Anonymization transforms data to reduce the chance that people can be identified directly or indirectly. It is used as a privacy-enhancing technology (PET) in AI systems and can preserve useful data while lowering reidentification risk. section 3.17 section 4.3
Anonymization helps with privacy, but its protection has important limits:
- Removing direct identifiers does not prevent all inference; seemingly harmless information can still reveal personal details, particularly at large scale. section 3.17
- Detecting and masking personally identifiable information (PII) can miss sensitive data and may disrupt the natural language patterns that language models learn. section 10.18
- Anonymization is preferable to doing nothing, but it is not foolproof and works alongside clearer privacy policies, consent, deletion, and opt-out practices. section 3.17 section 10.18
Protecting witnesses in documentary film
In Welcome to Chechnya, filmmakers used deepfake technology to anonymize the faces of LGBTQ+ activists. The technique preserved the emotional force of their stories while reducing the risk of persecution, and similar approaches could protect faces in surveillance footage or on social platforms. section 4.3
Anonymization therefore involves a trade-off rather than a guarantee: privacy may improve, while data utility, model behavior, or confidence in protection can remain uncertain. section 3.17 section 10.18
