Apple’s Private AI Training Is Coming to More Features

Apple is using new privacy-friendly AI methods to improve features in Apple Intelligence—without ever seeing your private data.

According to Apple’s machine learning research website, these tools help the company understand how people use features like Genmoji and email summaries, but they don’t collect or store anything personal.

To figure out what kinds of Genmoji people like, Apple uses a system called differential privacy. This means your iPhone only shares vague, anonymous signals that are mixed with random noise, so Apple can’t link anything back to you. It helps Apple learn general trends—like lots of people asking for “dinosaur in a cowboy hat”—without seeing exactly what you typed.

For longer content, like emails, Apple takes a different approach. Instead of looking at real emails, Apple creates fake (or “synthetic”) ones that feel realistic. Then, your phone privately compares these fake messages to your real ones and sends back a simple vote—never the actual content.

This helps Apple figure out which types of messages are most common, so it can train its AI to do a better job summarizing or writing text. Again, this only happens if you’ve opted in to share analytics, and the content of your messages never leaves your device.

“Apple currently uses differential privacy to improve Genmoji, and in upcoming releases we will also use this approach, with the same privacy protections, for Image Playground, Image Wand, Memories Creation and Writing Tools in Apple Intelligence, as well as in Visual Intelligence,” explained the company on its machine learning website.

According to Bloomberg’s Mark Gurman, an upcoming beta version of iOS and iPadOS 18.5 and macOS 15.5 will include this new system. Yesterday we saw second betas of this software released for developers.