App Store: Apple provides insight into how AI summaries are created

iOS 18.4 introduced AI summaries of user reviews in the App Store. The complex creation process is now explained in a blog post.

listen Print view
Section of an iPhone screen with the App Store app icon.

(Image: BigTunaOnline / Shutterstock)

3 min. read

At first glance, Apple's new summary for app reviews looks like a simple task for artificial intelligence. The function introduced with iOS 18.4 is intended to give App Store users a quicker overview of what others have to say about an app. On closer inspection, however, it is much more complex, as older and incorrect reviews can give a skewed picture – Apple has now explained in a blog post how it intends to overcome this challenge using various AI models.

Regular readers of app reviews know that the ratings of an app often include entries that miss the point – For example, in the case of a delivery service app, criticism of the quality of the food at a participating restaurant or resentment towards the app developer. Critical reviews for older versions with bugs that have long since been fixed also threaten to distort the picture.

Apple is therefore turning the two sentences into a veritable science, as now explained in the company's own machine learning blog. A multi-stage system of Large Language Models (LLMs) is designed to deliver the most useful summaries possible with a length of between 100 and 300 characters.

In a first step, all reviews are sent through a filter to filter out spam, insults and fraudulent content. A specially tuned LLM then extracts the core statements from each review. In the next step, these statements are sorted into different topics by another model. In the third step, the most relevant topics are then selected. Another AI model then produces the summary in the desired length and in the style of Apple.

The publication of iOS 18.4 was preceded by a review by human testers who, according to Apple, checked the summaries for various aspects. This includes ensuring that the content is not harmful or offensive and that the ratings used are truthful. Further steps were taken to check whether the grammar is correct and whether the summaries are useful when making download and purchase decisions.

Videos by heise

Apple also sees its own approach as a blueprint for similar applications where large amounts of user feedback are available. The summaries are not yet available in the German App Store. They are being rolled out gradually and are currently only available for selected apps in the US store.

(mki)

Don't miss any news – follow us on Facebook, LinkedIn or Mastodon.

This article was originally published in German. It was translated with technical assistance and editorially reviewed before publication.