The Open Source Initiative (OSI) has introduced the first official Open Source AI Definition (OSAID), providing a structured framework to help identify when AI systems can be classified as truly open-source. Launched at the All Things Open conference, this definition is the result of years of collaboration among global industry stakeholders, including tech organisations, government bodies, and open-source advocates.
The OSI, renowned as a leading authority on open-source standards, crafted this framework to apply the same open-source principles to AI as it does to software. The new framework is intended to ensure that AI models, including their parameters, weights, and training data, meet the criteria of open-source, allowing for unrestricted use, study, modification, and sharing.
An AI system must fulfil four key freedoms to be considered open-source:
- Freedom of Use: Users must be able to use the AI system for any purpose without needing additional permissions.
- Freedom to Study: The system’s components should be accessible, allowing users to understand and analyse its inner workings.
- Freedom to Modify: Users must be able to alter the AI for any purpose, including to adjust its outcomes.
- Freedom to Share: The system should be sharable, both in its original form and with modifications, for any purpose.
For these freedoms to be realised, users must have access to all necessary resources, such as source code, data, and model parameters, which are essential for full transparency and modifiability.
Carlo Piana, chair of OSI’s board, praised the process that led to the definition’s release, describing it as inclusive and rigorous. “Version 1.0 of the Open Source AI Definition embodies the same high standards as traditional open-source software, ensuring that AI systems adhere to the same essential freedoms,” he said.
A key requirement of the OSAID is transparency around training data, with the definition mandating that open-source AI systems disclose sufficient information about their data sources. This stipulation aims to enable skilled individuals to recreate similar systems using comparable data, a principle highlighted by Ayah Bdeir, Mozilla’s head of AI strategy. Bdeir acknowledged the framework’s practical approach, aiming to balance idealism with real-world viability.
The Digital Public Goods Alliance (DPGA), an advocate for accessible digital resources, voiced strong support for OSI’s work. DPGA CEO Liv Marte Nordhaug announced that the new definition would be integrated into updates to their own standards for AI, underscoring the value of transparency and accessibility in the AI field.
Also endorsing the definition is the EleutherAI Institute, a non-profit AI research organisation. Executive Director Stella Biderman called the OSAID “an essential step in promoting open-source principles within AI,” and emphasised that the framework supports independent researchers and promotes transparency among large AI developers.
The release of the Open Source AI Definition also clarifies the legal considerations involved, specifying that while no specific licence is required, AI models may rely on licences or legal mechanisms to ensure parameters are freely available to the public.
Stefano Maffulli, OSI’s Executive Director, noted the challenges in formulating the framework, including differing opinions among contributors. Despite this, he highlighted that the process led to a balanced result. “This is just the beginning; the OSAID will evolve with further input from the community,” Maffulli said.
With the new OSAID, OSI has laid a foundation for a more transparent and accessible AI industry, inviting further community input as they refine the standards to reflect advances in both AI and open-source policy.
Reference: https://www.artificialintelligence-news.com/news/industry-leaders-back-open-source-ai-definition/
