Giannandrea said Apple focused on reducing hallucinations in its models, in part by using curated data. “We put a lot of energy into training these models very thoroughly,” he said. “So we can be sure that we are using this technology responsibly.”
The training wheels approach to artificial intelligence applies across Apple’s portfolio. If it works as promised, it should mean that Apple Intelligence is less likely to fabricate or suggest something inappropriate. In its blog post, Apple said testers were more likely to find its models more useful and less harmful than competing models on devices from OpenAI, Microsoft and Google. “We don’t take this teenager and tell him to go on a plane,” Federighi said.
Apple’s hotly anticipated tie-up with OpenAI will keep ChatGPT at arm’s length, with Siri and a recent writing assistant called Writing Tools only touching it for certain hard queries and with user consent. “We’ll ask you before you go to ChatGPT,” Federighi said. “From a privacy standpoint, you’re always in control and you have complete transparency when you leave Apple’s privacy sphere and use a different model.”
An Apple deal with OpenAI once seemed highly unlikely. The startup has seen explosive growth thanks to the brilliance of its chatbot, but has repeatedly courted controversy in the form of legal battles, boardroom drama, and relentless promotion of its powerful but unreliable technology. Federighi said Apple may include a Google Gemini flagship model in the future, without providing further details.
Apple has been ridiculed for being slower than its competitors in developing generative AI and has yet to reveal anything as powerful as OpenAI’s ChatGPT or Google’s Gemini, but the company has published some noteworthy AI research, including details about the company’s multimodal models running on devices .
Apple once seemed to be at the forefront of using artificial intelligence in personal computers, after launching Siri in 2011. The assistant used the latest breakthroughs in artificial intelligence to better recognize speech and tried to turn a confined range of voice commands into useful actions on the iPhone.
Competitors such as Amazon, Google and Microsoft soon followed suit with their own voice assistants, but their usefulness was fundamentally confined by the challenge of parsing the meaning of convoluted and ambiguous language. The huge language models that underpin programs like ChatGPT represent a significant advance in the ability of machines to handle language, and Apple and other companies hope to exploit artificial intelligence to improve their personal assistants in many ways. LLMs could make helpers like Siri better able to understand convoluted commands and conduct relatively sophisticated conversations. They could also enable assistants to exploit the software by writing code on the fly.
“They are committed to personal, private and contextual AI,” says Tom Gruber, an artificial intelligence entrepreneur who co-founded the company that developed Siri, which was acquired by Apple in 2010. Gruber says he was ecstatic to see the company’s demo exploit cases highlighting these features.
Other observers say Apple’s announcements are an attempt to keep up with the competition without making too many gaffes. “Apple is great at offering great new capabilities and showing us new ways of doing things,” says David Yoffie, a professor at Harvard Business School. “None of the things that have been announced seem like that, which is not surprising as they are trying to catch up.”
Yoffie says Apple’s focus on data privacy and security wasn’t surprising, given people’s concerns about sharing data with programs like ChatGPT. “Generative AI is a complement to the iPhone,” he says. “I think it’s important for them to show that they’re not behind the Android world, which I think they did today.”
Still, generative AI is by definition unpredictable. Apple Intelligence may have behaved during testing, but there’s no way to account for all the results when they’re rolled out to millions of iOS and macOS users. To deliver on the promises made at WWDC, Apple will need to empower AI with a feature no one else has managed yet. You have to make him behave.