AI2 ImpACT Licenses
The AI2 ImpACT Licenses get their name from AI2’s four core organizational values: Impact, Accountability, Collaboration, and Transparency. The aspiration of the AI2 ImpACT Licenses is to put each of these values into practice through an AI license that serves the common good. The AI2 ImpACT Licenses are designed specifically for AI artifacts by a multidisciplinary group at AI2, including stakeholders from legal, ethics, and research. Unlike other AI licenses, the AI2 ImpACT Licenses set license rights and permissions based on the potential risks of a particular artifact; they also provide tools to increase transparency about AI development and improve flexibility to address potential and unforeseen harms from AI releases.
We plan to use the AI2 ImpACT Licenses for the first time in connection with the AI2 OLMo project. As a next step, we intend to create templates and additional resources so that the AI2 ImpACT Licenses may be used externally by any researcher or AI developer interested in responsible AI.
Access the full text of each of the AI2 ImpACT licenses:
What do the AI2 ImpACT Licenses cover?
The AI2 ImpACT Licenses were developed by AI2 to license AI artifacts for use with machine learning and developing AI technology. The first version of the AI2 ImpACT Licenses specifically cover data (raw data curated into a dataset, such as data for training, evaluation, or instruction tuning) and models (model weights and parameters that comprise a machine-learning assembly, including the parameters for fine-tuning an algorithm and/or features of a model).
The Licenses do not cover code (executable code used in conjunction with an AI artifact will be separately licensed under a well-known permissive open source license such as Apache 2.0 or MIT) or services (AI technology provided as a hosted service, e.g. through a web UI or API, will have separate terms of service that apply to developers and users of the hosted service).
How do the AI2 ImpACT Licenses work?
AI2 classifies each artifact we release under a license by risk: low, medium, or high.
Unlike traditional software and content licenses where use rights and restrictions are structured around a specific artifact, the AI2 ImpACT Licenses are artifact-agnostic and are instead structured according to the risk level we’ve assigned a given artifact. In other words, what you are permitted to do with an artifact depends on its risk category (low, medium, or high), rather than the type of artifact (e.g., data, a model, or something else). Each risk classification results in different license terms tailored for that risk category as summarized here:
|Risk-Based License Terms||AI2 ImpACT-LR License |
|AI2 ImpACT-MR License |
|AI2 ImpACT-HR License |
|Install & use artifacts|
|Create & use derivatives|
|Distribute artifacts||Internal Use Only||Internal Use Only|
|Distribute derivatives||Internal Use Only|
|Use-Based Restrictions (artifacts & derivatives)|
|Prominent notice of artifact modifications with derivative distribution||N/A|
|Derivative Impact Reports||N/A|
What do the AI2 ImpACT Licenses say?
We understand that the AI2 ImpACT Licenses are different from what you typically expect to see, so we have provided a more simplistic (but not comprehensive) overview of the AI2 ImpACT Licenses here – please always refer to the full legal text of the AI2 ImpACT licenses to make decisions about your use case.
First, we’ll start with the basic use rights and restrictions that apply across the board in all of the AI2 ImpACT Licenses for low, medium, and high risk:
- Use, install, and create derivatives: You can install and use any artifact as-is and create derivatives of any artifact. “Derivatives” can mean many different things in the AI context, so we intentionally defined it broadly to cover both known and unknown types of derivative works that can be generated from a data artifact or model artifact. Some examples include:
- Data derivatives: modifications to our dataset; a new dataset that incorporates some or all of our data; a model trained on some or all of our data using any code; any application that embeds or incorporates a model trained on our data.
- Model derivatives: iterations of our model weights by modifying the algorithm or parameters (the “diff”); our model weights fine-tuned with other data; an application that uses our model weights with any code; another model based on training data derived using our model weights; any of the foregoing incorporated and/or made part of a larger application, tool, or product.
