The inner workings of large language AI models can feel a little mysterious at the best of times. The learning and limitations that make the tech work are often closely guarded secrets (although leaks have revealed a a few troubling practices), but beta testers of the upcoming Apple Intelligence AI for iOS and macOS have discovered a bunch of text prompts designed to keep the tech behaving properly.
A Reddit user has spotted a plain text file containing some backend prompts for Apple Intelligence – and it’s both fascinating and somewhat hilarious to read what is essentially a motivating pep talk designed to keep the AI in line, and prevent it from getting up to any funny business on the iPhone 16 and beyond.
“Do not hallucinate,” “Do not make up factual information,” “You are an expert at summarizing posts,” “You are a helpful mail assistant,” read some of the commands. When reading the list, I half expected to find one that said, “You are a strong, independent AI.” Are these the kinds of positive affirmations AI needs to hear in order to get up in the morning?
Apple Intelligence is due to arrive later this year (Image credit: Apple)
Some users are questioning how these instructions, written in plain english rather than any kind of code, could possibly work. “Hilarious. These prompts are far too verbose to be effective,” one Redditor comments, while another adds, “‘Do not hallucinate’ – unless they’re using plain English as some sort of code that they previously trained it on this is hilarious.”
But as previous AI debacles have proven, it’s important to keep the tech in line. From Google Gemini’s diversity fail to infamous instances of AI giving questionable advice (such as suggesting one user add glue to their pizza), we’ve seen plenty of evidence that AI can go a little off the rails.
Apple Intelligence is due to start rolling out with iOS 18 later this year. And from customisable homescreens to some juicy animations, it’s already looking like a bumper update for iOS users.
