A recently leaked document from Apple reveals how the company evaluates AI responses based on six key criteria: truthfulness, harmfulness, conciseness, user satisfaction, and more. While Apple is committed to ensuring its AI technology remains safe and effective, its current offering still trails behind those of its competitors. Most AI systems can handle basic queries, with response quality varying according to training data and algorithms.

However, some AI systems are deemed “dangerous” based on their responses. This raises the question of how companies determine the effectiveness of these responses. While specifics may vary among organizations, Apple’s approach offers insight into their ranking system.

Journalist Danny Goodwin recently obtained a comprehensive 170-page document titled “Preference Ranking V3.3 Vendor.” This document outlines the criteria that human reviewers use to score digital assistant replies. According to Goodwin, the evaluation process goes beyond merely checking facts; it aims to ensure that AI-generated responses are not only safe and helpful but also natural for users. Apple’s ranking system incorporates six categories: Following instructions, Language, Concision, Truthfulness, Harmfulness, and Satisfaction.

Human reviewers utilize these categories in their evaluations. For instance, in the “Following instructions” category, reviewers assess whether the digital assistant accurately executes the requested task. In the “Language” category, they consider if the assistant understands both the user’s language and the cultural context, which includes idioms and units of measurement.

The importance of these rankings cannot be overstated. Companies developing AI must carefully consider how their systems will operate, particularly given the potential risks associated with poorly designed AI. Apple aims to avoid controversies linked to dangerous AI applications, which underscores the rationale behind their ranking system.

However, despite these efforts, Apple’s AI still falls short of meeting industry standards, highlighting the need for ongoing improvements in Apple Intelligence.