Privacy and preparation
Purpose-specific dataset privacy: one HR file, five releases, measured
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Purpose-specific dataset privacy means judging a file against the use a recipient will make of it, not against a generic clean label. In a generated database of 240,124 employees, dropping names and IDs still left 99.95 percent singled out by fields a coworker knows. Field lists for three named uses and one generalized variant singled out 0 to 38 people; in a 297-person slice the pay-audit list singled out 42.09 percent.
Redaction
Purpose specific dataset
Key takeaways
- A privacy verdict belongs to one use, one recipient and one population; change any of the three and the verdict has to be redone.
- Dropping direct identifiers is a step, not a verdict: the generic release of the generated file left 240,008 of 240,124 current employees alone on birth date, gender, hire date, title and department.
- The same table needed different fields for each use: a title pay benchmark singled out nobody, while a pay-equity audit singled out 38 people, all nine managers among them.
- Size moves the answer as much as purpose: in a 297-person slice the pay-audit fields singled out 125 people, and 34 even after hire year was banded.
- Keep two records per use, a necessity reason for every field and a group count on the fields that leave, plus the triggers that reopen the review.
What does purpose-specific dataset privacy decide?
It decides which fields leave the company, in what form, for one named use by one named recipient. Two records settle it, kept side by side: a necessity record that says why each field is needed for that use, and a count of how many people the remaining fields single out in the population the recipient will actually receive.
Rule-makers in three jurisdictions tie necessity to purpose in this way. The ICO's guidance on the UK GDPR data minimisation principle treats adequate, relevant and limited as tests against a stated purpose: enough to fulfil it, a rational link to it and nothing beyond it, and it notes that the regulation does not define the terms. In the US, the HHS minimum necessary standard asks HIPAA covered entities to make reasonable efforts to limit protected health information to what an intended purpose requires, and to justify any use of an entire record. In the EU, the EDPB's Opinion 28/2024 on AI models asks supervisory authorities to verify that a controller relying on legitimate interest documented a necessity step, including whether a less intrusive route exists and whether the amount of personal data is proportionate.
None of these texts calls a file safe in general. Each poses a question that cannot be answered until someone writes the use down. This article shows what changes when they do, on a generated HR database where a generic release and four purpose-built releases give five different answers. It is general information, not legal advice; counsel and a privacy specialist decide for a real file.
Why a clean-data label settles nothing on its own
A label such as names removed or PII stripped records what left the file, not what stayed or who will hold it. The fields that remain can still single people out to anyone who knows a few facts about them, and the label carries no use, recipient or population to judge that against.
The ICO frames the question as a motivated intruder test: could someone who wants to identify people do it with practical steps and the means reasonably likely to be used? In an HR file the obvious intruder is a colleague. A coworker knows who joined in March, who works in which department and who holds the manager title, so birth date, hire date, title and department are the fields to count, even though none of them is a direct identifier.
ISO/IEC 27559:2022 describes its subject as a framework for identifying and mitigating re-identification risks across the lifecycle of de-identified data. A lifecycle has stages, recipients and uses; a label has none of them. The measurement below puts numbers on that gap.
Five releases of one generated HR database: what each exposed
Purpose changed the exposure by four orders of magnitude, from 240,008 people singled out to 16 or none. The file is the MySQL employees sample database, whose README says its data was generated; it holds 300,024 employee records, and 240,124 people have a current salary, title and department. For each release I counted how many current employees are the only person with their combination of the coworker-knowable fields that use keeps.
The generic release, which drops employee numbers and names and keeps everything else, left 240,008 of 240,124 people singled out, 99.95 percent. A monthly hiring forecast needs only department and hire month, and as person-level rows it singled out 16 people. A title pay benchmark keeps one knowable field, the title, and singled out nobody. A pay-equity audit needs gender, title, department and hire year beside pay, and singled out 38 people, including all nine current employees titled Manager.
The slice column is the warning. Suppliers SourceX works with start at 50 full-time employees at peak, a long way from 240,000, so every count was repeated on a fixed slice of 297 current employees: those whose employee number is divisible by 800. There the hiring-forecast rows single out 218 people (73.40 percent) and the pay-audit fields 125 (42.09 percent). An identical field list becomes a different privacy decision at a different size.
Each release's field list and its counts at both sizes, measured on 2026-10-28 by the script named in the methods note:
| Release (intended use) | Fields that leave | Singled out, 240,124 people | Singled out, 297-person slice | In groups under 5, slice |
|---|---|---|---|---|
| Generic release, direct identifiers dropped | Birth date, gender, hire date, title, department, salary | 240,008 (99.95%) | 297 (100.00%) | 297 (100.00%) |
| Monthly hiring forecast | Department, hire month | 16 (0.01%) | 218 (73.40%) | 297 (100.00%) |
| Title pay benchmark | Title, salary | 0 | 0 | 4 (1.35%) |
| Pay-equity audit | Gender, title, department, hire year, salary | 38 (0.02%) | 125 (42.09%) | 279 (93.94%) |
| Pay-equity audit, generalized | Same, with hire year in three bands and the 9 managers removed | 0 | 34 (11.45%) | 135 (45.45%) |
The privacy-utility review matrix, step by step
Every use gets its own pass through six steps, and every step leaves a record a second reviewer can recheck. The matrix pairs the test that settles each step with the evidence to keep, and shows what the step produced for the pay-equity audit in the generated file.
