Apple has announced a new approach to training its artificial intelligence models, designed to protect user privacy while improving the accuracy of features like email summaries and writing tools. The company is shifting away from relying solely on synthetic data, which can struggle to understand complex trends in long sentences or entire emails.
To address this limitation, Apple will use a novel technology that compares synthetic data to a small sample of recent user emails. This method allows Apple to learn overall trends without compromising individual user privacy. Here’s how it works:
Apple creates a large set of synthetic messages on various topics and derives a representation, called an embedding, of each message. These embeddings are then sent to a select group of devices that have opted in to Device Analytics.
Participating devices choose a small sample of their own recent emails and compute the embeddings for these samples. Apple then uses differential privacy to analyze the most frequently selected synthetic embeddings across all devices.
This approach enables Apple to refine its dataset, generating new training data or testing data that better reflects real-world language patterns. By protecting user privacy while improving AI model accuracy, Apple aims to provide more effective features in future iOS and macOS updates.
Source: https://9to5mac.com/2025/04/14/apple-details-on-device-apple-intelligence-training-system-using-user-data