USA Trends and Developments Contributed by: Jeffrey Harvey, Randall Parks, Andrew Geyer and Cecilia Oh, Hunton Andrews Kurth LLP
states as well as Puerto Rico, the Virgin Islands and the District of Columbia. Out of these jurisdictions, 38 states (up from 18 one year ago) and Puerto Rico have adopted resolutions or enacted legislations. Similar to its adoption and implementation of a sophis - ticated legislative framework relating to privacy and personal information, the EU is also currently ahead of the United States in terms of adopting a legisla - tive framework at the federal or national level. The Artificial Intelligence Act, which was signed into law by the European Union (EU) on 13 June 2024 and published on 12 July 2024, establishes a national framework geared towards regulating the ethical use and implementation of AI. The Artificial Intelligence Act also created the European Artificial Intelligence Board, which is charged with the promotion of co- operation across the EU on matters related to AI and designed to promote compliance with the Act itself. At the federal level in the United States, there has been little action on AI beyond a recent Executive Order that generally promotes the adoption of various standards for AI safety and security. Intellectual property, traditional AI and generative AI Of primary importance when determining the legal risk associated with nearly all forms of AI is: who owns the intellectual property in the AI learning and its out - puts? The answer to this question differs depending on the type of AI solution deployed. Traditional AI sys - tems process data based on a predetermined set of rules and logic, and generally perform a specific task to increase efficiency through repetition. Generative AI and agentic AI process data against a base data set, and develop creative or new content as a result. While, strictly speaking, agentic AI is not generative AI, there is a good deal of overlap. Accordingly, they will be viewed in the same manner for purposes of this section. Buyers of traditional AI systems must disclose their trade secret processes and historical data to establish the predetermined set of rules and logic noted above. While this raises conventional issues of confidential - ity and ownership of the disclosed IP, the customer must also consider who owns the insights or outputs generated by the AI in processing the customer’s data
and how the vendor is permitted to use and profit from the AI that the customer has helped to train (this becomes even more tricky in the agentic AI context). The nightmare for the category-leading customer is that the provider takes the AI-generated insights or outputs and the newly trained AI, and turns them into a category-killing product in which the customer has no financial participation. Savvy providers recognise this concern and are willing to address it effectively. Buyers of generative AI solutions are less concerned with the development by the provider of a category- killing product than they are the source and creation of the output itself. Generative AI solutions generally “scrape” publicly available sources of data in order to deliver new output that is responsive to various queries from end users. The data resulting from the query is typically based on any number of other data sources, the origin of which is unknown. For example, a generative AI solution may be trained by using sev - eral of a famous artist’s greatest works. If an end user then requests that the solution create a brand new image, as if this author painted it, the generative AI solution will fulfil the request. The famous artist neither trained the AI solution nor painted the new image, but the generative AI solution used this author’s style of painting and previous works, in combination with oth - er data, to develop the new image. Is the new image a derivative work of the author’s images used to train the generative AI solution? Is “training” a generative AI model a “fair use” or a permissive use? Consider the impact on this author’s career (and their incentive to produce creative works) if users can obtain works of any image that appears as if the artist painted them. Similarly, buyers of generative AI solutions must understand the risks associated with treating output as if it is owned by the buyer. If 1,000 separate buyers each asks their own instance of the solution to per - form the same task, then the output may be exactly the same or substantially similar for each of the 1,000 buyers. Can any one of the buyers legitimately claim ownership? Providers of the generative AI solutions generally make it clear that all risk associated with the use of the output, including any risk of infringement, is borne by the end user.
107 CHAMBERS.COM
Powered by FlippingBook