A brief note on machine learning models, including generative neural networks such as large language models and most of what's been misleadingly characterized as “AI” these past few years.

These models essentially consist of a lossily-compressed version of their training data (for example). The degree of lossiness varies, and all such models can be induced to output arbitrarily large portions of their training data verbatim. Furthermore, because these are black box models, there is no way to convincingly demonstrate that they will never reproduce verbatim any given portion of their training data.

Therefore, I consider machine learning models to be a derivative work of their training data. In particular, any machine learning model trained upon this blog is a derivative work thereof, and must therefore abide by the blog's license (CC BY-SA 4.0) and be released under that same license.