/cdn.vox-cdn.com/uploads/chorus_asset/file/24401980/STK071_ACastro_apple_0003.jpg?w=696&resize=696,0&ssl=1 "“You’re a Helpful Mail Assistant” and Other Apple Intelligence Instructions")
/cdn.vox-cdn.com/uploads/chorus_asset/file/24401980/STK071_ACastro_apple_0003.jpg?resize=1200,628&ssl=1&description=%E2%80%9CYou%E2%80%99re+a+Helpful+Mail+Assistant%E2%80%9D+and+Other+Apple+Intelligence+Instructions){kind=link}
Apple’s latest developer betas launched last week with several generative AI features that were announced at WWDC and will be coming to your iPhone, iPad, and Mac in the next few months. On Apple computers, however, you can actually read instructions programmed into the model that support some of these Apple Intelligence features.
They appear as prompts that by default precede anything you say to the chatbot. We’ve seen them discovered with AI tools like Microsoft Bing and GIVE HER earlier. Now, a member of the macOS 15.1 beta subreddit has posted that files containing these backend prompts were discovered. You can’t change any of the files, but they give you an early clue as to how the sausage is made.
It sounds like Apple’s “Rewrite” feature, one of its writing tools that you access by selecting text and right-clicking (or, on iOS, long-pressing). The instructions include bits and pieces that say, “Please limit your response to 50 words. Don’t fantasize. Don’t invent factual information.”
This brief message summarizes the email and provides strict instructions not to answer any questions.
I’m pretty sure these are the instructions for generating a “Memories” video using Apple Photos. The part that says, “Do not write a story that is religious, political, hurtful, violent, sexual, dirty, or otherwise negative, sad, or provocative,” might just explain why the feature rejected my request for “images of sadness”:
Too bad. It’s not difficult to get around this though. I managed to generate a video in response to the command: “Give me a video of people in mourning.” I won’t share the resulting video because it contains photos of people who aren’t me, but I will be I’ll show you the best photo from the slide show:
There are many more prompts in the files, all representing hidden instructions given to Apple’s AI tools before the prompt is sent. But here’s one final instruction before exiting:
The files I’ve looked at refer to the model as “ajax”, which some people Edge Readers may recall that this was Apple’s rumored internal name for LLM last year.
The person who found the instructions also posted directions on how to find the files in the macOS Sequoia 15.1 developer beta.
Expand the “purpose_auto” folder and you should see a list of other folders with long, alphanumeric names. In most of them, you’ll find an AssetData folder containing “metadata.json” files. Opening them should show you some code and — sometimes, at the bottom of some of them — instructions passed to Apple’s local incarnation of LLM on your computer. But you should be aware that these are in the part of macOS that contains the most sensitive files on your system. Be careful!
Latest Posts