Traditional end-to-end encryption cannot be directly applied to AI conversations under the current cloud-based inference architecture, because large language models must read user input on the server side in order to generate a response.