AssociationAI / AI Literacy
AssociationAI team Published

Article

Beginner

Why Removing Names Doesn't Make Member Data Anonymous

A 2000 study of 1990 Census data found that ZIP code, gender and birth date alone likely singled out most Americans. Deleting the name column from a member export leaves those traits behind.

AI Literacy Privacy Member Data Data Quality

Three columns that most member exports carry without a second thought were enough, in one well-known study, to point to a single person for most of the country. In a 2000 working paper from Carnegie Mellon University, the privacy researcher Latanya Sweeney reported on 1990 U.S. Census data: “It was found that 87% (216 million of 248 million) of the population in the United States had reported characteristics that likely made them unique based only on {5-digit ZIP, gender, date of birth}.”

That figure is about the U.S. population in 1990, and the word “likely” matters. It says nothing directly about any association’s members, and it is not a claim about people today. It does describe a pattern that association staff run into every time a vendor asks for “some real-looking data” for a pilot.

What the number means for a membership desk

Here is how the request usually arrives. The Kettle Prairie Society of Records Managers (an invented organization) is trying an AI tool that drafts renewal reminders. The vendor wants a sample file. A staff member exports last year’s renewals, deletes the name, email and phone columns, and calls the result anonymous. The file still has each member’s exact birthday, five-digit ZIP code, gender, chapter and credential. In one small town, exactly one member was born on a particular day in 1971. Anyone who knows that member can find her row in a few seconds.

This is a teaching example, not a case study.

Nothing in that file looks personal at a glance, which is the problem. Sweeney’s summary of the research is blunt: “It was found that combinations of few characteristics often combine in populations to uniquely or nearly uniquely identify some individuals.” The column that gave the row away was not a name. It was the overlap of three ordinary facts.

Her conclusion follows directly: “Clearly, data released containing such information about these individuals should not be considered anonymous.” In the introduction she makes the same point about the habit our invented staff member followed. Stripping out the obvious fields, like a person’s name, street and phone, and then roughly coarsening the rest, she argues, falls short of anonymity, since the leftover attributes can still combine to re-identify someone.

And the number of traits it takes is small. “In general, few characteristics are needed to uniquely identify a person.”

An AI trial makes this sharper, not softer. A file pasted into a chat window or uploaded to a vendor portal has left the membership system, and staff rarely know who at the vendor can open it, how long the vendor keeps it or whether it will be joined with anything else. Once a file is out, the society cannot pull it back or control what it is matched against. So the review belongs before the upload, while the file is still on a staff laptop and every column can still be changed.

Why a partial or sampled file is not safe either

A common reply is that a test file is only a slice of the membership, so a match is unlikely. Later research tested that idea. In a 2019 paper, researchers writing in Nature Communications built a statistical model to estimate how often a person in an incomplete dataset could be correctly matched. Their headline estimate: “Using our model, we find that 99.98% of Americans would be correctly re-identified in any dataset using 15 demographic attributes.”

That is the authors’ model estimate for Americans described by 15 demographic attributes, not a count of real people who were found. Few member exports have 15 such fields. Many have six or eight, and a few of those, such as chapter, credential and join year, are rare enough inside a professional society to do the work of several census fields.

The same authors take on the sampling argument directly: “Our results suggest that even heavily sampled anonymized datasets are unlikely to satisfy the modern standards for anonymization set forth by GDPR and seriously challenge the technical and legal adequacy of the de-identification release-and-forget model.” GDPR is the European Union’s data protection law, and many U.S. associations are not bound by it. We read the line as the researchers’ view of what anonymization has to mean, not as a legal test for your organization.

Taken together, the two papers say the same thing from two directions. Sweeney shows how few ordinary traits it takes. The 2019 team shows that handing over only part of a dataset does not fix it.

Where member exports carry these traits

Association data has its own versions of ZIP, gender and birth date. In our reading of typical exports, these are the columns to look at before anything leaves the membership system:

  • Full birth dates kept for age-based dues, student rates or retirement categories.
  • Five-digit ZIP codes and city fields kept for chapter assignment.
  • Chapter, credential, specialty and join year, which narrow quickly in a small society.
  • Member numbers, which feel safe since they are not names, even though anyone with access to the membership system can look them up.
  • Notes and comment fields, where a staff member may have typed a name, an employer or a health detail years ago.

A corporate member’s contact rows add another layer: a job title plus an employer can be a single person.

What to do before a test file leaves the building

Our advice is to treat “names removed” as “not yet anonymous” and to check what is left, not only what was taken out. Three moves cover most of the risk for a pilot file.

First, decide whether the tool needs member data at all. Many AI trials work just as well on invented rows that share the export’s column layout, and an invented file has nothing to re-identify. Our earlier article on what an AI system does with what you type covers why the destination matters too.

Second, if real data is needed, coarsen the traits that combine. Keep birth year instead of full birth date and the first three ZIP digits instead of five, and fold rare chapters or credentials into broader groups. These are our suggestions, not a standard.

Third, count before you send. For each mix of the traits that remain, find the smallest group of rows sharing it. A group of one is a person. Our companion tutorial on checking a test file for identifiers before trying an AI tool walks through that count step by step and comes with a lab that grades invented test files as STOP, FIX or PASS against teaching rules.

What a PASS does not mean

None of these steps make a file anonymous in a legal sense, and the research above is a reason for modesty about any single check. A file that clears a group-size rule can still be matched with outside information nobody on staff thought of. The safest test file is the one with no real members in it, and the next safest is the smallest real one that still lets the pilot answer its question.

Sweeney’s paper is more than 25 years old, and the habit it describes is still easy to fall into: delete the obvious columns and stop. The traits left behind are usually the ones that identify. This is general information, not legal advice.

Sources (2)