A generative AI model can repeat sensitive information from its training data. Privacy-enhancing technologies (PETs) are methods that reduce privacy risks, including the risk that information can be traced back to a person. They include anonymization, pseudonymization, data masking, and differential privacy. section 3.4 section 3.17
Sensitive data in email
Research involving the Enron Email Dataset found that credit card and Social Security numbers could be extracted with minimal effort. The finding illustrates why unintended memorization is a privacy concern. section 3.4
One approach is to filter training data using blocklists of potentially sensitive sequences. Blocklists can remove known patterns, but they cannot protect against memorization of unseen or rare content. More generally, sensitive information can be hard to find in large datasets, and masking it can disrupt the language patterns a model learns from. section 3.4 section 10.18
Differential privacy introduces statistical noise during training to limit the influence of any individual data point. It can be difficult to implement for large language models without sacrificing usefulness, and it does not prevent memorization of unique, infrequent strings. section 3.4
PETs reduce risk rather than eliminate it. Even after direct identifiers are removed, personal information may be inferred from apparently harmless data. Technical protections therefore sit alongside measures such as privacy risk assessments, internal review, and clear data policies. section 3.17 section 10.18

