There are some carve-outs for FOSS, but. The biggest problem is that the copyright lobby got in a body blow. It won’t be enough to make the IP fanatics of lemmy happy, but it’ll make some people money.
Model makers must have a policy in place to ensure compliance with the EU copyright directive. That means that websites can use a machine-readable opt-out from AI training. I think this is a big reason why we’re now hearing about these deals with reddit and other copyright holders.
Also, model makers must create a summary of the copyrighted training data used. The AI office, which is supposed to enforce this act, is to provide a template for a sufficiently detailed summary. A lot will depend on the AI office and the courts.
A couple problems are obvious. Many enthusiasts will not bother with the paperwork. What will that mean for Hugging Face? Or the AI Horde? The EU is certainly the wrong place to build a business around hosting AI models.
The current open models will likely become “illegal”. The makers would have to retroactively provide then necessary documentation, but why would they bother? Even if they did, there is the question about the policy regarding the opt-out. Mind, that doesn’t outlaw possession or use. It simply means that businesses may be fined for providing downloads or inference to EU residents.
I think this will likely have a chilling effect on open models. EG Meta could simply say that “Llama 4”, if it were to come, is off-limits in the EU. But that might not be enough to indemnify them. Or they could try to comply, which would cost them money for no clear gain. And/or they’d have to leave out data without regard for quality.
Research institutions are not bound by the opt-out. They might become a source of open models.
The carve-outs also do not apply to so-called high-risk AI systems. That would make sense if the act made sense.
LLMs, image diffusion models, and such are termed GPAI (general purpose AI). They are not considered high-risk, by default. They are considered high-risk only once they are adapted to a high-risk purpose. Of course, the line isn’t that clear. Say, a teacher could use an LLM to grade tests. Regulators might cause problems there.