Apple plans to train AI using your on-device data without accessing personal content or emails: Report

Apple will soon update its AI training approach by using on-device data to improve features while prioritising user privacy, set for release in upcoming iOS and macOS updates.

Signage for Apple Intelligence inside an Apple store in Walnut Creek, California, US, on Monday, Feb. 24, 2025. Apple Inc., as it seeks relief from US President Donald Trump's tariffs on goods imported from China, said that it will hire 20,000 new workers and produce AI servers in the US. Photographer: David Paul Morris/Bloomberg (Bloomberg)

Apple is introducing a new approach to training its artificial intelligence (AI) models, which aims to enhance their performance while maintaining user privacy. This shift is set to roll out in the upcoming beta versions of iOS 18.5 and macOS 15.5, according to a recent Bloomberg report. Apple has also provided details in a blog post on its Machine Learning Research website.

Currently, Apple relies on synthetic data, artificially generated information rather than data from actual users - to train various AI features, such as writing assistants and email summaries. While this approach ensures privacy, Apple acknowledges that synthetic data can only go so far, especially when trying to capture trends in the way people write or summarise longer messages.

New Privacy-Focused Training Method

To address this limitation, Apple has developed a method that involves comparing synthetic data to real user data, without ever accessing the actual content of user emails. Here's how it works: Apple generates thousands of fake emails covering various everyday topics, such as a message asking to play tennis at a specific time. Each email is converted into an embedding—a data representation capturing its content, like its topic and length.

These embeddings are then sent to a limited number of devices whose users have opted into Apple's Device Analytics program. These devices compare the synthetic embeddings to a sample of the user's recent emails, selecting the most similar synthetic message. Crucially, Apple assures that the actual emails and their matching results remain entirely on the device.

Differential Privacy Ensures Data Protection

The company uses a privacy technique called differential privacy, ensuring that only anonymous signals are sent back from the device. Apple then analyses the frequency with which certain synthetic messages are selected, without identifying which devices chose which data. This anonymous information helps refine AI features by making them more aligned with the types of content people write, all while preserving user privacy.

Apple has already applied this method to other features, such as Genmoji, its custom emoji tool. In this case, Apple tracks which emoji prompts are popular, like “an elephant in a chef's hat,” to improve its AI model. However, rare or unique prompts are excluded from analysis, ensuring that user-specific data is never tied to any individual device.

In the future, Apple plans to expand this privacy-focused technique to more AI tools, including features like Image Playground, Image Wand, and Visual Intelligence, all while continuing to prioritise user privacy.