As always, the devil is in the details. They defined an AI system as software that uses any of: (a) Machine learning approaches, including supervised, unsupervised and reinforcement learning, using a wide variety of methods including deep learning; (b) Logic- and knowledge-based approaches, including knowledge representation, inductive (logic) programming, knowledge bases, inference and deductive engines, (symbolic) reasoning and expert systems; (c) Statistical approaches, Bayesian estimation, search and optimization methods.

That could very reasonably be interpreted to mean all software. The act is mostly concerned with high risk applications. These are based on intended use, which is obviously tricky in the context of general purpose LLM’s. Most of the covered intended uses make sense, but in combination with their definition above it could mean that, for example, any developer of EdTech software in the EU could be regulated under the act. More broadly, the act was clearly not written with general purpose foundation models in mind. For example, Article 10 requires that all training data be “relevant, representative, free of errors and complete.” That is pretty much impossible to comply with if you’re training on The Pile or similar.