Apple to enhance AI using user data, assures no compromise on privacy

Currently, Apple uses synthetic data that mimic the structure and characteristics of user-generated content to improve Apple Intelligence-based features

Apple has outlined how it plans to use data from participating users to improve its AI models—without compromising individual privacy. In a recent research blog published by the US-based technology company, it explained how it will gather “usage trends” and “aggregated insights” to refine Apple Intelligence features, while ensuring that no personal content or unique user behaviour is revealed.

 

Currently, Apple uses synthetic data—artificially generated examples that resemble real user data—to train its AI systems. These synthetic samples mimic the structure and characteristics of actual data but do not include any user-generated content. However, Apple acknowledges that synthetic data alone can limit how well AI understands real-world trends and context, since it is randomly generated.

 

 

How Apple will use user data

 

To make synthetic data more useful, Apple plans to compare large sets of synthetic messages with real user data—but only on the user’s device. This will help identify patterns and improve the relevance of training data, without ever seeing or storing the real content itself.

 

For example, to train summarisation or writing tools, Apple will generate a wide range of synthetic messages such as: “Would you like to play tennis tomorrow at 11:30AM?”

 

These messages are turned into embeddings—mathematical representations that capture details like language, topic, and message length. These embeddings are then sent to a small number of devices from users who have opted in to Apple’s Device Analytics.

 

Each of these devices will then compare the synthetic messages with a small sample of the user’s recent emails, on-device. The device identifies which synthetic messages are most similar to the user's data, and only shares which types of messages were commonly selected—not the content itself. Apple then uses this information to fine-tune the quality of synthetic data for future training.

What this means for Apple Intelligence features

 

Apple says this technique helps improve features like email summarisation by enhancing the realism and variety of the synthetic training messages. A similar process is already being used for Genmoji, Apple’s AI tool that generates custom emojis based on text prompts. By analysing which synthetic prompts are most similar to user-generated ones, Apple can improve Genmoji’s accuracy and creativity.

 

Apple calls this approach “differential privacy” and will soon apply it to more Apple Intelligence tools—including Image Playground, Image Wand, Memories creation, Writing Tools, and Visual Intelligence features.