Researchers reveal text anonymization fails in practice: hidden personal details remain recoverable even when 90% of sensitive data is masked

Academic researchers published SPIA (Subject-level PII Inference Assessment), a new benchmark that tests how well text anonymization actually protects people's private information. Instead of measuring whether obvious personal details like names and phone numbers are removed, SPIA evaluates whether someone could figure out who you are by reading between the lines — using context clues in the remaining text. The benchmark includes 675 real documents from legal cases and online platforms.

Experiments revealed a critical weakness: even when over 90% of explicitly labeled personal information (names, addresses, dates) is masked, someone reading the anonymized text can still infer facts about the original person 67% of the time. This means anonymization tools that appear to work well by traditional measurements are leaving your identity and details vulnerable through subtle contextual clues that remain in the text.

This matters if your personal information ends up in shared datasets or legal documents — companies and researchers often anonymize documents before sharing them publicly or with third parties, assuming the anonymization is safe. This research shows those assumptions are wrong, meaning your health records, legal filings, or financial histories could be re-identified by someone with time and context knowledge, even after professional anonymization. Organizations that rely on anonymization for GDPR compliance or data sharing now have evidence their current methods are insufficient.