Defining "Open-Source AI": A Legal Analysis of Copyleft Principles, OSI Standards, and Corporate Licensing Practices

“Open-source AI” is a commercial and regulatory term that is gaining in popularity, but AI systems don't fit the source-code model of traditional open source software. This paper explores copyleft, the Open Source Definition, the Open Source AI Definition 1.0 (OSAID), and some key examples of corporate licensing policies, through a doctrinal and comparative legal lens. It states that copyleft is compatible with open source AI where the shared freedom includes share-alike provisions that ensure downstream freedoms; however, it does not include reciprocal language that turns into user restrictions, field of use restrictions, competition restrictions or scale restrictions. The parallel comparison between the weights of Meta Llama 4, Phi from Microsoft, xAI Grok-1, and the licensing of OpenRAI makes it clear why open weights, permissively licensed components, and responsible-use licensing should not be considered open-source AI. Three tests are suggested: Completeness of the artefacts, Freedom of licence, Copyleft discipline. The paper concludes that legal labelling is crucial for research reproducability, procurement, compliance and the integrity of the open AI ecosystem.