Apple has introduced a new approach to enhance its AI models that protects user privacy while addressing complaints about poor performance in features like email and notification summaries. In a blog post, the company outlined its plan to use synthetic data and on-device comparisons as part of a privacy-first training process.
Instead of using genuine user data, Apple creates synthetic data—fake messages that have the structure and properties of real emails or texts but contain no actual user content. These synthetic messages are intended to address a wide range of themes, lengths, and styles. Apple then sends these synthetic communications, in the form of "embeddings" that record linguistic and structural characteristics, to select a number of user devices that have opted into Device Analytics.
The participating devices compare the synthetic embeddings to actual user data stored locally—such as recent emails or messages—to determine which synthetic inputs are the most similar. Instead of transferring any data back to Apple, the devices just return a signal indicating which synthetic sample was the closest match. This feedback is then used by Apple to refine and improve its artificial intelligence models.
According to the company, this method depends on its long-standing usage of "differential privacy," a process that injects random noise into data sets to prevent people from being identified while still offering valuable insights. This technique guarantees that sensitive information never leaves the user's device.
Apple has previously used this technology to develop the Genmoji function and plans on bringing it to additional features like Image Playground, Image Wand, Memories Creation, Writing Tools, and Visual Intelligence. The system is being evaluated on beta versions of iOS, iPadOS 18.5, and macOS 15.5.
With this change, Apple plans to improve the relevance and quality of its AI-generated outputs while keeping its track record for strong user privacy, an essential step in its attempt to keep up with competitors in the AI area.