An Overview of Key IP Issues in AI

Given the recent explosion of interest in Generative AI that came along with the introduction of ChatGPT in November and the fact that Proskauer is advising clients on AI issues across almost all practice areas, the firm put together an integrated cross-disciplinary series of webinars designed to answer many of the questions presented by this emerging technology.

Webinar: A Series Focused on Legal Issues Associated with Emerging Generative Artificial Intelligence

The first webinar in the series, “An overview of key IP issues in AI,” tackled several emerging IP issues related to generative AI (“GAI”), the latest example of new technologies bumping up against existing IP laws and challenging the established concepts of IP registration, infringement, and fair use.  The panel ran through multiple IP-related issues that are being raised by stakeholders. Content owners are wondering if GAI developers have already infringed their works during the AI training process and, in some cases, also when allegedly derivative works are produced in the GAI output; at the same time, content creators are wondering how GAI might help them create new work or content or provide a new source of licensing revenue. In addition, businesses across all industries are already using GAI to help in various functions but have concerns about whether such use could expose them to potential risks or liabilities.  As stressed by the panel, unless new regulations are enacted, these IP questions surrounding GAI will necessarily be decided by courts and test the flexibility of our existing IP laws.

The panel broadly outlined the state of law in this area and also broke down the issues from a practical perspective, noting the uncertain risks for entities developing and using generative AI.  Topics included:

• AI terminology: The panel provided a brief overview of AI terminology and a reminder of the differences between “software” and “AI” and “AI” and “GAI,” as those terms are sometimes incorrectly intertwined. On a related front, the panelists discussed the FTC’s recent statements surrounding “AI hype” and the potential for agency enforcement over false or unsubstantiated claims about a product’s efficacy or AI capabilities.

• Copyright infringement (developer liability and “input”): The big question is whether developers of GAI applications infringed the IP rights of content owners based on the web content that was ingested and used to train the AI large language models and whether such developers can be contributorily liable for user prompts that specifically reference copyrighted works to produce new content. These issues are the subject of several ongoing lawsuits involving visual art GAI and coding assistants GAI, with the claimants asserting copyright claims based on the input process (training data ingestion) and the output process of generating content in response to user prompts.  As to the possible outcome of the copyright-related suits against GAI developers for the input of web-based training materials, the panel stated: “It’s really going to depend on how each of these different GAI technologies works…which will be integral to the infringement analysis,” noting that discovery will have to be made clear, for instance, if a GAI technology may have just scanned, read and analyzed text and stored the relationship between such texts instead of engaging into traditional copying and storing.  Secondary liability, the panel suggested, might turn on an interpretation of the Second Circuit’s Google Books case, which involved a finding of fair use of a search engine for millions of copyright-protected books that provided snippets for users to read, and the landmark 1984 Supreme Court Sony-Betamax decision that found a VCR maker could not be liable for providing a tool that had a substantially non-infringing use and allowed viewers to engage in fair-use protected time-shifting.

• Copyright infringement (user liability): The panel also delved into the potential circumstances where a user or enterprise using GAI models might be held to have committed direct or secondary copyright infringement. The panel was skeptical of secondary liability of users for the input or training of the AI model: “A secondary liability case against users for the training of the platform is probably a difficult case.” Though, the panel suggested that there are possible scenarios where a user might be liable for direct infringement, such as an individual asking ChatGPT to write a script based on a popular novel, which could be deemed an unauthorized derivative work.

• Patent issues: Amid the rapid proliferation of AI-related patent applications, the panel noted that “the rise of GAI raises novel issues of patent infringement.” One interesting wrinkle was the issue of who should be liable for patent infringement when originally there is no infringement when the AI model is created, but the infringement occurs when the AI evolves and learns. The panel stated that AI developers perhaps might be deemed the operators of the products and would likely be targeted in any patent litigation, but since there are likely to be multiple entities involved in the AI lifecycle chain, a theory of “divided” patent infringement or joint infringement could be pleaded in certain cases. Lastly, the panel discussed how even though current case law states that an AI cannot be an “inventor” under the Patent Act, there is an open question as to the patentability of inventions made by humans using AI, an issue that the USPTO is currently studying with public input.

• Trademarks: Unlike the sometimes-tricky registration issues relating to works or inventions that are created with both human and AI involvement, trademarks, on the other hand, need not be created by a human author to be protectable. A mark can include any word, name, symbol, device, or combination and registration would depend on the typical legal requirements (e.g., “use in commerce” or “intent to use”, not generic and being capable of distinguishing the applicant's goods from those of others). Thus, not surprisingly, generative AI has become a useful resource for businesses that want to conjure possible ideas for marks that are available.

• Confidentiality and trade secrets: While the law recognizes that at times a trade secret owner may need to share the secret with third parties (typically subject to an NDA), the panel discussed the question of whether inputting confidential information into a GAI could weaken the position that information is a trade secret (especially if the user did not opt-out of a GAI using input data for training purposes). We will likely hear future litigants make this argument when challenging a user of GAI asserting trade secret protect