Apple is creating new, different techniques to analyze user data trends and collective insights to improve its AI capabilities. On Monday, the Cupertino tech giant unveiled its use of differential privacy technologies, pointing out that user privacy is completely secure. Additionally, Apple is focusing on collecting data such as usage patterns and embeddings to improve its text generation features and Genmoji. However, the company noted that this data will only be collected from devices where users have consented to share Device Analytics.

Apple wishes to profit from User Data While Preserving Privacy

In a recent update on its Machine Learning Research platform, Apple introduced a technique developed to boost specific Apple Intelligence capabilities. The company accepts that its AI performance has been lacking, pointing to its commitment to ethical practices in data collection and training processes.

According to Apple, its generative AI relies on synthetic data, which is produced by algorithms or digital systems, not collected from real people. Although this method allows LLMs to gain a general understanding of the world, the lack of human-written content means their outputs often seem lifeless or formulaic, a phenomenon sometimes referred to as 'AI slop'.

To improve  AI output and avoid human-like natural responses, Apple is looking to responsibly learn from user behavior without compromising data privacy. The method is known as “different privacy “ according to Apple.

To improve Genmoji, Apple is dependent on differential privacy to track popular prompts and user behaviour, with data collected only from those who have chosen to share device analytics. The iPhone provides a mathematical assurance that unique or unusual prompts will not be identified, and that no specific prompts can be accessed by individual users.

By collecting this data, the company is focused on analyzing prompts that reflect actual user interactions. Apple will fundamentally analyze the types of signals that produce positive results and those signals where users need to adjust their input multiple times to get the desired outcome.

This approach will be expanded by the company to include Image Playground, Image Wizard, Memory Creation, and Writing Tools in Apple Intelligence, with future applications in Visual Intelligence.

Another important area where Apple is applying this technique is in text generation. Apple's plans for email generation are somewhat different from those of Genmoji. The company has developed a collection of emails on common topics to assess the performance of its devices . This company generated several variations for each topic, then derived email representations that captured area dimensions such as language, topic, and length.

Currently, Apple is applying this method to enhance text generation in emails and has stated that it will use the same approach for email summaries in the future.