Why does TLS not actually hide what I am asking my AI assistant?

Streaming LLM responses emit tokens one at a time over an encrypted connection, and each packet arrival changes size and timing in ways that correlate with topic. The Whisper Leak paper published in November 2025 demonstrated topic classification with above 98% AUPRC across 28 major providers using only encrypted traffic metadata. Several providers issued patches but each mitigation trades query privacy for added latency and there is no agreed specification for what a sufficient defense looks like. The structural problem is that any low-latency streaming protocol leaks information through timing, so padding and noise are palliatives rather than solutions.