Oh! You have misunderstood the whole concept of privacy. I have a thought experiment for you:-
Let's assume Microsoft is not lying 🤥. The data (screenshot) remains on device, which is passed to some AI model like Image-to-text etc. This model generates text on-device. But no where Microsoft guarantee's that the text generated or output from those AI models won't be sent to the Microsoft. They only say the screenshots and AI models remain on-device, but the output/metadata can be sent to Microsoft.
That is the issue. Earlier there were many apps where Microsoft couldn't pry because they were encrypted etc. Now they don't need to break any encryption they just need metadata. That's easy to transfer and use.