Apple’s latest developer betas launched last week with some generative AI features that were announced at WWDC and will be heading to its iPhones, iPads and Macs in the coming months. On Apple computers, however, you can read instructions programmed into the model that supports some of these Apple Intelligence features.
They appear as prompts that precede anything you say to a chatbot by default, and we’ve seen them discovered in AI tools like Microsoft Bing and DALL-E before. Now, a member of the macOS 15.1 beta subreddit posted that they discovered the files containing these backend prompts. You can’t change any of the files, but they give you an advance hint as to how the sausage is made.
In the example above, an AI bot for a “helpful email assistant” is being told how to ask a series of questions based on the content of an email. It may be part of Apple’s Smart Reply feature, which can suggest possible responses for you.
Screenshot: Wes Davis / The Verge
This sounds like Apple’s “Rewrite” feature, one of the writing tools you can access by highlighting text and right-clicking (or, on iOS, long-pressing) it. Their instructions include passages that say: “Please limit your answer to 50 words. Don’t hallucinate. Do not invent factual information.
Screenshot: Wes Davis / The Verge
This brief notice summarizes the emails, with careful instructions not to respond to any questions.
Screenshot: Wes Davis / The Verge
I’m pretty sure this is the set of instructions for generating a “Memories” video with Apple Photos. The passage that says, “Do not write a story that is religious, political, harmful, violent, sexual, dirty, or in any way negative, sad, or provocative,” may just explain why the resource rejected my suggestion to ask for “images of sadness ”:
A shame. It’s not difficult to get around, however. I was able to generate a video in response to the request: “Give me a video of people grieving.” I will not share the resulting video because it contains photos of people who are not me, but go show the best image included in the slideshow:
There are many more prompts contained within the files, all of which feature the hidden instructions provided to Apple’s AI tools before your prompt is sent. But here’s one last instruction before you go:
The files I browsed refer to the model as “ajax”, which some Border readers may remember it as the supposed internal name of Apple’s LLM last year.
The person who found the instructions also posted instructions on how to locate the files in the macOS Sequoia 15.1 developer beta.
Expand the “purpose_auto” folder and you will see a list of other folders with long alphanumeric names. Inside most of them, you will find an AssetData folder containing “metadata.json” files. Opening them should show some code and – occasionally at the bottom of some of them – instructions passed to the local incarnation of Apple’s LLM on your machine. But you must remember that they live in a part of macOS that contains the most sensitive files on your system. Tread carefully!