Washington
CNN
—
Apple's splashy announcement at its Worldwide Developers Conference this week that it's adding artificial intelligence to its products and partnering with OpenAI, the creators of ChatGPT, has raised a lot of questions about how Apple's AI products will work.
The confusion is understandable: Apple is releasing its own suite of AI models while simultaneously integrating ChatGPT into its devices and software, raising questions about where one ends and the other begins, and perhaps more pressingly, what the two companies will do with the personal information they receive from users.
The risks are especially great for Apple, which has made security and privacy hallmarks of its brand.
Here's what we know:
Differences between Apple Intelligence and ChatGPT
Why does Apple need ChatGPT when they have their own AI? The answer is that they each serve different purposes.
Apple Intelligence — A Collective Brand “Google Assistant,” the collective name for Apple's AI tools, functions above all as a personal assistant — with an emphasis on “personal.” The assistant gets specific information about your relationships and contacts, the messages and emails you've sent, the events you've attended, the meetings on your calendar, and other highly personal data about your life. Apple hopes it can then use that data to make your life just a little easier, by digging up photos you took at a concert years ago, finding the right attachment to add to an email, or ranking your mobile notifications by priority and urgency.
But Apple Intelligence may know that you went on a hiking trip last year, Executives call it “world knowledge” — general information that doesn't directly concern you, such as history or current events. That's where ChatGPT comes in. Users can have Siri forward questions or prompts to ChatGPT (on an opt-in basis), or let ChatGPT help them with document creation within Apple apps. Apple said it also plans to eventually integrate with other third-party AI models. This integration will essentially eliminate the steps to access ChatGPT, making it more seamless for Apple users to access its platform.
Because Apple Intelligence and ChatGPT serve very different purposes, the amount and type of information users send to each AI may also differ.
Apple Intelligence will have access to a wide range of personal data, from written communications to photos and videos you take, to calendar event records. There appears to be no way to prevent access to this information other than to not use Apple Intelligence features. An Apple spokesperson did not immediately respond to questions about the matter.
ChatGPT doesn't necessarily or automatically have access to highly personal details of its users, but if you decide to use ChatGPT through Apple, you can choose to share some of this data, as well as other data, with OpenAI. In a demo on Monday, Apple showed that Siri would ask users for permission before sending a prompt to ChatGPT.
As part of its deal with Apple, OpenAI made an important concession: it agreed not to store any prompts from Apple users and not to collect IP addresses. However, knowingly logging in and connecting to an existing ChatGPT account will ruin everything. Some users may choose to do so in order to take advantage of their ChatGPT history and the perks associated with ChatGPT paid account plans.
Now that we know what OpenAI will and won't do with your data, what about Apple?
When using ChatGPT, Apple users will need to send personal information and AI queries to OpenAI, but in most cases Apple Intelligence will not send user data anywhere, Apple said. Whenever possible, Apple will strive to use smaller AI models to process AI prompts directly on the device.
This is similar to how Apple already handles FaceID and other sensitive data: the idea is that by processing the data directly on the device, it limits risky exposure — if the data never actually goes anywhere, it can't be intercepted or hacked from a central server.
If an AI task requires more processing power, Apple Intelligence sends the query and data to an Apple-managed cloud computing platform, where a more capable AI model fulfills the request.
Here, Apple claims to have achieved a major privacy feat. It's a breakthrough, and although the company's announcement only got a relatively brief airing during Monday's packed keynote, the company is clearly proud of the progress it's been meticulously planning.
Apple on Monday announced that it has developed a new method of cloud computing that will allow the company to perform calculations on sensitive data without anyone knowing what data is being processed or what calculations are being performed, not even the company itself. Apple's new architecture, called Private Cloud Computing, borrows certain hardware and security concepts from the iPhone, such as the Secure Enclave that already protects sensitive user data on Apple's mobile devices.
With private cloud computing, “your data is never stored by or accessible to Apple,” Craig Federighi, Apple's SVP of software engineering, said during Monday's keynote. After fulfilling a user's AI request, private cloud computing deletes all user data involved in the process, Apple said.
Apple claims that private cloud computing is “possible” because it tightly controls the entire technology ecosystem, from its specialized, proprietary computer chips to the software that ties it all together.
If it's true that Apple can't see what personal data its large-scale AI models are processing (Apple claims it invited researchers to test it themselves because it intends for the system's design to be scrutinized), then Apple's implementation stands out from other companies' implementations. For example, when using ChatGPT, OpenAI has made it clear that it uses users' data to further train its AI models. With Private Cloud Compute, in theory, you don't have to take Apple's word that it doesn't use users' data for AI training.
Apple's AI models didn't just appear out of nowhere: Like models from other companies, they had to be trained, which raises questions about whose data Apple used and how.
In a technical document published this week, Apple said its models are trained “on licensed data, including data selected to power specific features.”
“We never use your private personal data or interactions when training our underlying models,” the company added, “and we apply filters to remove personally identifiable information, such as Social Security numbers or credit card numbers, that are publicly available on the internet.”
But Apple has acknowledged that it scraped the data it used to train its own models from the public internet, somewhat similar to other AI companies, some of which have faced copyright lawsuits and sparked debate over whether AI startups are unfairly profiting from human work.
Apple hasn't disclosed what web-based information it captured, although it said publishers can add code to their sites to prevent data collection by Apple's web crawlers, but that puts the onus on publishers, and not Apple, to protect their intellectual property.
