Could a language model reveal something private about a person whose data it learned from? Differential privacy is a privacy-enhancing technique designed to limit how much any one data point affects a model’s behavior. It introduces statistical noise during training, giving a mathematical basis for protecting individual contributions rather than relying only on finding and removing sensitive details. section 3.4
That distinction matters because simpler filters have blind spots. A blocklist can remove known sensitive patterns but cannot catch every rare or unexpected piece of private information. Differential privacy offers a more principled protection, yet it is difficult to apply effectively to large language models: stronger privacy can come at the cost of model performance, and the technique does not prevent every form of memorization, including rare, unique strings. section 3.4
Differential privacy is one part of a broader approach to data protection, not a substitute for responsible data practices. Anonymization and masking also have limitations, while organizations still need clear policies explaining what data they collect, how they use it, and how people can opt out or request deletion. Users should remain cautious about entering private information into AI tools. section 3.17 section 10.18
