The Open-Source Initiative (OSI) is planning to form a definition of “Open Artificial Intelligence” (not to be confused with OpenAI, a company selling proprietary autocomplete software whose technical details only grow less open with each iteration). Unfortunately, odds of the definition requiring the release of training data are slim: the OSI’s executive director isn’t keen on the idea himself.
I see libre/open-source software as a means to reduce dependence on a vendor, and mitigate the risk of user domestication. As long as training data is out of the community’s reach, it’s impossible for the vendor to be replaced. Yes, it’s possible to customize or re-train the model, but the vendor remains in control of its future development.
Recent decades have tested the effectiveness of liberating source code as a defense against user domestication, as I explain in another blog post. But to re-define Open Source to allow labelling a model that is impossible to competitively fork would be to miss the whole value of FOSS in my eyes: to allow users to own not just their tools, but those tools’ futures.