Apple has announced a new way to improve its AI models without directly using or copying user data from iPhones or Macs. The company’s method involves comparing synthetic data to real-world examples—like emails or messages—from users who have opted into its Device Analytics program. Instead of sending actual data, devices send a signal indicating which synthetic sample most closely resembles the real data.

How It Works: Privacy-Preserving Comparison

Devices analyze synthetic inputs locally and determine which ones best match real samples. Only the choice—not the data itself—is shared with Apple. This allows the company to refine its AI responses (like email summaries) using the most commonly selected fake examples, without accessing personal content. Apple claims that user data never leaves the device during this process.

Apple’s Challenges With AI Performance

Until now, Apple has relied solely on synthetic data to train its AI models. According to Bloomberg’s Mark Gurman, this has sometimes led to less effective performance. Apple has reportedly delayed some features of its “Apple Intelligence” suite and even replaced the head of its Siri team. The new training method is now being tested in beta versions of iOS and macOS.

Differential Privacy Still Plays a Key Role

Apple continues to use its long-standing method of differential privacy, introduced in 2016, to protect user information. This involves injecting random data into the training process to prevent individual users from being identified. The same method is now being applied to its new AI training strategy, ensuring privacy while improving performance.