Anonymization alters information to prevent it from being linked to an identifiable person. It can reduce privacy risk before data enters a training or fine-tuning pipeline, where removing learned information later may be difficult. But removing names and other direct identifiers is not enough to guarantee anonymity: combinations of attributes can still identify someone. section 5.2.2
Search logs re-identified
In 2006, AOL published search logs with numeric IDs in place of usernames. Journalists and researchers identified users from their queries, including a woman in Georgia. The detail in the searches could be linked to individuals despite the removal of usernames. section 5.2.2
Generative AI adds another exposure path. A model trained on documents with redacted names may still learn identifying context, such as job titles, neighborhoods, and events. A later prompt containing part of that context might lead it to reconstruct or guess an identity. Models can also memorize and reproduce training data, so a claim that inputs are anonymous should not be treated as proof that outputs cannot expose people. section 5.2.2 section 1.1
Apply anonymization alongside data minimization, not as a reason to collect everything. An enterprise policy chatbot, for example, may need only de-identified policy files rather than an entire HR drive containing résumés and ID scans. For an API-based system, an organization-controlled pre-call pipeline can validate inputs, anonymize sensitive data, and encrypt protected fields before a request crosses the trust boundary. section 5.2.2 section 2.3.1








