Device-local AI is being sold as a clean break from cloud surveillance, and the sales pitch is working faster than the technical caveats can catch up. The framing around Qwen 3.5 and Apple hardware [1][4] treats local processing as synonymous with private processing—a conflation that complicates immediately. Apps can still sync results and send metadata even when the primary model runs locally. The pitch is not wrong that local inference reduces exposure; it is wrong that local inference eliminates it. Compliance-focused content reinforces…