Apple Intelligence is instructed not to hallucinate
Apple's most recent developer betas, released last week, have brought a range of innovative generative AI features. These advancements will be integrated into iPhones, iPads, and Macs in the coming months. Initially announced at the Worldwide Developers Conference (WWDC), these features are now accessible for exploration on Apple computers. Users can also view the pre-prompt instructions embedded into the model, like "do not hallucinate" and "avoid religious, political, harmful, violent content" that support some of these Apple Intelligence features.
Instructions and backend prompts
The instructions for Apple's AI features appear as prompts that precede any user interaction with a chatbot. Similar prompts have been used in AI tools such as Microsoft Bing and DALL-E. A member of the macOS 15.1 beta subreddit recently found files containing these backend prompts on Apple's system, offering an early look at how these AI features work.
A look at 'Rewrite' tool's code
The discovered files include an AI bot for a "helpful mail assistant" that asks questions based on email content, potentially part of Apple's Smart Reply feature. Another feature is the "Rewrite" tool, part of the Writing Tools accessible by highlighting text and right-clicking or long-pressing on it. The instructions for this tool include passages such as "Please limit the answer within 50 words. Do not hallucinate. Do not make up factual information."
AI features for 'Memories' and model identification
The files also contain instructions for creating a "Memories" video with Apple Photos. These guidelines specify not to create a story that is religious, political, harmful, violent, sexual or in any way negative. The files refer to the model as "ajax," which was rumored to be the internal name for Apple's LLM last year. Instructions on how to locate these files within the macOS Sequoia 15.1 developer beta, were also shared by the person who discovered them.
