As we continue to see AI steadily and increasingly be incorporated into service offerings, businesses should pay special attention to previously “standard” provisions when contracting for the provision and use of services that incorporate AI. This is especially true considering there may be situations where service providers use AI at some point in the workstream without the recipient even realizing.
Similar to how the pandemic pressure-tested certain contractual provisions (such as force majeure provisions), the unique issues surrounding AI are also testing how we draft and negotiate other contractual provisions. In this blog post, we explore some particular contractual issues and quirks in IP ownership provisions and data usage rights provisions in connection with the incorporation of AI.
Intellectual Property Ownership
Notably, it is important to know that the US Copyright Office has been applying the requirement for “sufficient human authorship” to refuse registrations for images and texts that are produced by generative AI. As you may also know, in the United States a copyright registration is generally a prerequisite to bringing an infringement action.
Taking the above into account, our first “gotcha” when negotiating agreements is not ensuring that the IP assignment language is appropriate for a given transaction. While we are generally trained (in the United States) to look for present tense assignment language to ensure the effectiveness of IP transfers (i.e., “Supplier hereby assigns” instead of “Supplier shall assign”), the advent of generative AI and the US Copyright Office’s aforementioned positions present an additional issue for consideration that may not have previously been considered as vital to IP assignments: what is the subject matter of the assignment?
A provision that “Supplier hereby assigns all of ITS ownership interests in and to work product” may be interpreted very differently than “Supplier hereby assigns all of THE ownership interests in and to work product.” One word can certainly change the meaning—and potential effectiveness of the assignment. To further clarify, we would consider also adding a sentence that provides that the applicable work product was generated in a manner that allows the transfer of such rights.
Data Usage Rights
Following a similar theme as above, in many negotiation scenarios parties may accept language that allows usage of certain data sets on an “aggregated, anonymous, de-identified basis solely for purposes of improving the products.” The original intent of such language was to use data sets, such as usage data, to improve the user interface of a product and assist in feature development and similar improvements that will increase the overall quality of the service across a user base.
However, consider whether your organization would feel differently about accepting the same language if “improving the products” meant training the models or algorithms that inform AI decision-making or substantially contribute to the probability values that generate AI outputs.
Unfortunately, the fix to this gotcha is not quite as simple as changing a word or two. If the goal is to restrict a data set—even usage data— from training AI products, organizations should simply say so, by adding language such as “provided, however, that no such data may be used to train, modify or improve any artificial intelligence algorithms or models without data owner’s prior written consent.”
The goal of this blog (and the parts still to come in this series) is not to have a chilling effect on the use of AI in business, but rather to consider some areas where this exciting new technology has caused us to increase our precision and thoughtfulness in reviewing the terms and conditions around such use. Morgan Lewis lawyers stand ready to assist companies navigating this evolving landscape.