It will at least greatly hinder LE's capability to do massive minority report type dragnets, targetted violence incitement campaigns, or grading workers or schoolchildren based on their facial expressions etc extremely nasty stuff.
The banned use cases are very specific and concerns systems explicitly designed for such dystopian shit. AI giving advice how to build weapons or explosives is not banned here. The "unacceptable risk" category does not concern companies like MistralAI or Black Forest Labs. This is not the same idiocy.
I agree with you, but how do you effectively prevent it? Standards vary across countries; what's not acceptable in Europe might be acceptable elsewhere.
For instance, discussing or questioning Nazism is illegal in Germany but allowed in many other countries. Should every LLMs be restricted globally just because Germany deems it illegal?
Similarly, certain drugs are legal in the Netherlands but illegal in other countries, sometimes even punishable by death. How do you handle such discrepancies?
Let's face it: most of the time, LLMs follow US-centric anti-racism guidelines, which aren't as prominent or necessary in many parts of the world. Many countries have diverse populations without significant racial tensions like United States and don't prioritize African, Asian, or Latino positivity to the same extent.
Moreover, in the US, discussions about the First or Second Amendment are common, even among those with opposing views, but free speech and gun rights are taboo in other societies. How do you reconcile this?
In practical terms, if an LLM refuses to answer questions because they're illegal in some countries, users will likely use uncensored models instead, rendering the restricted ones less useful. This is why censorship is never successful except by North Korea and China.
Take Stable Diffusion as an example: the most popular versions (1.5, XL, Pony) are flexible for unrestricted use, whereas intentionally censored versions (like 2.1 or 3.0) have seen limited adoption.
The "unacceptable risk" category is not about chatbots or LLM capabilities in general. It's not about censorship or banned topics. It's about certain specific applications and systems tailored for them.
The banned use cases are very specific and concerns systems explicitly designed for such dystopian shit. AI giving advice how to build weapons or explosives is not banned here. The "unacceptable risk" category does not concern companies like MistralAI or Black Forest Labs. This is not the same idiocy.