1
$\begingroup$

As I was reading the AI act I was wondering how it will apply to current open LLM (say Mixtral of Mistal AI).

It seems that there are some specific arguments regarding open LLM that would put them outside of the AI act scope:

  • They are free (are they still "put on the market" ?)
  • They are tools (not an AI system with a usage), it's there reusage in an AI system that would

Was the topic of LLM covered somewhere that I am missing ? Bonus question: given their prevalence and potential impact, will they be considered high risk ? (or its their reusage ?)

$\endgroup$
2
  • 1
    $\begingroup$ I don't see a fundamental boundary in LLMs as tools or agents. If the system only answers questions, but does it always to the best of their "knowledge" it is a tool. But if it sometimes asks questions out of some curiosity (objectively describable as missing information "wanted"), sometimes answers but sometimes refuses to answer or lies on purpose due to some game constraint, is it an agent? I don't think the changes needed for those very different behaviours are any deep. $\endgroup$ Dec 12, 2023 at 11:15
  • 1
    $\begingroup$ By "AI act", are you referring to europarl.europa.eu/news/en/headlines/society/20230601STO93804/…? Can you please include the link to the AI act in your own question? Thanks. $\endgroup$
    – nbro
    Dec 13, 2023 at 0:59

1 Answer 1

1
$\begingroup$

Yes.

The EU AI act has a specific section about generative AI. I'll quote it.

Generative AI, like ChatGPT, would have to comply with transparency requirements:

  • Disclosing that the content was generated by AI
  • Designing the model to prevent it from generating illegal content
  • Publishing summaries of copyrighted data used for training

So, clearly, LLMs (open or closed) are subject to the EU AI Act. So, I don't really know why you have doubts about it. Have you maybe read some unofficial document or are you referring to a different act?

Moreover, in the official linked document, these rules are not placed under the "High risk" section, but I suppose generative AI is also high-risk because in the "High risk" section we also have e.g. "Education and vocational training", which generative AI has been used for.

The act will probably evolve in the next months or years. See also the previous detailed one here.

$\endgroup$
2
  • $\begingroup$ My question related to open LLM (Mistral AI distributing weights), for which the provider might be someone else. As such its not entirely clear that those bullet points apply to mistral / the entity using weights or both. $\endgroup$ Dec 15, 2023 at 10:34
  • $\begingroup$ I think the AI Act and these rules still apply (there's nothing that suggests otherwise). I think your question is different. Your question is "who is responsible for ensuring this specific generative AI complies with these rules". If this is not your question, then I think that's the most interesting question anyway. $\endgroup$
    – nbro
    Dec 15, 2023 at 13:01

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .