Defining an Open Source AI

The traditional view of open source code implementing AI algorithms may not be sufficient to guarantee inspectability, modifiability and replicability of the AI systems. The Open Source Initiative is leading an exploration of the world of AI and Open Source, diving around the boundaries of data and software to discover how concepts like copy, distribution, modification of source code apply in the context of AI.

AI systems are already deciding who stays in jail or which customers deserve credit to buy a house. More kinds of “autonomous” systems are appearing so fast that government regulators are rushing to define policies.

Artificial Intelligence/Machine learning, explained at a high level, is a type of complex system that combines code to create and train/tune models, and data used for training and validation to generate artifacts. The most common tools are implemented with open source software like TensorFlow or PyTorch. But from a practical perspective, these packages are not sufficient to enable a user to exercise their rights to run, study, modify and redistribute a “machine learning system.” What’s the definition of open source in the context of AI/ML? Where is the boundary between data and software? How do we apply copyleft to software that can identify your cats in your collection of pictures?