The rhetoric of open source AI has risen alongside the conversation around generative AI. This despite the obvious challenges to reproducibility and extension of AI systems, such as the cost of compute, the inscrutability of the dominant machine learning models, and need to preserve data privacy. These constraints come in the way of using, studying, modifying and distributing AI models- central tenets of open source software. What then do we mean by open source AI?

In this talk we will critically analyze the different descriptions of openness deployed by and for AI products. What aspects of openness do they highlight and why? By recognizing upfront the material challenges to reproducibility, transparency and extension in machine learning, we will surface strategies that can continue the values of open source software in machine learning development.