AI: File discoveries allow insight into Apple Intelligence prompts

Files found in the current beta of macOS 15.1 provide detailed insights into how Apple Intelligence works and the so-called prompts that Apple uses for this. Since they became known, the instructions that tell the AI model not to hallucinate have been the subject of controversy in AI forums. Some experts believe it is questionable whether the AI can reliably follow these instructions at all.

Apple's new artificial intelligence is set to make its debut on iPhones, iPads and Macs this fall with iOS 18.1, iPadOS 18.1 and macOS 15.1. The current test versions only support US English as a language, with other languages to follow later. Users in the European Union and China are also excluded from using it – it is still unclear whether and when Apple will reach an agreement with the EU and China to release Intelligence there too.

Mentions of Apple's AI "Ajax"

The prompt templates that have now been found- templates that are given as instructions to the AI together with user input – were found in the /System/Library/AssetsV2/com_apple_MobileAsset_UAF_FM_GenerativeModels folder in various JSON files. In some places, the name "Ajax" is mentioned, which was already mentioned as Apple's project name for its own Large Language Model (LLM) in the run-up to the presentation of Apple Intelligence.

In one of the prompt templates, Apple instructs the AI to be "a helpful mail assistant" that is supposed to generate questions based on mail texts, with which users can quickly create answers by tapping buttons (Smart Reply). The questions should be short and no longer than eight words, and the answers should only be two words long. The template also instructs the AI to only provide "valid JSON" as output - the software is then apparently supposed to evaluate this JSON response to offer the selection options based on it.

Nothing sad, sexual or religious

Other findings point to the rewrite function, with which existing texts can be reformulated in a formal tone, for example, and to the summary function for emails (according to the prompt, a maximum of three sentences, less than 60 words).

According to commentators, it is questionable whether the AI can actually be persuaded by instruction not to simply invent facts. Some say that the AI cannot provide a reliable answer to the request to only deliver valid JSON by calculating the highest possible probabilities. It is therefore more important that the software then checks the JSON response to see whether it is valid and simply starts new attempts until the desired result is available or a maximum number of attempts has been exceeded.

The supposed prompt for creating photo reminders in Apple's future operating systems is also interesting. It calls on the AI not to create any religious, political, violent, sexual or negative stories. These should also not be sad or provocative, it continues.

(mki)