 30 additional claims, which we classified as “potential” claims. We consider this class of potential claims to be more speculative because they have not been included in many, or sometimes any, lawsuits to date.
We also sub-categorized these potential claims into the 19 that are most likely to be made in relation to the pre-deployment of GenAI models and 11 that are more likely to be made in connection with the post-deployment of the GenAI models.
For this potential class of claims, we describe the requirements of the claims and, to help GenAI entities determine their legal risk, we also provide the potential penalties plaintiffs and prosecutors may request (money, injunctions, imprisonment, etc.).
The paper also includes a few helpful additions to assist a broader readership than just lawyers and legal scholars. Our goals are to make GenAI legal issues accessible by explaining the issues with plain language, to promote transparency, and to encourage a more open exchange of ideas around AI and the law. We purposely avoid high-minded and complex legal discussions.
We also include a legal glossary for key terms, identify several practical ways to potentially mitigate legal risks, and we note several open questions in the GenAI legal field to help delineate the known risks from the unknown.
What Now?
The paper proposes some ways the taxonomy can drive further research, but we mostly intend for it to promote AI literacy by providing a useful practical guide to a wide range of AI researchers, business leaders, lawyers, and legal scholars everywhere.
Some of our next steps include building out a more thorough analysis of the claims, more detailed mitigation strategies, and identifying gaps between the law as it exists today and how GenAI may require fundamental changes to how it’s applied tomorrow. We invite any comments, questions, or feedback you may have!