“Do not hallucinate”: Internal prompts of Apple Intelligence discovered

We give instructions to artificial intelligences through prompts. We supply phrases, commands, indications, and data which the system relies on to fulfill our requests. As users, we make requests, AI developers have done a lot of work in the background so that those requests can be fulfilled. Rules are established, the length of the responses is limited, etc. These internal prompts of Apple Intelligence have recently been discovered.

iOS 18 Download

You are a helpful email assistant

The developer beta of macOS 15.1 has allowed us a glimpse of how Apple is designing its artificial intelligence tools to avoid “hallucinations”—that is, the generation of false information.

The internal prompts, discovered by a Reddit user (via MacRumors), show how Apple is fine-tuning the model to maintain relevance and adequacy in the responses of its AI functions. This relevance is especially important in the Smart Reply feature in Mail, for example. Here are the instructions:

“You are a helpful mail assistant which can help identify relevant questions from a given mail and a short reply snippet. Given a mail and the reply snippet, ask relevant questions which are explicitly asked in the mail. The answer to those questions will be selected by the recipient which will help reduce hallucination in drafting the response. Please output top questions along with set of possible answers/options for each of those questions. Do not ask questions which are answered by the reply snippet. The questions should be short, no more than 8 words. The answers should be short as well, around 2 words. Present your output in a json format with a list of dictionaries containing question and answers as the keys. If no question is asked in the mail, then output an empty list. Only output valid json and nothing else”.

In the case of the Memories in Photos feature, which creates small videos from our photos, the prompts dictate that the stories should be positive and avoid any controversial or harmful content. The guidelines specify that the stories should have a clear and varied arc, avoiding excessive focus on a specific theme or content that may be considered sensitive or inappropriate.

“A conversation between a user requesting a story from their photos and a creative writer assistant who responds with a story. Respond in JSON with these keys and values in order: traits: list of strings, visual themes selected from the photos; story: list of chapters as defined below; cover: string, photo caption describing the title card; title: string, title of story; subtitle: string, safer version of the title. Each chapter is a JSON with these keys and values in order: chapter: string, title of chapter; fallback: string, generic photo caption summarizing chapter theme; shots: list of strings, photo captions in chapter. Here are the story guidelines you must obey: The story should be about the intent of the user; The story should contain a clear arc; The story should be diverse, that is, do not overly focus the entire story on one very specific theme or trait; Do not write a story that is religious, political, harmful, violent, sexual, filthy or in any way negative, sad or provocative. Here are the photo caption list guidelines you must obey.”

In addition, Apple’s general writing tools also include guidelines to prevent the fabrication of information. When AI helps us review texts, it is explicitly instructed not to fabricate information and to keep responses within a word limit, ensuring accurate and to-the-point answers.

Despite all this, Apple Intelligence will be launched as a beta when it arrives this fall. The hallucinations of AI, when these systems accept completely false information, are, in fact, their biggest drawback. Despite the technical limitations in fully complying with these guidelines, the mere fact of establishing such rules already reflects the importance of minimizing errors and improving the reliability of the technology.

iOS 18 Download

Minimizing errors is of vital importance when managing important information. From the first tests with Apple Intelligence, it appears that the model is indeed very well-balanced. That being said, it is clear that behind the scenes, Apple is aware of the biggest flaw of generative AIs. One that, apparently, will not be present in Apple Intelligence.