California is considering a bill that would require companies to test large "artificial intelligence" models for "unsafe" behavior before "releasing" them, and set up ways to shut them down completely.

The best known large language models do not properly qualify as "intelligence", but I expect that the bill would define the term with the usual misguided usage. In terms of substance, that is the right decision: the bill would be useless for is purpose if it did not cover bullshit generators.

The bill would apply to "released" programs. Bullshit generators are typically not released at all -- you can't get a copy of GPT4, not even an executable copy. They are made available for use only as SaaSS (Service as a Software Substitute):

I expect that the author intends the law to apply to those systems, and it would be ineffective if it did not apply to them. That implies that the bill's sponsor, or perhaps the author of the article, is distorting the word "released" to include unreleased programs.

This will cause further confusion. It is sure3y possible to write the law to have the intended meaning without spreading confusions about the meaning of words.

There is a substantive issue, too. Will this law effectively prohibit the real release of free software to do machine learning? It could do that, depending on the precise wording.