According to a new report by Bloomberg's Mark Gurman, Apple is quietly shifting gears on its AI game, and it hopes to do it in a way that stays true to its privacy-first philosophy. The company now plans to analyze user data directly on devices to sharpen its AI models without ever sending that data back to its servers.
So what does that mean in plain English? Well, instead of relying mostly on synthetic data (which is basically fake but realistic-looking text created by Apple), the company is going to start checking that synthetic data against snippets of real-world items like recent emails stored right on your iPhone, iPad, or Mac.
Apple explained it like this:
When creating synthetic data, our goal is to produce synthetic sentences or emails that are similar enough in topic or style to the real thing to help improve our models for summarization, but without Apple collecting emails from the device.
Apple's whole AI platform, dubbed Apple Intelligence, has been dragging behind the likes of OpenAI, Microsoft, and Google for a while now. A big part of the reason was that it was trying to build next-gen AI tools using data that wasn't quite up to snuff. Sure, synthetic data helps dodge privacy issues, but when that data doesn't feel like actual user interactions, the results can get clunky.
Case in point: Apple's writing and summarization tools have been known to fumble. Notifications don't always make sense. Summaries are often off-mark. And as for Siri, that's a story on its own. Internal tests showed the updated Siri failed one out of every three tasks. That was enough to spark a shake-up in leadership, delays in release timelines, and a general feeling of "we're not quite ready yet."
So now, with this new system baked into the upcoming beta versions of iOS 18.5, iPadOS 18.5, and macOS 15.5, Apple's hoping to clean things up.
The idea is to let the system peek at what kind of emails you're actually dealing with without reading or storing them elsewhere and use that to calibrate the AI's synthetic training data. That could improve everything from message summaries to writing suggestions.
Apple's also using this approach to improve features like Image Playground and Memories Creation.
For things like Genmoji, the company is relying on differential privacy, a system that helps identify trends across users without exposing individual behavior. Gurman noted that:
The idea is to track how the model responds in situations where multiple users have made the same request, say, asking for a dinosaur carrying a briefcase, and improving the results in those cases.
Importantly, these enhancements will only kick in for users who have opted into device analytics and product improvement settings. You can toggle those in the Privacy and Security tab on your device if you're curious.
1 Comment - Add comment