Using BigCode Open & Responsible AI License

Responsible AI licenses are mechanisms part of -and interacting with- a broader system of AI governance instruments and processes, such as Model Cards, Ethical Charters and upcoming AI regulations. Open Responsible AI Licenses (OpenRAIL) are AI-specific licenses designed to permit free access, re-use, and downstream distribution of AI artifacts while setting specific use restrictions at the same time. See more information here. 

Compared to its predecessor (BigScience OpenRAIL-M), the BigCode OpenRAIL-M license agreement has been designed to better address: (i) concerns from commercial companies when using a model under a RAIL agreement; (ii) fostering documentation along the AI value chain. 

Companies might want to use the model and commercialize it in some of their products and services, but the set of use restrictions might act as a deterrent, as their legal teams might have a hard time interpreting the restrictions and trying to match their wording with already existing restrictions they have in other contracts. Moreover, the license might force companies to alter some of their existing contracts due to the obligation to include the same set of use restrictions when distributing the ML model or a modified version of it. BigCode has adapted the license to give flexibility to commercial companies to respect the use restrictions while presenting them in a different format better suited to their internal contractual wording and tools.

The BigCode project enabled the Model License Taskforce to set a licensing sandbox and test some of the clauses of the OpenRAIL against some commercial contracts. This served to identify pain points from commercial entities when facing the adoption of an ML model licensed under an OpenRAIL. Moreover, licensing discussions focused on how the use restrictions in the license should be passed across the value chain. 

On a different note, documentation practices in AI are increasingly becoming a core governance factor capable of bringing transparency and information about the core characteristics of the AI products released. Model cards are the perfect example of documentation tools in the AI space. These documentation tools are meant to inform the user about the basic technical characteristics of an ML model, such as its intended use and technical limitations (somewhat like a technical specification).

Consequently, inspired by some open source licenses which require source code users to include a copy of the license when they distribute the source code, the BigCode license agreement does the same with model cards. It requires users of the LLM when distributing it to share the same “explanatory documentation” (e.g. model cards, data cards, etc), or, if distributing a modified version, to share a model card of similar or better quality. 

In this way, by compelling users of the model to share essential documentation when distributing the model, the license promotes better documentation standards and understanding of the technology along the AI value chain. This might prove especially helpful for upcoming AI regulations such as the EU AI Act, where documentation requirements will be required for LLMs.

From a regulatory perspective, the BigCode OpenRAIL-M license agreement was created with upcoming AI regulations in mind, such as the EU AI Act. The AI Act could end up regulating the distribution of general-purpose AI systems (GPAIS) by requiring stakeholders to forbid the use of the GPAIS for the identified high-risk scenarios (as it currently stands under Article 4(c) of the European Council’s General Approach to the AI Act as of December 2022). Henceforth, designing a license with similar restrictions in mind to those stipulated in the AI Act might serve as a trigger to start articulating both public and private AI governance endeavors. 

There is a need for progressive and organic standardization, as has happened with Open Source and Creative Commons licensing phenomena, in order to bring more clarity to users. 

The core challenge of this type of license is to strike a balance between enabling open access and use of the ML model and placing a set of use restrictions at the same time. It can become a challenge for the licensor to control how the ML model is used outside of its technical infrastructure or platform. Consequently, other complementary governance mechanisms are being assessed to see how to optimize the enforcement of this new type of license.

Responsible AI Licenses are a new AI governance tool part of a broader AI governance system. The use of AI-specific licenses is becoming an urgent need to govern economic interactions in the AI industry, especially in light of new regulatory trends, such as the EU AI Act. The core learning for the use of RAILs, and in this case for BigCode, has been the need to adapt the license for all stakeholders to be able to use it, including commercial entities for commercial purposes. It is by taking an open and collaborative approach to AI licenses that existing RAILs will be better understood and improved in the mid-run. Further practical information on RAILs can be found at the RAIL Initiative.