Would CC BY-SA do? It requires attribution, and there is no real way to attribute text from an LLM to the training data. Assuming that LLM output counts as a derivative work of the training data, which a consensus on has not be reached. (It would also be required for any other copyright based solutions)

Edit: reading the licence, it requires "a credit identifying the use of the Work in the Derivative Work", which may allow companies to just post a huge page showing all the training input, but I doubt this counts as "reasonable means".

Given that for-profit immitative ML companies give exactly 0 fucks about copyright, there isn't a single license that could possibly protect you.

I think you can just take the text of a license like Creative Commons but modify it to say that you're not giving permission to use your content to train AI.

However, if you go this route, remove the Creative Commons name from the license. CC does not permit their name/trademark being used on less permissive licenses.

I mean you can make whatever terms you want in a license if it's yours. As for standard licenses like MIT or GPL. I'm not aware of any that currently exist.

This is what we are going to find out in the next three years. Someone could come up with a license, but we won't know if it's viable until it gets tested by the courts. There's still debate on whether scraping from copyrighted works then using that data to train a LLM (or any other ML model) is an allowed use. Technically, that is transformative of the original works, but it uses vast numbers of copyright works. The New York Times is looking at suing OpenAI over this exact issue.

In any case, if you did come up with a license, you would also have to enforce it. With no clear legal precedents that would require filling suit against any violator.

I've actually been trying to train a model on CC (the attribution for the millions of entries is the open source dataset) - I do think you'd want to just use MIT or similar?