FAQ

Disclaimer: The RAIL Initiative is not a law firm and does not provide legal advice. The presentation and proposition by RAIL Initiative of licenses and/or the FAQs does not create a lawyer-client or other relationship. RAIL Initiative makes the licenses and related information available on an “as-is” basis. RAIL Initiative gives no warranties regarding the licenses, any material licensed under their terms and conditions, or any related information. The RAIL Initiative disclaims all liability for damages resulting from their use to the fullest extent possible.

 
 

What is RAIL and OpenRAIL?

Responsible AI Licenses (RAIL) are a class of licenses designed to encourage the responsible use of an AI artifact being licensed by including a set of use restrictions applied to AI artifact. RAILs can be more or less restrictive depending on the aims of the licensor. For instance, a license can be RAIL while being a proprietary license, or a license just allowing the use of the AI feature for research purposes and without allowing distribution of derivative versions. 

In contrast, Open & Responsible AI Licenses (OpenRAIL) are a subclass of RAIL licenses that permit free-of-charge open access and re-use of AI artifacts for commercial purposes, while including usage restrictions. Note that usage restrictions in RAIL Licenses also apply to any derivatives of AI artifact. 

RAILs can be used to license data (D), Apps (A), models (M), and source code (S). depending on the AI feature(s) you are licensing, you will add suffix D, A, M, or S. If you want to know more about the difference and naming conventions, please take a look at this article.

What is the motivation behind OpenRAIL?

OpenRAILs are inspired by previous open licensing movements that have transformed the world and the perception on how society at large, and more precisely scientific communities, interact around a core essential value: the sharing of knowledge. The main sources of inspiration have been Open Source, Creative Commons, but there are many initiatives which stand for similar values and complementary goals. We do not want to reinvent the wheel, and are taking an evidence-based approach towards adapting current open licensing practices to the use of ML artifacts.

How does that work?

This is enabled by striking a balance between broad and permissive contractual and intellectual property rights grants (similar to Apache 2.0 license) on the one hand, and by including a specific set of restrictions governing critical use-case scenarios for the artifact (eg: a model)  being licensed, on the other hand. At the same time, there is also space to let licensors specify their own sets of usage-based restrictions; which respects that particular AI artifacts come with specific capacities which may need to be accounted for.

What’s the main point of using them?

We want RAILs to be for AI artifacts, such as datasets or models, what open software licenses are to code.

For instance, we feel there is a need to empower model developers to promote the responsible downstream use of the artifacts they release, by giving them the opportunity to control the potential of inappropriate use of the AI artifact. This is enabled by including usage restrictions clauses in a license. 


Which are the available RAIL licenses? What are the differences between them?

  • RAIL-A: license here

Who governs the license? RAIL Initiative.

  • SIL AI RAIL-M: license here

This RAIL-M License was created by SIL International, to facilitate the public release of NLP models representing indigenous languages while safeguarding against downstream usage that might harm those in indigenous communities.

Who governs the license? SIL International.

  • OpenRAIL-S: license here

This license is designed for AI-specific source code. Whereas the license has a permissive character providing the user with a permissive copyright grant, the user also has to comply with the set of use restrictions when using the code and developing derivative works based on the code.

Who governs the license: RAIL Initiative.

  • BigScience BLOOM RAIL 1.0: license info here 

This license is the first of its kind, created under the BigScience Workshop. It was the first OpenRAIL-M license and, broadly, the first open license specifically devoted to the licensing of an ML model, in this case, a 176B parameter large language model (BLOOM).

Who governs the license? BigScience (Decisions taken under Legal & Ethical WG; Model Governance WG)

  • BigScience OpenRAIL-M: license info here

This license is an updated version of the BigScience BLOOM RAIL 1.0, as the latter was  designed for BLOOM models. It was then updated in order to be generally used by any stakeholders when licensing their ML models, and therefore not just applicable to BLOOM. Anyone license models with this license.

