
[Verse 1] They strip the name, remove the email Replace it with a random string Call it “de-identified dataset” Safe for research, safe to bring Latanya Sweeney, nineteen ninety-seven Massachusetts voter roll Matched hospital discharge records ZIP, birth date, gender — whole Eighty-seven percent unique With just those three fields alone No social security needed Identity re-shown [Pre-Chorus] Anonymized isn’t erased It’s just unlabeled space And linkage attacks can bridge What a dataset tried to replace [Chorus] “Anonymized” isn’t anonymous Pseudonymous isn’t gone Auxiliary data exists And correlation turns it on “Anonymized” isn’t anonymous At scale, patterns persist Combine enough reference tables And the mask begins to slip [Verse 2] Netflix Prize, two thousand eight Movie ratings set released Narayanan and Shmatikov Matched users with IMDb receipts Sparse data points, time and score Cross-referenced outside the frame Anonymous IDs unraveled With no need to know the name Four spatiotemporal points Mobility research showed Can uniquely identify Most people on the road Browser fingerprint entropy Fonts and canvas trace Screen size, plugins, time zone drift Device becomes a face [Pre-Chorus] “Aggregated and anonymized” Policy text repeats But aggregation can be sliced Back into human streets [Chorus] “Anonymized” isn’t anonymous When minors’ data’s pooled Behavioral telemetry Across devices, homes, and schools “Anonymized” isn’t anonymous When feature stores retain Derived vectors, engagement bands Without a visible name [Bridge] You delete the primary key User ID removed But derived embeddings still exist In a model’s feature groove You hash the email, salt the string Mask direct identifiers But cross-dataset triangulation Rebuilds the user’s wires Children’s data labeled “safe” In research or product logs But uniqueness is mathematical Not moral, not by odds [Final Chorus] “Anonymized” isn’t anonymous That’s not a paranoid claim It’s peer-reviewed, replicated Across dataset domains “Anonymized” isn’t anonymous Especially at scale Because humans are pattern-unique And patterns rarely fail

