Apple is counting on your emails to resolve issues with its malfunctioning AI.

      Apple's AI initiatives have not achieved the same level of influence as Google's Gemini, Microsoft Copilot, or OpenAI's ChatGPT. The company's AI framework, referred to as Apple Intelligence, has not significantly improved functionality for iPhone and Mac users, leading to an internal management crisis.

      User data could potentially help reverse this trend. Today, the company released a Machine Learning research paper outlining a new method to train its on-device AI using data stored on your iPhone, beginning with emails. These emails will be utilized to enhance features like email summarization and Writing Tools.

      A brief overview of AI training

      Nirave Gondhia / Digital Trends

      Before delving into the details, here’s a quick summary of how AI tools operate. The initial phase involves training, which essentially means supplying a vast amount of human-generated data to an "artificial brain"—such as books, articles, research papers, and more. The more data it receives, the more accurate its responses become.

      This is because chatbots, technically referred to as Large Language Models (LLMs), aim to grasp the patterns and relationships among words. Tools like ChatGPT, which are now incorporated into Siri and Apple Intelligence, function essentially as predictors of word use.

      However, the amount of available data for AI training is limited, and the entire process is quite time-consuming and costly. So why not utilize AI-generated data for training the AI? Research indicates that this would effectively “poison” AI models, resulting in more inaccurate answers, nonsensical information, and misleading results.

      How is Apple planning to enhance its AI?

      Nadeem Sarwar / Digital Trends

      Rather than depending solely on synthetic data, AI responses can be enhanced through refinement and fine-tuning. The optimal way to train an AI assistant is to provide it with more human-generated data. The content on your phone is one of the richest sources for this information, but a company cannot simply access it.

      Doing so would violate privacy rights and invite legal challenges.

      Apple’s strategy involves indirectly analyzing your emails without copying or transferring them to its servers. In essence, all your data will remain on your device. Furthermore, Apple will not technically “read” your emails; instead, it will compare them to a set of synthetic emails.

      The key aspect is identifying which synthetic data closely resembles a human-written email. This would help Apple understand how humans typically converse. According to Bloomberg, Apple has predominantly used synthetic data for AI training.

      "This synthetic data can then be used to evaluate the quality of our models with more representative data and pinpoint areas for enhancement in features like summarization,” the company states. This could lead to meaningful improvements in the responses you receive from Siri and Apple Intelligence in the future.

      Nadeem Sarwar / Digital Trends

      By utilizing insights from authentic human data, Apple aims to refine its email summarization system and enhance some Writing Tools features. “The content of the sampled emails never leaves the device and is never shared with Apple,” the company assures. Apple has already implemented similar privacy-centric training systems for its Genmoji system.

      Why is this a significant advancement?

      Currently, the summaries generated by Apple Intelligence in Mail can often be confusing and, at times, completely nonsensical. The notification system has faced similar issues, leading Apple to temporarily suspend it after receiving criticism from the BBC for misrepresenting news articles.

      The situation has deteriorated to the point that the summarized notifications have become a source of humor in our team chats. Apple Intelligence often combines random sentences in its attempts to summarize conversations or emails, resulting in nonsensical outputs or completely altered meanings.

      The fundamental issue is that AI still struggles with context and human intent. The most effective way to address this is by training it using more contextually-aware material with a better understanding of the situation. Recently, AI models with reasoning capabilities have emerged, but they are not a complete solution.

      Nadeem Sarwar / Digital Trends

      The method proposed by Apple appears to strike a balance. “This approach allows us to enhance the subjects and language of our synthetic emails, assisting us in training our models to produce improved text outputs for features like email summaries, while ensuring privacy,” the company states.

      Here’s the good news. Apple will not be scanning all emails stored on iPhones and Macs globally. Instead, it will adopt an opt-in model. Only those users who have actively consented to share Device Analytics data with Apple will participate in the AI training process.

      You can enable this feature by navigating to: Settings > Privacy & Security > Analytics & Improvements. The company is expected to initiate these plans with the upcoming iOS 18.5, iPad 18.5, and macOS 15.5 beta updates. A corresponding build for developers has already been released.