Apple has introduced a new AI training method that aims to improve its AI models while maintaining user privacy, a move that could enhance its AI text outputs, such as email summaries.
Apple’s new AI training method involves comparing synthetically generated data with samples of real-world data taken from users who have opted into the Device Analytics program. Devices will compare synthetic inputs to samples of recent emails or messages, determining which synthetic data points are most similar to the real-world samples. The device then sends a “signal” to Apple, indicating only which synthetic variant is closest to the real data, without sending the actual user data. This approach ensures that Apple does not access user data, and the data never leaves the device.
Apple will use these signals to identify the most frequently picked synthetic samples. The company will then use these frequently picked “fake” samples to enhance its AI text outputs. Historically, Apple has solely trained its AI models on synthetic data, which, according to Bloomberg’s Mark Gurman, might result in suboptimal model training compared to using real-world data.
Apple has experienced issues implementing its “Apple Intelligence” features, including feature delays and leadership changes within the Siri team. The new AI training system is being introduced in beta versions of iOS and iPadOS 18.5 and macOS 15.5. This move is part of Apple’s efforts to turn things around and improve its AI capabilities.
Apple has been using a method called differential privacy since iOS 10 in 2016, which involves adding randomized information into the broader dataset to prevent linking data to any individual. This method is also applied to the company’s new AI training plans, ensuring that user data remains private. Apple has already used differential privacy to improve the AI-powered Genmoji feature.
