Apple has unveiled a novel AI training method designed to bolster its AI models while safeguarding user privacy, potentially leading to enhanced AI-generated text outputs, such as email summaries.
The tech giant’s new approach entails comparing synthetically generated data with real-world data samples from users who have opted into the Device Analytics program. On the device level, synthetic inputs are juxtaposed with samples of recent emails or messages to determine which synthetic data points bear the closest resemblance to the actual user data. Subsequently, the device transmits a “signal” to Apple, indicating the synthetic variant that most closely aligns with the real data, without conveying the actual user data itself. This methodology ensures that Apple remains unaware of user data, as the data never leaves the device.
These signals will be utilized by Apple to identify the most frequently selected synthetic samples. The company will then leverage these frequently chosen “fake” samples to refine its AI text outputs. Historically, Apple has relied solely on synthetic data to train its AI models, a practice that, according to Bloomberg’s Mark Gurman, may result in suboptimal model training compared to utilizing real-world data.
Apple has faced challenges in implementing its “Apple Intelligence” features, including delayed feature rollouts and leadership changes within the Siri team. The new AI training system is being rolled out in beta versions of iOS and iPadOS 18.5 and macOS 15.5, as part of the company’s efforts to revamp and enhance its AI capabilities.
Since iOS 10 in 2016, Apple has employed a method called differential privacy, which involves incorporating randomized information into the broader dataset to prevent linking data to any individual. This method is also being applied to the company’s new AI training plans, thereby ensuring the privacy of user data. Apple has already utilized differential privacy to improve the AI-powered Genmoji feature.
