Apple Upgrades Smart Reply in macOS 15.1 Beta to Prevent Hallucination" Issues

Apple Upgrades Smart Reply in macOS 15.1 Beta to Prevent Hallucination" IssuesApple has upgraded the Smart Reply feature in AppleIntelligence in the latest macOS 15

Apple Upgrades Smart Reply in macOS 15.1 Beta to Prevent Hallucination" Issues

Apple has upgraded the Smart Reply feature in AppleIntelligence in the latest macOS 15.1 beta, introducing a series of internal instructions to enhance the accuracy and reliability of the feature. These instructions, stored in JSON files, outline the workflow of the Smart Reply feature. Notably, Apple explicitly demands that the system "must not fabricate facts or generate hallucinations."

Smart Reply is a feature in Apple's Mail app that automatically generates potential reply options by analyzing email content, providing users with a convenient communication experience. Built upon Apple's own intelligent technology, this feature, like other generative AIs, suffers from "hallucination" issues.

The term "hallucination" refers to the generation of false or misleading information by generative AI models when creating responses. To prevent Smart Reply from producing erroneous information, Apple's internal instructions require the feature to solely extract explicitly stated questions from emails and provide corresponding reply options.

This measure helps reduce the error rate of AI in response generation. By limiting the AI's interpretation of email content, Apple aims to avoid AI generating inaccurate information due to comprehension bias.

Although Apple has taken steps to prevent Smart Reply from fabricating information, industry experts point out that completely eliminating "hallucination" issues remains challenging. This is because generative AI models inherently lack comprehension of text content and struggle to accurately judge the authenticity of information.

The "hallucination" problem in generative AI fundamentally stems from the limitations of its training data. Existing AI models are typically trained on massive amounts of text data, but this data may contain errors, biases, or incomplete information. Therefore, even though AI models can learn language patterns, they cannot fully grasp the truthfulness and accuracy of text content.

Moreover, there are issues with the training process of generative AI models. For instance, some models over-rely on specific types of text data during training, leading to biases when handling other text types.

To address "hallucination" issues, researchers are exploring various approaches. Some researchers are attempting to integrate external information, such as knowledge graphs, into AI models to enable them to better understand text content. Others are working on developing new training methods to enhance the reliability and accuracy of AI models.

The upgrade to the Smart Reply feature in macOS 15.1 beta demonstrates Apple's cautious approach to AI technology. However, fully resolving the "hallucination" problem requires more in-depth research and exploration. The future direction of AI technology hinges on whether researchers can overcome these challenges and develop more reliable and trustworthy AI models.