This Leaked Report Shows Apple’s Preference Ranking For AI Responses
Summary: A leaked Apple document reveals how the company ranks AI responses based on six key criteria, including truthfulness, harmfulness, conciseness, and overall user satisfaction. While Apple aims to ensure its AI remains safe and helpful, its current Apple Intelligence still lags behind competitors.
Most AI systems are capable of answering basic questions you throw at them. The quality of their responses is based on a variety of factors, such as the data used to train them. However, some AI systems earn a “dangerous” label based on the way they respond. So, the question is, how do companies decide what’s a good or bad response? While we can’t speak for others, we do know what Apple thinks based on its preference ranking.
Apple’s response ranking system
Danny Goodwin at Search Engine Land managed to obtain a 170-page Apple document called “Preference Ranking V3.3 Vendor.” According to Goodwin, “It lays out the system used by human reviewers to score digital assistant replies. Responses are judged on categories such as truthfulness, harmfulness, conciseness, and overall user satisfaction. The process isn’t just about checking facts. It’s designed to ensure AI-generated responses are helpful, safe, and feel natural to users.”
Goodwin doesn’t share the actual document, but he did a very comprehensive summary of the various criteria Apple uses when it comes to ranking its preference for AI responses.
According to the document, Apple has six rating categories. This consists of Following instructions, Language, Concision, Truthfulness, Harmfulness, and Satisfaction. Human reviewers base their ratings on these categories. For example, under “Following instructions,” they assess whether the digital assistant completed the task as requested.
Under “Language”, reviewers not only need to consider if the digital assistant understood the request based on the user’s language, but also if it understands the cultural and regional context behind it. This includes things like idioms, units of measurement, and more.
Importance of these rankings
So, why are these rankings important? Obviously, it’s critical that companies that are looking to develop or implement AI need to consider how it will be used. We’ve seen how some poorly designed AI systems even allow users to generate malware, give instructions on making weapons, and so on.
Apple wants to avoid this kind of controversy, so it makes sense for them to have a preference ranking. That said, Apple’s AI efforts have been underwhelming, and Apple Intelligence still lags far behind the competition. Hopefully, in addition to ensuring some kind of quality control in its AI responses, Apple is also working to make sure Apple Intelligence is at least useful.
