Apple has revealed its approach to improving AI models by analyzing user data privately using synthetic data. The company is generating synthetic messages and using “differential privacy” to compare their accuracy with real user data.
To do this, Apple creates synthetic messages that mimic the format of user-generated content but contain no actual user data. These messages are then used to derive embeddings, or representations of key message dimensions like language and topic. The embeddings are sent to a small number of devices that have opted-in to share device analytics with Apple.
The devices compare these embeddings with real emails to determine which ones are most accurate. This approach will be used to improve Genmoji models, as well as other AI-powered tools like Image Playground, Visual Intelligence, and Writing Tools. Additionally, users who opt-in to share device analytics will also receive synthetic data to help improve email summaries.
By using this private analysis method, Apple aims to refine its AI models while maintaining user privacy.
Source: https://techcrunch.com/2025/04/15/apple-details-how-it-plans-to-improve-its-ai-models-by-privately-analyzing-user-data