Apple outlines its plans to enhance its AI models by privately analyzing user data
Apple enhances AI with user data, synthetic data, and differential privacy.

In the face of recent criticism over its AI products' performance, particularly in the area of notification summaries, Apple has unveiled its strategy to refine AI models by leveraging user data privately. This initiative involves a privacy-focused technique named 'differential privacy', which aims to merge user data analytics with rigorous privacy measures. By first creating synthetic data, Apple can simulate user data properties without utilizing actual user-generated content.
To construct a representative dataset of synthetic emails, Apple begins with generating a wide array of synthetic messages across diverse topics. From each message, the company derives an 'embedding'—a format capturing essential dimensions like language, topic, and length. These embeddings are then sent selectively to user devices that have consented to participate in Device Analytics.
The main objective of this initiative is to allow devices to compare these embeddings against actual email samples to determine which are most accurate. Hence, this method, in turn, improves the efficacy and accuracy of Apple's AI models. Currently, Apple plans to deploy this technique primarily on Genmoji models in addition to further plans of improving tools such as Image Playground, Image Wand, and Writing Tools.
By polling users who agree to share analytics with synthetic data, Apple seeks to advance the performance of their email summary features. This first phase marks a significant step in Apple's commitment to technology that balances innovation with privacy.
The use of synthetic data ensures that Apple's research balances user experience enhancement with maximum privacy. Future implications of this project may influence broader capabilities within Apple's software offerings.
Sources: Apple Newsroom, TechCrunch, Wired