The launch of AI-generated review summaries on the App Store with the release of iOS 18.4 has introduced a new way for users to quickly understand the collective opinion on apps and games. In its machine learning blog post, Apple has detailed the intricate process behind this functionality, revealing how AI works to synthesize thousands of reviews into a concise and useful summary.
This feature leverages a multi-stage Long Language Model (LLM). Apple’s primary goal is to provide overviews that encompass a diversity of opinions, maintain a balance between positive and negative points, and accurately reflect what users truly think. The company emphasizes that security, fairness, truthfulness, and usefulness are fundamental pillars in creating these summaries.
AI comes to App Store reviews
The process begins with an intelligent screening phase. The LLM is programmed to automatically identify and discard reviews containing unwanted content such as spam, offensive language, or any indication of fraudulent activity. In addition, given a pool of legitimate reviews, the system processes them through a sequence of modules, each powered by LLMs specifically trained for their task.
At each stage, the AI extracts the most relevant information from each individual review. This allows it to identify frequently recurring themes in user opinions. The system then evaluates the sentiment associated with these themes, differentiating between positive and negative comments.
Finally, all this information is condensed into a concise summary, typically between 100 and 300 characters, providing a snapshot of the overall perception of the application.
A crucial aspect of Apple’s approach is the specialization of its language models. By using LLMs specifically trained for each step of the process, the company aims to ensure that the final summary is a faithful representation of user feedback. Furthermore, extensive human review was conducted during the development of this feature.
Thousands of abstracts were reviewed by human reviewers to determine their usefulness, consistency,a nd balance of content, and accuracy, ensuring they were free of bias or inappropriate information.
AI-powered App Store review summaries operate through a complex, multi-stage process involving extensive, specialized language models, supported by human oversight, underscoring Apple’s commitment to truthfulness. Hopefully, it can work without losing much of the “sauce” that authors who take the time to write reviews often produce.
