The joke is that the model itself runs locally, but it still is basically spyware that then will be used for data harvesting. Typically you run local LLMs to avoid this exact purpose / risk. So MS isn't even willing to give you the performance benefit of running the LLM on their servers for doing this.