Who governs the license? BigScience (Decisions taken under Legal & Ethical WG; Model Governance WG)

  • CreativeML OpenRAIL-M: license here

This license is an updated version of the BigScience OpenRAIL-M license. It was designed with text-to-image models in mind, such as Stable Diffusion. There are 2 main modifications: 1. The Preamble of the license; 2. The deletion of 2 use restrictions that are present in the BigScience OpenRAIL-M license, restrictions (e) and (g).

Who governs the license? Stability.ai and CompVis

  • OpenRAIL++-M: license here

This license was proposed by Yanick Kilchner as an updated version of the CreativeML OpenRAIL-M.  The main modification is the deletion of part of clause 7 of the CreativeML OpenRAIL-M. Clause 7 of the latter requires users of the model to undertake reasonable efforts to use the latest version of the model. The clause was initially thought within BigScience for potential critical model failure scenarios that could lead to unforeseen consequences and harm. In these exceptional circumstances, the user would have to undertake reasonable efforts to use the latest version (i.e. a new version tackling prior failure). However, some in the AI community pointed out that the clause was a stumbling block for users, as it could be interpreted as requiring users to undertake the costs of always using the last updated version of the model each time there was one, even if of lesser quality. 

This version is a blank template and therefore does not include a specific preamble or specific use restrictions. The user has to fill these sections at their discretion. 

Who governs the license? Yanic Kirchner.

  • CodeML OpenRAIL-M 0.1: license here

This license is being drafted under BigCode in order to license code generation models. The license will be an updated version of the BigScience OpenRAIL-M and will also serve as an interim preliminary version of the upcoming CodeML OpenRAIL-M IN 2023.

The main modifications compared to the BigScience OpenRAIL-M are likely to be, as with OpenRAIL++-M, getting rid of the model update requirement and introducing specific restrictions to code generation models.

Who governs the license? BigCode

Are OpenRAILs considered open source licenses according to the Open Source Definition? NO.

​​THESE ARE NOT OPEN SOURCE LICENSES, based on the definition used by  Open Source Initiative, because it has some restrictions on the use of the licensed AI artifact. 

That said, we consider OpenRAIL licenses to be “open”. OpenRAIL enables reuse, distribution, commercialization, and adaptation as long as the artifact is not being applied for use-cases that have been restricted.

Our main aim is not to evangelize what is open and what is not but rather to focus on the intersection between open and responsible licensing.

And what about Creative Commons licenses, is OpenRAIL any different from these?

RAILs ARE NOT CC licenses. We are aware CC licenses are used for model and data sharing, however, similar to open source licenses, current CC licenses do not provide use restriction clauses which we argue are essential to promote the responsible use of AI artifacts. 

What about subsequent versions of the licensed AI artifact, do use-based restrictions apply to them? 

Yes! RAIL licenses are designed to be applicable for downstream licensing terms of any derivative versions of the AI artifact offered and/or released by a downstream user. In other words, the analogy could be made to the so-called "copyleft" or “share alike” clauses from open source and creative commons licenses, meaning that the use of any derivatives of the AI feature (as defined in the license) should be governed -at minimum- by the same use restrictions.

 Can you give me an example? 

Imagine a company wants to use your model in order to develop a version for a commercial chatbot. The company accesses the model, modifies it, and finetunes it to be the technical backbone of the chatbot app. Firstly, these actions will be governed by the OpenRAIL license. Secondly and worth to note, according to the terms defined in the OpenRAIL License this is considered a derivative of the Model. Thus, the use of the chatbot will be governed by the use-based restrictions defined in the OpenRAIL license, and accordingly, when commercializing the new version of the Model by means of a commercial license (or any other type of legal agreement), the latter will have to integrate these use-based restrictions as part of the subsequent license. 

