What is Apple’s AI doing with your data?

Apple’s major announcement at its Worldwide Developers Conference introduced artificial intelligence integration into its products and a partnership with OpenAI, the maker of ChatGPT. This move has raised several questions about how Apple’s AI features will function.

The confusion is understandable. Apple is simultaneously launching its own suite of AI models while also incorporating ChatGPT into its devices and software. Naturally, users are curious about how these two systems will work together and how their personal data will be managed by both companies.

This is especially significant for Apple, a company known for its focus on security and privacy.

Here’s what we know:

**Distinguishing Apple Intelligence from ChatGPT**
Why does Apple need ChatGPT if it has its own AI? The answer lies in their distinct purposes.

Apple Intelligence, the brand for Apple’s AI tools, is designed to serve as a personal assistant. It leverages specific information about your contacts, messages, emails, events, and other personal data to make your life easier, such as finding photos, attaching the right files to emails, or prioritizing notifications.

However, Apple Intelligence lacks “world knowledge”—general information about history, current events, and more. This is where ChatGPT comes into play. Users can opt to have Siri forward questions to ChatGPT or use it to help write documents within Apple apps. This integration offers a seamless way for Apple users to access ChatGPT.

**Data Management**
Given their different purposes, the type and amount of information sent to Apple Intelligence and ChatGPT will vary.

Apple Intelligence will access a wide range of personal data, including your communications, photos, videos, and calendar events. There doesn’t appear to be a way to prevent Apple Intelligence from accessing this information aside from not using its features; Apple did not immediately respond to questions on this topic.

ChatGPT, on the other hand, won’t automatically have access to your personal details unless you choose to share them. During a demo, Apple showed Siri asking for permission before sending a prompt to ChatGPT.

As part of their agreement, OpenAI will not store prompts from Apple users or collect their IP addresses, although this changes if users log in to their existing ChatGPT accounts.

**Trusting Apple with Your Data**
Most of the time, Apple Intelligence will process AI requests directly on your device using smaller AI models, similar to how FaceID and other sensitive data are handled. This method minimizes data exposure, as your information doesn’t leave your device.

When more processing power is needed, Apple Intelligence will send your query to Apple’s cloud computing platform. Here, Apple claims a significant privacy breakthrough with its new Private Cloud Compute technology. This technology allows Apple to perform computations on sensitive data without the company, or anyone else, being able to see the data or the computation process. After processing, any user data involved is deleted.

**Training Data for AI Models**
Apple’s AI models require training, just like those from other companies. This raises questions about the data used for training. Apple states that its models are trained on licensed data and data selected for specific features, never using users’ private personal data. Apple also applies filters to remove personally identifiable information from public internet data.

Apple has admitted to using public internet data to train its models, similar to other AI companies. This practice has sparked debates and lawsuits over the use of such data. Apple allows publishers to block its web crawler from collecting data from their sites, but this places the responsibility on the publishers to protect their intellectual property.

In summary, Apple’s integration of AI involves both its proprietary tools and ChatGPT, each serving different functions. Apple has emphasized privacy and security in handling user data, employing on-device processing and new cloud computing methods to protect user information. However, the training of its AI models using public internet data remains a contentious issue.