Apple’s big plan for better AI is you

Apple has a brand new plan to make Apple Intelligence smarter. That plan is you — more specifically, your data.

A post on the company’s Machine Learning blog explains the plan to begin on-device analysis of your user data. Apple trains its own AI using synthetic data, but this is a limited approach because models trained this way can struggle to understand trends when supporting Writing Tools or summaries. The idea is that by combining your results data with this synthetic data Apple can build better models.

The way this works is interesting, as rather than grabbing all your information and uploading it to the cloud, Apple will analyze your information on the device, protect it using a technique called differential privacy, and share what data it does take in a form that does not directly connect to you. Apple isn’t taking everything, either – instead it’s just collecting information to show it how successfully its AI responds to enquiries (poll data). This will be in support of Writing Tools, Image Playground, Image Wand, Memories Creation, and Genmoji, and may be extended to text generation — surely part of the company’s big push to make better AI.

All your data R belongs to?

What’s being proposed is innocuously described.The idea is that Apple will combine this polling data with its own synthetically produced data to identify ways to optimize its own information and deliver better results. It’s basically a tool to ensure the information it uses to drive its AI isn’t garbage in order to make it far more likely that the answers/responses it generates aren’t rubbish either.

Not only is the whole system designed to maintain privacy, but people must opt into the Device Analytics program for this to happen. This is on by default. When enabled, the device will be able to figure out which inputs are closest to the real samples, enabling the company to identify ways in which to improve AI text outputs, including things like email summaries.

“These techniques allow Apple to understand overall trends, without learning information about any individual, like what prompts they use or the content of their emails,” the post said.

Your data is protected by differential privacy, which the company says is already being used to improve Genmojis; it means the company doesn’t know who you are, gains no access to your actual data and doesn’t see any contextual information. It just gets aggregated insights into which responses tend toward being the most appropriate in the real world.

“These most-frequently selected synthetic embeddings can then be used to generate training or testing data, or we can run additional curation steps to further refine the dataset,” Apple said.

Which end of the wedge?

I expect the broader effort is aimed at something more significant, and certainly more significant than the Genmoji Apple’s machine intelligence team spends so much time discussing in the latest report. It’s pretty clear that Genmoji just represents a cuddly face for machine-driven AI — Apple is training its devices to understand mood and context, which when combined with everything else your devices know about you looks increasingly like the creation of digital twins.

While the implementation seems to be currently focused on trivial things like email summaries, it’s pretty clear the system might help it in other domains, such as health. Your health data, for example, could be analyzed on device with only anonymized info shared with advanced medical AI in the cloud to support the AI-driven healthcare system the company is thought to be prepping. (Some reporters note that the system seems similar to the CSAM system Apple once considered; it was used to monitor your behavior, while this new attempt aims to augment it. The CSAM system was arguably more open to abuse.)

You can also turn this new AI system off:

On iOS: Settings > Privacy & Security > Analytics & Improvements, toggle Share iPhone Analytics to off.

On Mac: Settings > Privacy & Security > Analytics & Improvements, turn Share Mac Analytics off.

The future will be better tomorrow

If you close your eyes, put your fingers in your ears, and sing really loud you can perhaps pretend you don’t see the connection between rogue government’s such as the current UK administration demanding back doors into data security and news that Apple plans to analyze your data to train Apple Intelligence. But if you’ve read the same self-help manuals (1984, Brave New World, Ruling the Void, V For Vendetta) as the current crop, it’s hard to unsee the connection. 

Of course, these insights will most certainly generate major benefits — better health, higher productivity, economic growth — but all of those positive benefits can be undone if privacy and security are not universally protected. Differential privacy and data encryption show that Apple is trying to do something that combines the potential benefits of contextual AI while protecting your privacy. Sadly, not every company seems faithful to the same path. 

Ultimately, the real message here is that, by its actions, Apple illustrates the reality that AI development firms are hungry for all the data they can consume to optimize their models. That demand makes it inevitable that privacy wars concerning your rights to maintain control of your own information aren’t a series of skirmishes to look forward to tomorrow, they are already here. 

What happens if?

This raises plenty of questions: What happens when, for example, a state demands (and eventually gets) access to all your data, including encrypted data? What happens when that data stack leaks as they all do? What happens when other governments, terrorists, criminals get their hands on the same information? Will differential privacy be enough protection? What happens next?

You can follow me on social media! Join me on BlueSky,  LinkedIn, and Mastodon.