(Credit: Mojahid Mottakin/Shutterstock)
Apple appears to be willing (or demanding) that Apple Intelligence not make up false information by explicitly telling it not to hallucinate, according to recently discovered code strings.
One of the new features coming to iOS 18 and macOS Sequoia with Apple Intelligence is Smart Reply in Apple Mail. The feature tries to intelligently extract questions and context from emails and then provide recipients with relevant, or smart, replies to those items.
A user on Reddit, verified by The Verge, found prompts that mention hallucination and try to steer the AI away from giving Mail users incorrect information. One of the prompts was direct and explicit: "Do not make up factual information."
In June, after Apple Intelligence was first announced at WWDC, Apple CEO Tim Cook told The Washington Post that there may be times of unreliability and the ability for it to hallucinate. "I am confident it will be very high quality. But I’d say, in all honesty, that’s short of 100%. I would never claim that it’s 100%," he said.
It’s unclear if merely telling an AI chatbot not to lie is enough to keep it from making up false information, but it seems unlikely. Possibly one of the most fascinating elements of the exposed prompts might be seeing just how much encouragement artificial personalities need. Maybe they're not unlike humans after all.
Apple Intelligence arrives this fall, though we may not see it until iOS 18.1.