Apple is preparing to implement a new system that will analyse user data directly on its devices, aiming to strengthen AI capabilities without compromising privacy, Bloomberg has reported.
The move will mark a shift from the company’s prior exclusive reliance on synthetic data for training AI models, which can sometimes fall short of accurately representing real-world user behaviour.
Go deeper with GlobalData
The upcoming change targets limitations in Apple’s current approach, where synthetic data—designed to mimic real inputs without containing personal information—fails to fully capture the complexity of actual user interactions.
The new system, by contrast, will use anonymised samples of user emails from Apple’s Mail app on iPhone, iPad, and Mac to validate the quality of synthetic datasets.
Apple said this process will help it refine features within its Apple Intelligence platform, such as notification summaries, Writing Tools synthesis, and message recaps, without transmitting personal data off the device.
In a blog post, the company said: “When creating synthetic data, our goal is to produce synthetic sentences or emails that are similar enough in topic or style to the real thing to help improve our models for summarisation, but without Apple collecting emails from the device.”
Apple Intelligence, launched in 2024, is built on large language models, a foundational technology in modern AI.
In addition to synthetic data, Apple has used licensed third-party data and information from publicly available online sources to train these models.
However, the limitations of synthetic data have reportedly led to shortcomings in the system’s performance, including inaccurate summaries and misrepresented content in notifications.
The new data analysis system will be introduced in beta versions of iOS and iPadOS 18.5 and macOS 15.5.
Apple also confirmed it is expanding its use of privacy-preserving techniques to enhance other features within Apple Intelligence.
These include Image Playground, Image Wand, Memories Creation, and Visual Intelligence.
One such method, differential privacy, has already been employed to improve the Genmoji feature, which allows users to create custom emojis.
Differential privacy enables the identification of common user prompts while mathematically safeguarding individual data points.
