The ChatGPT integration was heavily gated by Apple and required explicit opt-in. That won't be the case with the Gemini integration. Apple wants this to just work. The privacy concerns will be mitigated because Apple will be hosting this model themselves in their Private Cloud Compute. This will be a much more tightly integrated solution than ChatGPT was.
This story says the custom model will run on-device and in Apple's Private Cloud Compute. The implication is that Google will not see the data. The "promise" of Private Cloud Compute is that Apple wants it to be trusted like "on-device".
Presumably cutting Google out of getting the data from this is part of why this story first was mentioned last year but is only now sounds close to happening. I think it's the same story/project.
And you don't think they will include an abstraction layer?