Stripping names and identification numbers from a dataset is the standard step before sharing it, and it is widely treated as sufficient. The protection it provides is weaker than the practice assumes, and the reason is combinatorial rather than technical.
Uniqueness comes from combinations
No single ordinary attribute identifies anyone. Many people share a postcode, a birth year, a job title.
Combining a handful of them collapses the population quickly. A postal area, a date of birth and a stated gender narrow most people to a very small group, and often to one.
The attributes involved are not sensitive on their own, which is exactly why they survive the anonymisation step.
The record has no name on it and still refers unambiguously to one person, which is what identification means in practice.
External datasets supply the missing link
Re-identification works by joining. The stripped dataset provides the sensitive column, and some other source provides the name.
Electoral registers, property records, professional directories, social media profiles and leaked databases all contain the same ordinary attributes attached to identities.
Anyone holding both can match rows, and the match does not need to be certain to be damaging.
The publisher of the anonymised data cannot assess this risk, because it depends on what other data exists in the world, which is unknown and growing.
Behaviour is more identifying than demographics
Location traces, purchase histories and browsing records are far harder to anonymise than a form with fields.
A small number of location points is often enough to isolate an individual, because the pattern of where someone sleeps and works is close to unique.
The same holds for the set of films someone has rated, the specific pages they read, or the times of day they are active.
These datasets have no name field to remove, and the identifying content is the data itself rather than an attached label.
Coarsening helps — rounding times, enlarging geographic areas — and it reduces the analytical value that motivated collecting the data in the first place.
Aggregate figures leak more than expected
Publishing counts rather than records seems safe, and it can still expose individuals.
If a table reports that a small geographic area contains three people with a particular condition, and a neighbouring breakdown reports two, the difference identifies someone to anyone who knows the area.
Repeated queries against the same underlying data compound this, since each answer constrains the possibilities further.
Statistical agencies have known this for decades and apply suppression rules, which is why official tables have gaps where cell counts are small.
Differential privacy changes the guarantee
The formal response is to add calibrated random noise, so that the published result would be nearly the same whether or not any particular person was included.
This provides a mathematical statement about exposure rather than a hope that nobody joins the right table.
The cost is accuracy, and the trade-off is set by a parameter that determines how much noise is added and therefore how much protection exists.
Choosing that parameter is a policy decision dressed as a technical one, and it is frequently made by whoever wants the data to remain useful.
Synthetic data is not automatically safe
Generating artificial records that reproduce the statistical properties of a real dataset is an appealing alternative.
The protection depends entirely on how the generator was built. A model that fits the original data too closely will reproduce unusual real records nearly verbatim.
Outliers are the specific danger, because they are the rows a model must memorise to reproduce and the rows that identify people.
Synthetic data with a privacy guarantee attached is meaningful. Synthetic data without one is an assertion, and it should be tested rather than believed.
Legal definitions lag the technique
Data protection regimes generally exempt data that is genuinely anonymous, and treat pseudonymous data as still regulated.
The line between them is drawn by whether re-identification is reasonably likely, which depends on available auxiliary data and on the effort someone might expend.
Both of those change over time, so a dataset released as anonymous can become pseudonymous through no action by its publisher.
Releases are irreversible. A file that has been downloaded cannot be recalled when the surrounding conditions change.
The practical posture that follows
Treating anonymisation as a binary property produces the failures. Treating it as a risk level that must be argued produces better decisions.
That means asking which external datasets plausibly exist, testing the release by attempting re-identification internally, and preferring controlled access over open publication for anything behavioural.
It also means accepting that some datasets cannot be shared usefully and safely at the same time, and that the honest answer is sometimes to publish less.
The organisations that get this wrong are rarely careless. They are following a standard step that was designed for paper records and applying it to data that behaves nothing like paper records.