Here's how Apple's AI model tries to keep your data private
What are the main features of Apple Intelligence as revealed at WWDC?
Apple Intelligence, revealed at WWDC, is a suite of features that brings generative AI tools to iPhone, iPad, and Mac. The main features include rewriting email drafts, summarizing notifications, and creating custom emoji. Apple Intelligence is built with an on-device philosophy that ensures privacy by keeping data local for common AI tasks such as transcribing calls and organizing schedules. For more complex AI requests, the system can reach out to cloud servers while maintaining data privacy. The integration of Apple Intelligence into Apple's operating systems and apps is a major selling point, as well as the optimization of models for power efficiency and size to fit on iPhones.
What is unique about Apple's approach to training its AI models compared to other companies?
Apple's approach to training its AI models is unique in that it does not train its models with private data or user interactions. Instead, the company uses a combination of licensed materials and publicly available online data scraped by its Applebot web crawler. Publishers must opt out if they don't want their data ingested by Apple. The company also omits feeding social security and credit card numbers found online, as well as ignores "profanity and other low-quality content." This approach sets Apple apart from other companies in terms of training AI models with a focus on user privacy.
How does Apple ensure the privacy of user data with its new AI system, Apple Intelligence?
Apple ensures the privacy of user data with its new AI system, Apple Intelligence, through a twofold approach. First, the system is built with an on-device philosophy, which means that common AI tasks are processed locally on the user's device, keeping data private. Second, for more complex AI requests that require the use of cloud servers, Apple has created Private Cloud Compute (PCC) servers that are designed to protect user privacy4.
When a user request requires a more capable AI model, it is sent to the PCC servers. These servers have their own secure boot and Secure Enclave to hold encryption keys that only work with the requesting device. The Trusted Execution Monitor ensures that only signed and verified code runs on the PCC servers. Apple says that the user's device creates an end-to-end encrypted connection to a PCC cluster before sending the request. Apple claims it cannot access data in the PCC since it is stripped of server management tools, meaning there is no remote shell. Furthermore, Apple does not provide the PCC with any persistent storage, so requests and possible personal context data pulled from Apple Intelligence's Semantic Index are deleted in the cloud afterward.
Each build of the PCC will have a virtual build that the public or researchers can inspect, and only signed builds that are logged as inspected will go into production. This ensures transparency and allows security researchers to verify the security and privacy guarantees of the system.
In addition to these measures, Apple's revamped Siri can send some queries to ChatGPT in the cloud, but only with the user's permission. This shifts the privacy question into the hands of OpenAI, which has its own privacy policies, and the user, who must agree to offload their query. Apple CEO Tim Cook said that ChatGPT would be called upon for requests involving "world knowledge" that are "out of domain of personal context."
Overall, Apple's approach to privacy in its Apple Intelligence system is based on a combination of on-device processing, secure cloud servers, and transparency measures that allow for independent verification of the system's security and privacy guarantees.