Apple Adjusts Data Policy to Explore Training AI Models on Anonymized User Data

Apple has adjusted its earlier commitment not to use personal customer data to train general-purpose artificial intelligence models, and has begun exploring the use of anonymized or authorized user interaction data within a specific framework to train and optimize its artificial intelligence algorithms and machine learning systems. The adjustment adopts a controlled, phased mechanism rather than comprehensive data scraping. Under this mechanism, users have clear rights to be informed and to choose whether to anonymously share de-identified text inputs, Siri interaction logs, or app usage feedback.

For a long time, Apple treated on-device priority and zero-data-collection training as core selling points, emphasizing that artificial intelligence features including Apple Intelligence mainly rely on local chip computing power. For complex computations, Private Cloud Compute promises that data is discarded immediately after reading, is not retained, and is not used for secondary training. Demand for high-quality real interaction scenario data from large language models and multimodal models has risen sharply, and Apple previously relied mainly on public web scraping and commercially purchased licensed datasets. In this adjustment, differential privacy and local synthetic data generation technologies will remain the front-line barriers in data preprocessing, to prevent sensitive personally identifiable information from becoming physically associated with model weights.