Apple AI system exposed: Engineers warn against speculation or fabrication

This netizen's comment succinctly compares someone to a "nanny". This metaphor suggests that the person in question may exhibit characteristics of being overly concerned, nagging, or meddlesome, similar to the traditional image of a housewife or housekeeper.

Apple AI (Apple Intelligence) backend prompts accidentally exposed. A developer discovered what appears to be Apple Intelligence backend prompts in MacOS 15.1 Beta 1 and shared the findings on Reddit.

The exposed prompts include content in the following aspects:

  1. Email assistant function:

    • Identify relevant questions in emails
    • Generate brief questions and answers
    • Limit the length of questions and answers
    • Output results in JSON format

  2. Safety classification function:

    • Classify text-to-image queries as safe or unsafe
    • List specific categories of unsafe content

  3. Email reply function:

    • Generate initial replies based on short reply snippets
    • Modify reply content according to provided questions and answers
    • Limit replies to within 50 words
    • Emphasize not to generate hallucinations or fabricate factual information

  4. Message summary function:

    • Use phrases rather than complete sentences
    • Limit summaries to 10 words or less
    • Do not answer any questions in the message

  5. Visual concept extraction:

    • Summarize text into a maximum of 5 topics
    • Represent each topic with a single word
    • Sort the topic list by relevance

These prompts demonstrate Apple's strict control over AI assistant functions, security considerations, and emphasis on accuracy and conciseness.