{"id":53662,"date":"2025-04-14T22:48:05","date_gmt":"2025-04-14T20:48:05","guid":{"rendered":"https:\/\/www.apfelpatient.de\/?p=53662"},"modified":"2025-05-12T00:45:16","modified_gmt":"2025-05-11T22:45:16","slug":"how-apple-improves-apple-intelligence-despite-data-protection","status":"publish","type":"post","link":"https:\/\/www.apfelpatient.de\/en\/allgemein\/so-verbessert-apple-die-apple-intelligence-trotz-datenschutz","title":{"rendered":"How Apple improves Apple Intelligence \u2013 despite data protection"},"content":{"rendered":"

Apple Intelligence is designed to make many features on iPhone, iPad, and Mac smarter. These include Siri, email summaries, Genmoji, writing tools, and image functions. To further develop these features, Apple needs data. At the same time, the company has a clear focus: protecting privacy. This presents Apple with a challenge \u2013 training AI models is difficult without user data. Nevertheless, Apple is working on solutions to further improve Apple Intelligence without collecting personal information.<\/strong><\/p>\n\n\n\n

Large language models like those used by Apple require a lot of data to function effectively. Other companies collect extensive information about user behavior for this purpose. Apple is taking a different approach: The company wants to improve the quality of its AI models without storing or analyzing personal data. A blog post<\/a> by the Apple Machine Learning Research Team has now described some methods it hopes to use to achieve this.<\/p>\n\n\n\n

Why Apple relies on data<\/h4>\n\n\n\n

Apple Intelligence relies on usage data to improve, including information about how users write emails or use Genmojis. But Apple doesn't want to collect real-world content that can be traced back to specific individuals. Therefore, the company relies on alternative methods such as synthetic data, local on-device processing, and differential privacy.<\/p>\n\n\n\n

How Apple works with synthetic emails<\/h4>\n\n\n\n

For example, Apple creates synthetic emails that resemble real messages. These can be emails with typical content, such as an invitation to play tennis at 3 p.m. These synthetic emails are converted into so-called embeds. These contain information such as the language, subject, and length of the message. Apple creates multiple versions with different lengths and wording. These embeds are sent to a small group of users who have consented to receiving device analytics. These users' iPhones compare the synthetic embeds locally with real emails on the device. This calculates which synthetic email is most similar to the real message.<\/p>\n\n\n\n

\"Apple
Image: Apple<\/figcaption><\/figure>\n\n\n\n

The device makes this decision independently, without Apple having access to the real content. Differential privacy is then used to determine which synthetic email was most frequently perceived as similar within the comparison group. This information is reported back to Apple anonymously. Neither Apple nor anyone else can track which device made which selection.<\/p>\n\n\n\n

What synthetic embeddings are used for<\/h4>\n\n\n\n

Apple can use the synthetic embeddings that prove to be particularly suitable in the comparison as training or test data. They also serve to further improve the quality of the synthetic content. This allows Apple to specifically train its own AI models without having to access sensitive user data.<\/p>\n\n\n\n

How it works with Genmoji<\/h4>\n\n\n\n

A similar approach is used with Genmoji \u2013 the creation of emojis based on text input. Here, too, Apple uses differentiated privacy to determine which inputs are used most frequently. The system only considers prompts that have been formulated similarly by many users. Content that is too specific or unique is filtered out. Apple receives no information about which prompts originate from which user. All data is anonymized, and random code is also incorporated to prevent tracking. There is no link to IP addresses or Apple IDs. Apple also does not see the generated Genmojis themselves.<\/p>\n\n\n\n

Participation is voluntary<\/h4>\n\n\n\n

It's important to note that only users who have explicitly agreed to share so-called device analytics with Apple in their settings participate in these processes. Users who do not want this can deactivate this option at any time. Then, no anonymized data from the device will be used for these analyses.<\/p>\n\n\n\n

Future application in additional functions<\/h4>\n\n\n\n

Apple plans to expand these privacy-enhancing practices to other areas. In the upcoming versions of iOS 18.5, iPadOS 18.5, and macOS Sequoia 15.5, they will also be used in the following features:<\/p>\n\n\n\n