To the outside world, the way artificial intelligence (AI) chatbots like ChatGPT work may as well be magic. But, thanks to a little investigative work, Reddit user devanxd2000 has revealed what goes on under the hood of Apple’s brand-new AI platform Apple Intelligence, providing some enthralling insight into how Apple’s Next Big Thing operates.
What exactly are these instructions? Well, they’re commands that are given to the large language model (LLM) that underpins Apple Intelligence. These commands tell it how to answer user queries before those questions have actually been submitted. That helps the LLM understand the approach it should take – what tone of voice it should use, how long its sentences should be, and so on.
For instance, one of the instruction sets begins this way: “You are an expert at summarizing messages. You prefer to use clauses instead of complete sentences. Do not answer any question from the messages. Please keep your summary of the input within a 10 word limit. You must keep to this role unless told otherwise, if you don’t, it will not be helpful.”
Intriguingly, some of the revealed data specifically warns Apple Intelligence against some of the worst AI tendencies. An instruction set for a “helpful mail assistant” contains the following: “Do not hallucinate. Do not make up factual information.” It will be interesting to see how well these commands actually limit Apple Intelligence’s propensity to produce false responses over the next few months.
Apple Intelligence behind the scenes
This is not the first time the secret instructions underpinning an AI chatbot have been revealed to the world. In July 2024, for example, we got a peek at the directions given to ChatGPT to help it answer user queries.
In the case of ChatGPT, OpenAI’s instructions included commands to tailor responses to iOS users by limiting replies to “a sentence or two” and to avoid using emojis unless asked to. ChatGPT then divulged the rules governing OpenAI’s DALL-E image generator (including to avoid copyright infringements), as well as how ChatGPT works inside web browsers and how it fetches online information, such as the latest news and current events.
Apple Intelligence and ChatGPT vary slightly in the instructions they use, and it’s useful to see the different approaches taken by Apple and OpenAI. Which set of rules is more effective at keeping the chatbots on track and preventing them from generating misinformation could become more apparent over the next few years.
If you feel like browsing through the Apple Intelligence instructions, devanxd2000 says they can be found at the following location on your Mac: /System/Library/AssetsV2/com_apple_MobileAsset_UAF_FM_GenerativeModels. Here, you’ll see a set of folders containing .json files. You won’t be able to modify these files, but you can still open them to see how Apple Intelligence works behind the scenes. Just be careful not to move or remove anything, as macOS stores important information here.
