Ask HN: Keystroke biometrics in various web-based LLM UIs?

Using PerplexityAI through the web UI, there is significant latency between keystrokes and text manifestation in the prompt field. Is this merely overly aggressive next-word prediction, or is there plausibly typing/keyboard biometrics being captured and used for xyz purposes?

If not, why not? If so, what is being done with the data and what methods beyond resisting canvasing are viable to prevent this?

Also, is there any known, effective OS software, that can securely act between keystrokes and the system to mitigate this?