How Apple Plans to Improve Its Artificial Intelligence Models While Protecting User Privacy
Even Apple’s most loyal users admit that the company is behind the curve when it comes to artificial intelligence, with Apple Intelligence’s big update to Siri now officially delayed (which had been heavily touted for the past year). In a new blog post, Apple outlines some of the ways it hopes to get back on track.
One potential reason for Apple’s problems with generative artificial intelligence may be that the company pays much more attention to user privacy than OpenAI and Google. Apple doesn’t collect any user data to train its large language models or LLMs (though it has trained its models on free text on the Internet) and relies heavily on synthetic data to generate AI text from suggestions and existing texts.
The problem with synthetic data is that it is artificial. It lacks the nuance and variation of human writing as it changes over time, and without text written by real people for comparison, it’s difficult to gauge the quality of what AI outputs. As mentioned in the blog post, Apple is now planning to improve text generation.
In general terms, the idea is that AI-generated synthetic text will be compared to a selection of users’ actual text stored on Apple devices, but with multiple layers of protection to prevent individual users from being identified or any personal correspondence being sent to Apple. Essentially, this approach scores synthetic text by comparing it to real writing samples, but only aggregate scores are returned to Apple.
In fact, what’s actually happening has nothing to do with actual words or sentences at all: both synthetic text and handwritten writing are converted into “embeddings,” which are essentially mathematical representations of the text. There is enough data to evaluate the quality of AI text without reaching the level of real reading.
All this information is encrypted in transit, and comparisons are only made on devices where users have enabled device analytics on their gadgets (for example, this option can be found under Privacy & Security > Analytics & Improvements in Settings on iOS). Apple never knows which AI text sample was selected by an individual device, only the samples that have the best rating among all devices tested.
Genmoji and other tools
According to Apple, this anonymous rating system can be used to improve text created or rewritten using generative AI models, and should also mean more accurate and intelligent writing . The highest rated results may be modified with a different word or two before being sent back for the next round of grading.
A simpler version of the same approach is already being used by Apple to implement its Genmoji AI feature , which lets you magically impersonate an octopus on a surfboard or a cowboy wearing headphones. Apple aggregates data from multiple devices to see which searches are popular, and also applies security measures to ensure that unique individual searches are not visible or tied to specific users or devices.
Again, this only happens on iPhones, iPads, and Macs that have device analytics enabled. By making devices report “noisy” signals without any specific user information, Apple can improve its AI models based on aggregate data without users having to worry about their Genmoji suggestions being detected.
Apple says similar techniques will soon be used in other Apple Intelligence features. These features will include Image Playground, Image Wand, Memories Creation, Write Tools, and Visual Intelligence—all of which were among the first Apple AI capabilities to actually be available on devices.
According to Bloomberg , the new and improved systems will be tested in the upcoming beta versions of iOS 18.5, iPadOS 18.5 and macOS 15.5. We may well hear more about them, and about Apple Intelligence in general, at this year’s Apple Worldwide Developers Conference, which starts on Monday, June 9 .
Meanwhile, Apple’s AI rivals show no signs of slowing down—and they have fewer qualms about using text written by their users to further train their AI models. In recent days, we’ve seen Microsoft release a number of updates to Copilot (including Copilot Vision and file search ), Google adding video generation to Gemini , and OpenAI updating ChatGPT’s memory capabilities .