Apple aims to generate synthetic data based on user data, provided they give permission.
Apple has announced in a statement that it wants to improve its AI models (without ChatGPT support), for example for summarizing emails or generating text and images. The company does this by keeping user data private if the user consents to device analysis.
Differential Privacy
The technique Apple uses is called ‘differential privacy’ and is used in text generation and email summaries, among other things. For example, Apple generates synthetic emails that contain the formatting and properties of real messages. Never personal data or real emails from the user themselves.
These emails are then sent to the iPhones and iPads of people who have given permission for analysis. These devices compare the synthetic emails with your real emails locally, without Apple having access to your mailbox. The device then sends a notification to Apple so it knows which emails most resemble the real ones. Apple uses this feedback to better train its AI.
Recently, the beta version of Apple Intelligence, Apple’s suite of AI functions in iOS, has become available in Europe. In addition to text generation and email summaries, this approach is also used to improve the Genmoji models. Soon, Image Playground, Image Wand, Memories Creation, and Writing Tools will also be improved in this way.
read also