After facing criticism about the underperformance of its AI products, particularly in areas like notification summaries, Apple announced on Monday its strategy to enhance its AI models by privately analyzing user data with the help of synthetic data.
The company explained that it would utilize a method known as “differential privacy.” In this process, synthetic data is generated first, and then Apple reaches out to users’ devices—if they’ve agreed to share analytics data—with portions of the synthetic data. These devices compare the synthetic data with real data to assess the accuracy of Apple’s models, allowing the company to make improvements.
“Synthetic data is designed to resemble the structure and key features of real user data, but it doesn’t include any actual content from users,” Apple stated in a blog post. “To create a diverse set of synthetic emails, we begin by generating a large collection of synthetic messages on various topics. We then produce an embedding, a representation that captures the key characteristics of each message, such as language, topic, and length.”
The company added that these embeddings are sent to a select group of devices whose users have opted into Device Analytics. The devices compare the synthetic data with actual emails, helping Apple identify the most accurate embeddings.
Apple mentioned that this method is being used to improve its Genmoji models and plans to extend it to other areas, including Image Playground, Image Wand, Memories Creation, Writing Tools, and Visual Intelligence. Additionally, Apple will collect feedback from users who have opted into sharing analytics to enhance its email summary features.
