Carlos Muñoz Ferrandis - BigScience Legal/Ethical Working Group
Danish Contractor - BigScience Model Governance Working Group
Disclaimer: This post is not intended to be legal advice from any of the authors.
In collaboration with the RAIL initiative, we are excited to release the BigScience OpenRAIL-M license – a license with behavioral use restrictions that can be applied to any AI model being released. We hope the AI community will find this license useful for releasing their AI Models.
Background
After the release of the BLOOM RAIL license one of the most recurrent questions was whether BLOOM was open and re-distribution and distribution of derivative versions enabled (which indeed it is). However, we realized the term “RAIL” did not clearly communicate the concept of free access and re-distribution and that we needed a clear and simple strategy to address this.
Simultaneously, we observed that other recently released ML projects such as OPT-175 and SEER were made available under their own licenses (OPT-175-license and SEER-license) both of which could be considered RAIL licenses, but they only permit the use of the model for research purposes. Consequently, it became clear that as the adoption of RAIL Licenses increased, we needed more clarity in the way RAIL licenses and the artifacts released under are organized. The result was the development of a naming convention by the RAIL Initiative, outlined in this article, which introduces a license class, called Open RAIL. Open RAIL licenses promote free use and re-distribution of the applicable artifact, while maintaining behavioral Use Restrictions.
The BigScience OpenRAIL-M License
Recently, models such as Stable Diffusion have also decided to implement OpenRAIL-M licenses, in this case, the Creative ML OpenRAIL-M, which has been adapted from the BLOOM RAIL License.
To help the AI community reuse the BigScience BLOOM RAIL License more broadly for distributing models, we adapted the terms to make the license applicable to any associated AI Model. Thus, this license is not just for NLP models, but it has been adapted so that it can be applied to other types, including multimodal generative models. We refer to this license as the “BigScience Open RAIL-M”.
Based on the questions we received after the release of the BigScience BLOOM Model, we include some FAQs below for the BigScience OpenRAIL-M License (available here). Additional FAQs pertaining to the usage restrictions are available here.
What is an Open RAIL license?
Open Responsible AI Licenses (Open RAIL) are licenses designed to permit free and open access, re-use, and downstream distribution of derivatives of AI artifacts as long as the behavioral-use restrictions always apply (including to derivative works).
Why Open?
The term “Open” is to indicate that the license enables royalty free access and flexible downstream use and re-distribution of the licensed material, and distribution of any derivatives of it.
To be clearer, you can use any type of license or legal agreement you want to re-distribute the model or distribute derivatives of it under the sole conditions of embedding the RAIL clauses related to the responsible use of the model (i.e. Section 5, Attachment A).
Responsible?
Responsible AI licensing is a mechanism that is part of -and interacting with- a broader system of AI governance instruments and processes, such as Model Cards and Ethical Charters.
Open RAIL licenses are designed to promote responsible downstream use and distribution of the model by including a set of use-based restrictions for which the model cannot be used. In the case of the BigScience OpenRAIL-M License, the restrictions are specific use-cases where the BigScience community believes models could be used as a harmful instrument, due to either the intent of the user or the technical limitations, thus going against BigScience’s values informed by its Ethical Charter.
What does the “M” stand for?
“M” stands for “Model”, which is the artifact being licensed under this license and subject to the use-based restrictions. See the naming conventions for RAIL Licenses here.
What has been modified from the BigScience BLOOM RAIL Licenses to make this one a generally applicable license?
In general terms, the license is almost the same, no critical modifications have been made. We adapted it to be applicable to other models, such as other NLP models or multimodal generative ones, and not just to the BLOOM set of models. We introduced the terms “information and/or content” along the definitions and use-based restrictions in Attachment A to emphasize the cases where the license may even be applied to multimodal generative models.
What if I do not agree with some of the restrictions in Attachment A?
Community feedback is essential for us, the “BigScience Open RAIL-M” license is the result of an ML community initiative and we are grateful to receive feedback.
My use of the model falls under a restriction, but I still think it’s not harmful and could be valuable…
In case your use of the model falls under a restricted case but you think your use is likely to be non-harmful, please contact the licensor of the model you are using/distributing for them to assess the case and see whether an authorization and/or license could be granted for you for this very specific case. We do not want the license to be a future constraint for new uses and novel/more efficient ML techniques, and thus both the license and the licensors using it should be flexible to adapt to constant changes and improvements made by the AI community while being accountable about their respective uses of the ML artifacts.
What about enforcement of the license?
The stakeholder adopting this license for one of their models should be aware that the enforcement of the license’s terms remains their decision and should be carried at their discretion, according to the stakeholder’s values and willingness to promote responsible use of the released AI artifact.
Note that, even in the absence of active enforcement, the use-based restrictions in the BigScience OpenRAIL-M serve as a deterrent as there is always an enforcement potential. However, stakeholders adopting the license should be always aware that openly released AI artifacts are also exposed to potential misuses for which the mere presence of the restrictions placed in the license might not be enough. In these specific instances, the licensor will be the one deciding whether to enforce the license and in which manner depending on how the AI artifact has been made available and how it is being misused. For instance, paragraph 7 of the license indicates that the licensor can at his/her discretion restrict the usage (remotely or otherwise) of the model in violation of the license.
Any further comments or questions? Please reach out to us:
Carlos Muñoz Ferrandis (carlos@huggingface.co)
Danish Contractor (danishcontractor@outlook.com)
License acknowledgments: Jenny Lee
Blog Acknowledgements: Giada Pistilli, Daniel McDuff, Britney Muller