Apple seeks to fix flaws in its AI by allowing it to learn from your emails.

Apple's efforts in artificial intelligence have not had the same impact as Google's with Gemini, Microsoft's with Copilot, or OpenAI's with ChatGPT. Its AI toolkit, known as Apple Intelligence, has failed to deliver significant improvements for iPhone and Mac users, even leading to an internal management crisis within the company. However, it has been suggested that user data could be the solution to reverse this precarious situation.

Recently, the company published a research paper on machine learning that describes a new method for training its internal AI using data stored on iPhones, starting with emails. This information would be used to refine features like email summarization and writing tools.

To understand how these AI tools work, it is important to know that the first stage involves feeding a large volume of human-generated data to an AI model, supplying it with texts from books, articles, and research papers. The greater the amount of data, the better the responses it can generate, as these models try to understand patterns and relationships between words. Tools like ChatGPT, which are now integrated into Siri and Apple Intelligence, act as word predictors. However, the training process can be slow and costly, raising the question of why not use AI-generated data to train the AI itself. Nevertheless, research indicates that this could "contaminate" AI models, resulting in inaccurate and misleading answers.

Instead of relying solely on synthetic data, Apple aims to improve the responses of its AI tool by refining it with concrete human data. The information stored on phones is a rich source for this type of data, but accessing it without authorization could violate privacy and lead to legal actions. Therefore, the company plans to perform indirect analysis of emails without copying or sending the information to its servers. This means that the data will remain on the user's device. Instead of reading the emails, Apple will compare them with a set of synthetic emails to identify which ones are most similar to those written by humans, thereby helping to better understand how people communicate.

So far, Apple has primarily used synthetic data for AI training. These techniques allow for testing model quality with more representative data and identifying areas for improvement. With insights gained from more realistic data, Apple aims to optimize its email summarization system and various features of its writing tools, ensuring that email information does not leave the device and is not shared with the company.

This advancement is particularly important given that, currently, summaries generated by Apple Intelligence in Mail can be confusing or even absurd. The situation has become so complicated that Apple had to temporarily pause app notifications after receiving criticism regarding how they represented news articles. Artificial intelligence continues to struggle with understanding context and human intent, and increasing the quality of the material it is trained on is essential for improvement.

The methodology proposed by Apple seems to combine the best of both worlds, enhancing the quality of generated texts without compromising user privacy. Furthermore, the approach will be optional: only users who have consented to share device analytics data will participate in this AI training process. This feature will be implemented in upcoming updates of iOS 18.5, iPad 18.5, and macOS 15.5, a developer version of which has already been released.