Apple is changing how it trains its AI models by avoiding the collection or copying of real user data from iPhones and Macs. Instead, it relies on synthetic data—artificially created to mimic user behavior—and uses differential privacy to enhance features like email summaries without accessing personal emails or messages.

For users who opt into Apple’s Device Analytics program, AI models generate synthetic email-like messages and compare them locally on the device with a small sample of real user content. The device then sends only aggregated information about the closest matching synthetic data back to Apple—no actual user data ever leaves the device.

This privacy-focused method extends Apple’s long-standing use of differential privacy, which adds randomized data to protect individual identities while improving AI features such as Genmoji. Apple plans to expand this approach to other Apple Intelligence features like Image Playground and Writing Tools.

The system is currently being tested in beta versions of iOS 18.5, iPadOS 18.5, and macOS 15.5. While it remains to be seen how well this approach performs, it clearly demonstrates Apple’s commitment to balancing strong user privacy with advancing AI capabilities.