- Obey the Use-Based Restrictions: You must always comply with the Use-Based Restrictions in Exhibit A of the AI2 ImpACT Licenses.
The risk categories of low, medium, and high inform what you can do with the respective artifacts and the derivatives you create from them.
- Use and Creation: as explained in the previous section, you may use and install the artifacts and create derivatives of the artifacts for your internal use regardless of risk category. Specifically:
Low Risk Medium Risk High Risk Artifact: Can you use artifacts as-is? Yes Yes Yes Derivatives: Can you create and use derivatives for your own benefit? Yes Yes Yes
- Distribution: to “distribute” means any external use or sharing by any means, including making something publicly available online or providing it to third parties. Each risk category has different distribution rights for (1) the artifact itself and (2) the derivatives you create, as shown below:
Low Risk Medium Risk High Risk Artifact: Can you distribute artifacts as-is? Yes (subject to Distribution Rules) No (internal use only) No (internal use only) Derivatives: Can you distribute derivatives you create? Yes (subject to Distribution Rules) Yes (subject to Distribution Rules) No (internal use only)
- Distribution Rules: As described above, if you distribute any low-risk artifact or low/medium-risk derivative, you must comply with the following distribution rules:
- Flow Down Use-Based Restrictions: The Use-Based Restrictions should be included in an enforceable legal agreement for all downstream use and/or further distribution by your end users. Our intent is for the Use-Based Restrictions to continue running downstream.
- Attribution: Include the applicable attribution notice with your distribution as provided in the legal text of the respective AI2 ImpACT Licenses. This attribution should continue to run downstream.
- Notices: Retain all other copyright, IP, and attribution notices that come with the artifact. Also include a prominent notice stating how the original artifact was used or modified.
Derivative Impact Reports
Derivative Impact Reports are intended to be a light-and-quick version of a model card or a dataset card that can be completed in a few minutes. The purpose is to increase transparency about the inputs for AI development and provide more visibility into downstream AI applications.
Creating a Derivative Impact Report:
Answer just a few short questions using our submission form before you release your derivative, or after you have substantively updated it.
- The form takes just a few clicks and asks for basic information about your derivative and how it was developed, such as your name, email, and intended use.
- Check the box at the bottom of the form to receive a copy of your responses by email – this is your completed Derivative Impact Report.
- Publish, post, or otherwise make your Derivative Impact Report available to the general public without any requirements or barriers to access (e.g., in a GitHub repo or linked to a model card or dataset card).
This is a good-faith self-reporting tool and we are relying on your honesty and integrity to make these submissions useful for future researchers and the public at large. The more we share together, the more we can learn from each other.
We include typical breach and termination provisions for violations of the AI2 ImpACT Licenses that are not remedied within 30 days. Further, if you violate the Use-Based Restrictions and the issue is not remedied or resolved with AI2 within 30 days, AI2 may post a prominent notice on its website (or otherwise make a notice publicly available) stating you violated those restrictions and that your license rights were terminated.
We welcome feedback on this project in general and on this iteration of the AI2 ImpACT Licenses in particular at firstname.lastname@example.org.
When should I submit a Derivative Impact Report?
Before you release your derivative publicly, or after you have substantively updated or modified your derivative.
I developed a Model Derivative and a Data Derivative as part of the same project. Do I need to do a Derivative Impact Report for each?
Yes – the questions for a Model Derivative and a Data Derivative are slightly different. Please submit a separate Derivative Impact Report for each one.
Where can I create a Derivative Impact Report?
Use our submission form. Check the box at the bottom of the form to get a copy of your responses by email. This is your completed Derivative Impact Report.
I completed a Derivative Impact Report. Where do I put it?
Anywhere that’s convenient, so long as it's freely available to the public without cost or access restrictions. We recommend a GitHub repo associated with your derivative or as additional information linked to your model card or dataset card if you’ve made one.
What will you do with my Derivative Impact Report?
We plan to show Derivative Impact Reports (ours, yours, and others) on our website in the future.