After careful evaluation, we determined that Google’s technology provides the most capable foundation for Apple Foundation Models and we’re excited about the innovative new experiences it will unlock for our users.

— Apple picks Google’s Gemini to run AI-powered Siri coming this year

If this reported news is true, it seems like an unambiguous upgrade for Siri users on iOS, macOS, and anywhere else Siri runs. I have a BigMe HiBreak Pro Android e-ink tablet, and even on relatively constrained hardware the experience of using Gemini and Gemini Live is drastically better than Siri.

This announcement provides more questions than answers though — questions that matter a lot to me.

If this model runs purely on device, then that’s great news. Gemini’s local models are terrific, and better than whatever Apple is doing today with Qwen models in their FoundationModels framework today.

If this runs in the cloud, then I have more questions. I care deeply about my privacy, so I abstain from using Gemini for anything private. By that I mean I won’t ask Gemini any health questions, discuss anything personally identifiable, and if I do use it for code it’s only for my open source projects.

I’ve drawn the line there because Google unlike OpenAI and Anthropic provides no way for consumers to opt out of having their data used for model training purposes. I simply do not want that, and will not use a Gemini-powered Siri if those promises are not there.

Either way I’m curious to see how this affects Apple’s FoundationModels frameworks. The framework provides a level of abstraction so developers have to care very little about the underlying models. The results obviously matter, and there are particulars that matter like whether a model is local or remote — but ultimately this is a generic set of APIs that you could easily slot a new model into.

These are just the first questions I had, and it all depends on the details. My real hope is that Apple respects their stated values in a world of AI, and I expect to have more answers soon — and more questions stemming from those answers.