Should OpenRAIL licenses be considered OS AI Licenses?

Part of the Deep Dive: AI Webinar Series

Advances in AI have been enabled in-part thanks to open source (OS) which has permeated ML research both in the academy and industry. However, there are growing concerns about the influence and scale of AI models (e.g., LLMs) on people and society. While openness is a core value for innovation in the field, openness is not enough and does not address the risks of harm that might exist when AI is used negligently or maliciously. A growing category of licenses are open responsible AI licenses ( which include behavioral-use clauses, these include high profile projects such as Llama2 ( and Bloom ( In this proposed session the panelists would discuss whether OpenRAIL ( licenses should be considered as OS AI licenses.

Topics will include: Whether the definition of OS is not adequate for AI systems; Whether OS of AI systems requires open-sourcing every aspect of the model (data, model, source) and whether that is feasible; How data use requirements could be included in such a definition; and therefore, whether inclusion of behavioral use restrictions is at odds with any future definition of OS AI. In responding to these questions the panelists will discuss how the components of AI systems (e.g., data, models, source code, applications) each have different properties and whether this is part of the motivation for a new form of licensing. The speakers have their own experience of building, distributing and deploying AI systems and will provide examples of these considerations in practice.

Webinar summary

In this webinar hosted by the Open Source Initiative as a part of the “Deep Dive: Defining Open Source AI” series, Daniel McDuff, Danish Contractor, Luis Villa and Jenny Lee discuss the evolution of AI licensing, particularly within the context of open source. Initially, RAIL licenses were conceived in response to concerns about AI technologies being released without proper ethical considerations, inspired by events like the Cambridge Analytica scandal. They aimed to self-regulate and establish norms within the AI community. However, as AI models became more complex, the question arose about how to protect and standardize the use of these models and data. Some argue that the term “open source” should not be diluted and that other terms like “responsible source” should be used to describe AI models with use restrictions. The tension between the desire for open access and the need for responsible use remains a key challenge in the evolving AI landscape, with calls for tooling and standardization to facilitate compliance and enforcement of use restrictions.