The Ministry of Digital Development published a draft law on AI regulation for public discussion. The main catch is in Article 17 on "cross-border technologies." In short: their operation in Russia can be legally restricted and banned. The law is set to take effect on September 1, 2027.
That means any model that sends your users' prompts to servers in California (ChatGPT, Claude, Gemini) is a target under the law.
Also, the concept of a "trusted model" appears. For your AI to be allowed in the public sector or critical information infrastructure facilities, the model must process data exclusively within Russia.
The obvious solution, which corporate lawyers have already started discussing, is open weights. Various DeepSeek, Qwen, or Llama models that you download and deploy in a closed on-premises environment on your own servers within the country. Data doesn't leak out -> you don't fall under cross-border restrictions -> everyone is happy.
And there's also a wonderful Article 11. The AI developer and operator are liable for the result of generation that violates the law, "proportionally to the degree of fault." If your model generates instructions for creating something prohibited or produces a "discriminatory algorithm," you can't get away with saying "it's all the neural network, I just downloaded the weights." Although there is a caveat that you are exempt from liability if you have taken "exhaustive measures." What constitutes exhaustive measures will, of course, be decided by the court.
In short, it's time to learn to work exclusively with local models. If only we had the hardware to run them 🫠
Comments
0No comments yet.