Apple has unveiled a new initiative aimed at enhancing the performance of artificial intelligence models while strictly safeguarding user privacy. The cornerstone of this initiative is that Apple devices can complete crucial AI training processes without directly utilizing or transmitting user data.

In detail, Apple devices will compare synthetic datasets with samples of emails and messages voluntarily shared by users to identify synthetic inputs that closely resemble real data. This entire process takes place on the device, sending only a "signal" to Apple's servers instead of raw data, indicating which synthetic sample is most similar to the user’s data. Based on this, Apple optimizes AI text output functions, such as email summary generation.

Prior to this, Apple's AI models primarily relied on synthetic data for training, which, although ensuring privacy, might limit the practicality of AI responses. Recently, Apple's AI feature development has faced challenges, leading to delays in the release of some flagship features and the replacement of the head of the Siri team.

To turn things around, Apple is introducing this new AI training system through the beta versions of iOS and iPadOS 18.5, as well as macOS 15.5. This system integrates Apple's long-standing differential privacy technology, which ensures that data cannot be traced back to specific users by randomly adding information to datasets.

Since launching iOS 10 in 2016, Apple has consistently emphasized the application of differential privacy in protecting user data. This new initiative continues along that technological path. With this innovative solution, Apple seeks to strike a balance between improving AI performance and protecting user privacy.