Navigating Training, Improving And Competition Restrictions In Generative Artificial Intelligence (AI) Agreements

As organizations accelerate their adoption of generative AI, attention is increasingly shifting from the technology itself to the contractual frameworks that govern its use. What may appear to be standard licensing language can have significant implications for how businesses leverage AI tools, incorporate AI-generated content into their operations, and pursue future innovation strategies. At the same time, AI providers are seeking to protect the value of their underlying models, data, and intellectual property in a rapidly evolving competitive landscape. Against this backdrop, market approaches continue to develop, and key questions remain unresolved. This article examines one of the more consequential and evolving aspects of generative AI contracting, highlighting areas where organizations should pay close attention to ensure agreements align with both commercial objectives and long-term technology ambitions.

Software providers have long sought contractual restrictions prohibiting users from using the provider’s software to develop competing products or services (for simplicity, products and services are collectively referred to in this article as “offerings”). Sophisticated software providers and users alike generally understand how those “non-competition” restrictions operate in license agreements governing more traditional software offerings that don’t use AI (or that use only limited, non-generative AI models). However, software platforms leveraging generative AI (that is, AI capable of generating new content like text, code, images, data, or other output) raise unique considerations since users may derive meaningful value not only from the platform itself, but also from the output it generates.

As a result, some AI providers often seek express restrictions in generative AI software license agreements that prohibit users from using the AI platform – and sometimes also the output generated by the platform – to train, improve, or develop competing offerings. For AI providers, these restrictions can help protect the provider’s rights in its proprietary models, training datasets, and other technology underlying the AI platform. For users, however, overly broad restrictions may limit the user’s ability to utilize AI-generated output for its intended business purposes.

Accordingly, businesses negotiating for the use of generative AI software platforms should consider two linked questions: What does it mean to use an AI platform (or the output generated by that platform) to “train,” “improve,” or “develop” competing offerings, and does a restriction on use of the AI platform also restrict use of output generated by that platform? This article explores these questions and related drafting considerations in generative AI software license agreements.

Platform use vs. output use

License agreements for software offerings that do not utilize generative AI typically focus on a single category of user activity: the user’s authorized use of the licensed software. Generative AI software license agreements, however, often involve two distinct categories of activity: (i) the user’s use of the AI platform itself; and (ii) the user’s use of the AI output (that is, the text, code, images, data, or other output generated by the platform).

The distinction between platform use and output use can have practical consequences. Depending on the platform’s functionality and intended use case, AI-generated output may have meaningful independent value to the user. For example, a business may wish to use AI-generated output to create new offerings or train or improve existing offerings. As such, when an agreement prohibits a user from using an AI platform to “train,” “improve,” or “develop” a competing offering, an important question arises: does that restriction apply only to the user’s use of the AI platform itself, or does the restriction also apply to the user’s use of output generated by that platform?

Certain public-facing AI license terms posted online by major AI providers as of the date of this article reflect different approaches to this issue. Some providers generally limit their restrictions to use of the underlying AI platform itself, while others expressly restrict certain downstream uses of AI-generated output. These variations may reflect differences in the providers’ respective technologies, business models, and views regarding the competitive value of the AI-generated output. In any case, they suggest there is no clear consensus among major AI providers as to whether restrictions on use of an AI platform should also apply to use of AI-generated output.

It is also unclear whether a U.S. court would interpret a contractual restriction on use of an AI platform as extending to AI-generated output where the agreement does not expressly reference that output. U.S. courts generally interpret contracts based on their plain meaning and may be reluctant to read material restrictions into an agreement that the parties did not expressly include. Accordingly, where a restriction applies only to use of an AI platform, a user could have a reasonable argument that the plain meaning of the word “platform” refers to the provider’s underlying software and does not extend to new content generated by that software, such as text, code, images, data, or other output, which may reasonably be viewed as distinct from the software itself. However, contract interpretation principles can vary by jurisdiction and may depend on the specific contract language and the factual circumstances of the dispute. Given this uncertainty, AI providers and users should carefully evaluate these provisions in light of the particular platform, output, and intended use case at issue.

What activity is restricted?

Another key variation among the public-facing terms of certain AI providers is how broadly those terms define the offerings that the user is prohibited from training, improving or developing. Some providers focus on prohibiting developments that could compete with the provider’s AI platform (including, in some cases, using the AI platform to train competing AI models). For example, one leading AI platform is governed by public-facing terms that prohibit the user from using output to develop AI models that compete with the AI platform itself.

Other providers define the restricted activity more broadly. For example, the public-facing terms governing some leading AI platforms prohibit the user from using the AI platform (or, in some cases, the AI platform or its output) to develop (or, in some cases, using the AI platform to develop, train, or improve) offerings that are competitive with or similar to the provider’s AI platform.

Some providers take an even broader approach. For example, the public-facing terms of some leading providers of AI platforms prohibit users from using the AI platform’s output to develop machine learning models or “any related AI technology.” Similarly, the public-facing terms governing other leading AI platforms prohibit users from using either the AI platform or its output to “create, train, or improve (directly or indirectly) any AI technology.” In each case, these restrictions apply irrespective of whether the AI technology at issue is competitive with or similar to provider’s offerings.

Practical drafting considerations 

For AI providers, the appropriate scope of any training, improvement or competing-offering restriction will depend on the provider’s AI platform, business model, and the competitive value of its AI platform and the associated output. Accordingly, providers should carefully consider the following issues when drafting these restrictions:

• Whether the restriction should apply solely to use of the AI platform or also to use of output generated by that platform.
• Whether the restriction should apply only to offerings that directly compete with the provider’s AI platform – or more broadly to offerings that directly or indirectly compete with provider’s AI platform, “similar” offerings, or AI technologies generally.
• Whether certain activities (such as testing, evaluation, research, or benchmarking) should also be expressly prohibited.

For AI users, the appropriate scope of any training, improvement or competing-offering restriction will depend on the user’s anticipated use of the AI platform and output, including, specifically, the extent to which the user intends to use the AI platform or output to train, improve or develop any offering. Accordingly, users should carefully consider the following issues when evaluating these restrictions:

• Whether the restriction applies to the AI platform, AI-generated output, or both.
• Whether the restriction is limited to training, improving and/or developing offerings that directly compete with the licensed AI platform – or if the restriction extends more broadly to offerings that directly or indirectly compete with provider’s AI platform, “similar” offerings, or AI technologies generally.
• Whether the agreement expressly authorizes the user’s anticipated business uses of AI-generated output, particularly where the user may wish to incorporate that output into its offerings or to train or improve any of those offerings based on that output.

Bottom line

As businesses increasingly adopt generative AI technologies, both AI providers and users should carefully consider how training, improving and competing-offering restrictions are drafted in generative AI software license agreements. Given the uncertainty surrounding how courts may interpret provisions that restrict use of an AI platform without expressly addressing AI-generated output, generative AI software license agreements should clearly define whether these restrictions apply to solely the AI platform or if the restrictions should extend to AI-generated output. Thoughtful drafting in these areas can help ensure that providers adequately protect their proprietary rights and competitive interests while allowing their users to receive the intended benefit of both the AI platform and the output it generates.