Apple’s new AI upgrade gets smarter without reading your data

If you’ve been following Apple’s journey into the AI space, you’d know they’ve been treading with extra caution, prioritising privacy even when it slows things down. Now, the Cupertino giant is taking a significant step forward by refining its Apple Intelligence features using a novel combination of synthetic data and on-device analytics. The best part? User privacy remains untouched.

Apple has always touted privacy as a fundamental human right, and its latest innovation builds right on top of that belief. With competitors like OpenAI and Google pushing the boundaries of generative AI using vast datasets – many of which come from real user inputs – Apple had to find a way to compete without crossing its ethical lines. The result is a new way of analysing real-world usage trends without actually collecting personal data.

ALSO READ: How Apple Intelligence can transform our daily digital lives

You might ask, why are we talking about this, now? Well, Apple recently published a blog post, which was first reported on by Bloomberg. There, Tim Cook and Co. detailed how they plan on using the data generated by their users to train their AI. And, this sparked a wave of discussions, which we are here to give our two cents on.

Differential privacy meets AI

It all started with Genmoji, a feature that lets users create emojis from their own prompts. Apple began using differential privacy here. It was a method that lets it identify popular prompt patterns without ever seeing who typed what. The system works by having opted-in devices anonymously send “noisy” signals. Essentially, each device may respond truthfully or randomly, making it statistically impossible for Apple to track a specific user’s input.

For example, if multiple users typed “penguin in a top hat,” Apple’s system would only recognise the frequency of this phrase if enough people used it, and without knowing who did. That’s privacy-first innovation in action.

Now, Apple’s taking the same approach to features like Image Playground, Image Wand, Memories Creation, and Writing Tools. But when it comes to understanding longer texts, like email content for summarisation tools, Apple had to switch gears.

Enter synthetic data – With a twist

Here’s where things get interesting. Synthetic data isn’t new. Apple’s been using it to train its AI for a while. The challenge? It doesn’t always reflect how people actually write or talk. To bridge that gap, Apple now checks its synthetic data against real user data, but only on-device.

So how does this work? Apple creates a bunch of synthetic messages, like “Want to play tennis at 11:30?” Then, your device (if you’ve opted in) checks which synthetic messages most closely resemble your recent emails. Apple doesn’t get your emails. Not even fragments. Instead, your device just sends back a signal indicating which synthetic email felt closest to your writing, and even that is sent with differential privacy built in.

These anonymous signals help Apple refine which synthetic messages are most representative of real-world communication. From there, the company can improve how features like email summaries or writing suggestions function.

With this improved method, Apple is gearing up to roll out enhanced Apple Intelligence features in beta versions of iOS 18.5, iPadOS 18.5, and macOS 15.5. The second beta is already in developers’ hands.

Apple’s approach might not be the fastest route to cutting-edge AI, but it certainly is one of the most user-respecting one. While the rest of the AI world races ahead with aggressive data collection, Apple is betting on smart engineering and years of privacy research to carve its own path.

Privacy still at the core

It’s important to note that none of these improvements affect users who haven’t opted into Apple’s Device Analytics. And even for those who have, no content – not even anonymised content – ever leaves their device. All Apple sees are aggregated insights, without a single user’s identity in the mix.

What do you think about this move? Let us know in the comments below.