Anonymous is a myth¶

The claim that a dataset is “anonymised” because the names were removed has been falsified repeatedly, for nearly three decades, on exactly the terms that count: given a little outside information, the records name people again.

The classic demonstrations¶

The pattern was clear as early as 1997, when Latanya Sweeney re-identified the Massachusetts governor’s medical record in a supposedly anonymous release of state-employee hospital data, by matching it against the public voter roll on three fields: date of birth, sex and ZIP code. She went on to show that those three fields alone uniquely identify roughly 87% of the US population.

The demonstrations kept coming as datasets grew. In 2006 AOL published twenty million “anonymised” search queries tagged by a random user number; reporters identified individual searchers from the content of their own searches within days. In 2008 Arvind Narayanan and Vitaly Shmatikov re-identified subscribers in the anonymised Netflix Prize dataset by matching viewing histories against public IMDb ratings. In 2014 a released New York City taxi dataset, its medallion numbers hashed with a trivially brute-forcible scheme, was turned back into named drivers and traceable trips.

The general result¶

These were not lucky one-offs. In 2013 Yves-Alexandre de Montjoye and colleagues showed that four spatio-temporal points identify 95% of individuals in a mobility dataset of 1.5 million people. In 2019 Luc Rocher, Julien Hendrickx and de Montjoye generalised it: using a statistical model of uniqueness, they estimated that 99.98% of Americans could be correctly re-identified in any dataset using just fifteen demographic attributes, including age, sex and marital status.

The combination is the finding¶

Re-identification is not exotic. It follows from a plain fact about people: the combination of a few ordinary attributes is unique. That is the same logic the infrastructure-aggregation model applies to buildings and cables, the combination is the finding, turned on people. It is why the mitigations that rely on stripping identifiers, k-anonymity chief among them, degrade as soon as an adversary brings auxiliary data, and why the commercial datasets the surveillance model describes are dangerous even when they are sold as anonymous.

Last reviewed: 2026-07-08.