The Open Source Initiative (OSI) has introduced the first official definition of open source AI, known as the Open Source AI Definition (OSAID). After several years of collaboration with various stakeholders, including academia and industry, this definition aims to provide a clear standard for determining whether an AI model is genuinely open source. According to OSI Executive Vice President Stefano Maffulli, establishing a consensus on this definition is crucial for aligning policymakers and AI developers. The OSAID stipulates that an AI model must provide detailed information about its design, allowing for substantial recreation, and must disclose relevant details about its training data, including its origin and processing methods. For an AI model to be classified as open source, it must offer full access to its construction components, including training code and data filtering, and developers should have the freedom to use, modify, and build upon the model without seeking permission. Although the OSI lacks enforcement capabilities, it hopes the community will uphold this definition and correct misuse of the term "open source. " Many tech companies, including Meta, have claimed to offer open source models but often do not adhere to these new standards.

For example, Meta requires special permissions for certain users of its Llama models, which Maffulli and the OSI criticize. While Meta supports the OSI's efforts, it challenges the OSAID's definition, arguing that it fails to reflect the complexities of current AI models. Some experts believe the OSAID doesn't adequately address concerns about proprietary training data and copyright issues related to AI models. Maffulli acknowledges that updates to the definition may be necessary and has established a committee to monitor its application and propose future amendments. The OSI intends to maintain an open approach involving a broad range of stakeholders in this evolving landscape of open source AI.