Apple Uses New Tech That Compares Synthetic Data With Real Emails To Train AI Models, Then Applies Embeddings And Privacy Tools To Improve Text Output Quality

Apple Uses New Tech That Compares Synthetic Data With Real Emails To Train AI Models, Then Applies Embeddings And Privacy Tools To Improve Text Output Quality

Apple’s Personalized Siri feature has been delayed until next year, with a release window pushed beyond the initial iOS 18.4 launch. Despite this setback, internal confidence remains high that the feature will arrive later this year. The delay stems from complex technical challenges in training Apple Intelligence models, particularly regarding the limitations of using synthetic data for nuanced tasks like email summarization. To overcome these hurdles without compromising user privacy, Apple is employing a novel method that compares synthetic messages to user data locally on devices. The core of this new approach involves generating synthetic emails to represent common topics and then using device embeddings to identify the closest matches within a user’s recent messages. By leveraging differential privacy, Apple aggregates these selections to understand broader linguistic trends and improve model accuracy. This technique allows the company to refine its synthetic datasets based on real-world usage patterns while ensuring that no individual user data is exposed or collected centrally, effectively balancing performance improvements with strict privacy protections. This development is relevant to open data because it illustrates a sophisticated alternative to traditional data collection models. It demonstrates how large-scale AI training can occur by utilizing local device analytics and anonymized statistical feedback rather than centralized, personal data harvests. This highlights the evolving landscape of data governance, showing that high-quality model training does not necessarily require sacrificing user privacy or relying on opaque data extraction methods, offering a potential template for ethical AI development in privacy-conscious industries.

Related news

• Understanding Aggregate Trends for Apple Intelligence Using Differential Privacy
• Apple needs its AI to be better but wants you to know it still really, really cares about your privacy
• How Apple gets AI training data without violating anyone’s privacy
• On-device Apple Intelligence training methods seem to be based on a controversial technology
• Meta utilizará datos públicos de sus usuarios en la UE para entrenar a sus modelos de IA