New Tools are Needed for Tracking Adherence to AI Model Behavioral Use Clauses

TL;DR

This work analyzes the rise of licenses with behavioral-use clauses (RAIL) applied to AI artifacts, arguing that tracking both adoption and adherence is the next necessary step for effective governance. It presents two empirical studies: a case study of an open-source RAIL license generator that produced over 300 customized licenses in about a year, and a large-scale analysis of the HuggingFace model hub showing substantial but incomplete adoption of RAIL licenses (roughly 12% of models). The findings reveal convergence in clause configurations across licenses and frequent overlap in language, while highlighting gaps in enforcement and the need for provenance-based tooling and standardized formats. The authors discuss societal and regulatory implications, suggesting that improved tracking tools, model fingerprinting, and watermarking could support adherence, with broader impact on policy design and responsible AI deployment.

Abstract

Foundation models have had a transformative impact on AI. A combination of large investments in research and development, growing sources of digital data for training, and architectures that scale with data and compute has led to models with powerful capabilities. Releasing assets is fundamental to scientific advancement and commercial enterprise. However, concerns over negligent or malicious uses of AI have led to the design of mechanisms to limit the risks of the technology. The result has been a proliferation of licenses with behavioral-use clauses and acceptable-use-policies that are increasingly being adopted by commonly used families of models (Llama, Gemma, Deepseek) and a myriad of smaller projects. We created and deployed a custom AI licenses generator to facilitate license creation and have quantitatively and qualitatively analyzed over 300 customized licenses created with this tool. Alongside this we analyzed 1.7 million models licenses on the HuggingFace model hub. Our results show increasing adoption of these licenses, interest in tools that support their creation and a convergence on common clause configurations. In this paper we take the position that tools for tracking adoption of, and adherence to, these licenses is the natural next step and urgently needed in order to ensure they have the desired impact of ensuring responsible use.

Figures (4)

• Figure 1: License Generator. The RAIL License Generator enables a user to select a type of license and the asset they want to license and then customize the behavioral use clauses. The final license text can be exported in several ways along with a quick response (QR) code.
• Figure 2: License Generator Usage. [A] Number of Licenses Created. The number of RAIL, OpenRAIL and ResearchRAIL licenses created using our license generator has been accelerating over the year from April 2024. [B] Clause Adoption by Asset Type. Clauses select for every license generated for each type of asset. [C] Connectivity Plot of Non-Mandatory Clauses. The a circular plot highlights the heterogeneity across licenses in terms of the non-mandatory clauses selected. The opacity of each line reflects the number of licenses that include both clauses.
• Figure 3: RAIL License Release Timeline. Notable milestones in the adoption and standardization of RAIL and Open Source AI licenses.
• Figure 4: License Adoption. [A] Number of Models Licensed. The number of RAIL, OS and other licensed models on the HuggingFace Model Hub from 2023 to 2025. All other models did not have a license. [B] Bi-gram Overlap in Behavioral-use clauses. The percentage of two-grams that overlap between clauses. There is significant overlap in the text content of certain licenses and most licenses have at least 10% overlap with at least one other license.