Apple has announced a new approach to enhance its AI models without relying on or copying user data from iPhones and Macs. According to a blog post highlighted by Bloomberg, the company plans to utilize a synthetic dataset compared against samples from recent emails or messages of users who have opted into its Device Analytics program.
Apple devices will analyze which synthetic inputs most closely match real samples and will communicate this by sending only "a signal indicating which of the variants is closest to the sampled data" to the company. This process ensures that user data remains unseen and confined to the device. The frequently selected synthetic samples will then be used to enhance AI-generated text outputs, like email summaries.
Currently, Apple uses only synthetic data to train its AI models, which might result in less effective responses, as noted by Bloomberg’s Mark Gurman. The company has faced challenges with some of its Apple Intelligence features, leading to delays in launching certain capabilities and changes in its Siri team leadership.
To address these issues, Apple is set to implement its new AI training system in beta versions of iOS and iPadOS 18.5, and macOS 15.5, as reported by Gurman.
Since 2016, with iOS 10, Apple has promoted its use of differential privacy to protect user data. This method, which introduces randomized information into larger datasets, will also apply to its new AI training initiatives, helping prevent data from being linked to individual users.