“I think the model I want to release is likely not to cause harm in any event, I’m considering using an open source or creative commons license…”

You are the licensor, and therefore, you choose the license that will best fit your goals and values! RAILs should not be seen as the only way to go when licensing AI artifacts. If you think an open source license can be better for your case, please use one! In any case, please be aware that existing open source licenses do not cover the licensing of AI artifacts such as models or data, just source, and binary code, so sometimes it might be a bit confusing.

“I licensed my ML model under an open source license some time ago, and now I’m considering changing to an OpenRAIL. What considerations should I take into account?”

It will depend on the license you initially placed. For instance, if you were licensing your model under a permissive open source license, you should be aware that downstream users from now on when using a model with an OpenRAIL will not be able to re-distribute the model or derivatives of it under an open source license, as they will have to place the use restrictions of the initial license and therefore their license will be automatically not compatible with the Open Source Definition (principle 6 - no field of use restriction).

In case your AI artifact is widely used, the change from a permissive open source license to an OpenRAIL might deter some users from keeping on using the AI artifact (or not using the last version, or fork a former version in case the AI artifact is AI source code/library).

What about enforcement of A license? 

As a licensor, the license is your statement to potential users on how you would like them to use your model and under which conditions

Thus, in case you, as licensor, find a violation of the terms of the license, you are in a position to reach out to the licensee and try to communicate with them about the inappropriate use, in order to try to reach a common vision on the case at hand. But no one's perfect, so sometimes you won´t be able to reach an agreement. It happens and you should be ready for this. In this, scenario, as a licensor you can enforce the rights infringed in your license. But, how?

You might think about going to court as the only way to enforce the alleged misuse of the model. Contract and/or IP enforcement before Courts is definitely an option, but not the only one! If you are sharing your model in an open repository platform, and you are aware of a violation of the terms of use of the model, you can always reach out to the platform provider to assess the case and if needed take it down. Why? Providers of ML hosting services cannot afford to have unlawful material uploaded in their platform, and in some instances, unlawful means AI artifacts that have been developed by infringing IP or breaching a contractual agreement.

Could you illustrate with some examples? Yes! Here we go:


Case 1️⃣: Imagine you uploaded your model on a platform and you chose an OpenRAIL license for it. Suddenly, one day, you realize there is a model based on yours (so a derivative of your model) on the platform that is being openly released without integrating in its license the use restrictions of the OpenRAIL license.  At this point, perhaps you can first reach out to the user in violation of the terms-of-use of the model, maybe they committed a mistake! Maybe there was a simple misunderstanding! Let’s not assume the worst at first sight. If the user was definitely willing to release the model in this way for restricted uses, then you could reach out to the platform provider in order to assess whether to take down the model.


Case 2️⃣: Same case as above, however, this time you as a licensor are not the one spotting a misuse of the model, but rather a 3rd party aware of your OpenRAIL’ed model who wishes to report a misuse, as another stakeholder uploaded a derivative version and this 3rd party found it. The latter could always reach out to you to inform you, but they can also report the misused model.
Case 3️⃣: You OpenRAIL’ed your AI artifact and you realized someone is using it for an application not respecting some of the restricted use cases in the OpenRAIL license. You reach out to the licensee in a friendly way in order to discuss, reach some common points and solve the problem in a friendly manner benefiting both parties. In case this is not possible, you can look for legal advice and enforce the license in court. However, this should be seen as the exception, there’s always place for comprehensible dialogue and reaching an agreement! 

What if I want to use the AI artifact for a use case that should no longer be restricted because I have fixed a problem or a limitation of the model? 

 Please contact the author of the licensed AI artifact to review the use case and the changes made by you. The model at hand will need to be relicensed to you separately to permit that use case, if approved by the licensor.

Who decides on the use restrictions? Do restrictions cover every harmful use case? 

No! The list of use restrictions in RAILs does not conceivably represent “everything” one could possibly do with, for example, a model. The choice of use restrictions depends on the licensor, usually based on the awareness of the technical capabilities and limitations of an AI artifact.

💎Recommended Related Materials💎