Two steps carry the weight. The necessity test asks whether the use's output changes when a field is removed or coarsened; if it does not, the field stays home. The identifiability count runs on the delivery's real population, not on the warehouse, because the slice shows that size alone can move the answer from a few dozen people to four in ten.
The six steps, with the generated audit as the worked column:
| Step | Test that settles it | Evidence to keep | Pay-equity audit, generated file |
|---|---|---|---|
| 1. Name the use and recipient | A written statement of the decision the data serves and who receives it | Use statement, recipient, intended outputs | Gender pay comparison within title, department and hire cohort, for one audit team |
| 2. Necessity per field | Does the use's output change when the field is removed or coarsened? | Field list with a reason each, and refused fields with a reason each | Birth date, names and exact hire date add nothing the comparison uses, so they stay home |
| 3. Identifiability count | Group size on fields an outsider could know, in the population delivered | People singled out and in small groups, with the threshold set beforehand | 38 of 240,124 singled out; 125 of 297 in the slice |
| 4. Treatment | Coarsen, aggregate, remove rows or restrict the recipient, then recount | Counts before and after each treatment | Hire year in three bands, managers removed: 0 of 240,115; 34 of 297 |
| 5. Utility recheck | Does the treated set still give the use's output within a stated tolerance? | The output before and after treatment | Female pay coefficient -0.082 percent before, -0.105 percent after |
| 6. Sign-off and reopen rule | Who approved this use, and which change reopens the review | Signed record naming use, recipient, population and triggers | A new use, a new recipient or a smaller population reopens it |
What happens to the same fields in a 297-person company?
They single out far more people, and turning rows into counts does not rescue the smallest uses. The hiring forecast looks safest as a table of counts, yet in the 297-person slice every one of its 256 department-by-month cells holds fewer than five people and 218 cells hold exactly one. Anyone who knows who joined in a given month can read that person's department straight from the table.
A coarser time unit helps less than it seems. By hire year instead of month the slice table shrinks to 95 cells, but 71 still hold fewer than five people and 31 hold one. In the full workforce the monthly table has 1,623 cells, of which 39 hold fewer than five people, and all 39 fall in 1985, 1999 and 2000, the first and last stretches of the hiring history.
For a mid-size supplier the review outcome therefore differs from the warehouse answer. Deliver the forecast by quarter or year and by department group, release only cells above a threshold set for the recipient, or decide that this use does not justify a person-level extract at all. The pay audit in the slice still singles out 34 people after hire year is banded, which points toward delivering audit results rather than rows, or limiting who may hold the rows.
The strongest objection: once names are gone, why does the use matter?
Because removing direct identifiers changes who can read a record at a glance, not what the remaining fields reveal or what the recipient may do with them. The generated file shows the first point in numbers: the generic release dropped every name and employee number and still left 99.95 percent of people alone on fields a colleague knows.
US health privacy rules already write the use into the release. HHS describes a limited data set as protected health information with specified direct identifiers removed, which may be used or disclosed only for research, public health or health care operations and only under a data use agreement. The same fields are acceptable for one purpose and not for another; the agreement, not the file, carries the difference.
Use also decides whether a coarser field set costs anything. In the generated file the audit's pay estimate moved from -0.082 percent to -0.105 percent when hire year became a band and managers left, a shift of 0.023 points across 108 comparison cells instead of 555. That passes a reasonable tolerance here, but generated salaries follow simple rules. On real payroll the recheck could fail, and then the use, not the label, decides between keeping the field under tighter access and narrowing the use.
Privacy is also only one permission among several. A file can pass every count above and still be blocked by a customer contract or a software vendor's terms, so the privacy review sits beside a rights review rather than replacing it.
Final-delivery checklist for a purpose-bound release
Run the list for every use, at the size and in the form the recipient will receive, and keep the record each item produces:
Checklist · 9 items
- The use statement names the decision, the outputs and the recipient, and the delivered fields match it line for line.
- Every delivered field has a necessity reason on file, and every field that was requested but refused has one too.
- Group sizes were counted on the fields an outsider could know, in the delivered population, with the threshold written down before counting.
- Rare roles and edge rows, such as the only manager in a department, were found by rule rather than by eye, then removed, coarsened or justified in writing.
- Aggregate tables were checked cell by cell, because a count of one discloses as much as a row.
- The utility recheck ran after treatment, and its result is recorded against a tolerance set beforehand.
- The license scope repeats the use, so a recipient's new use arrives as a written request rather than an assumption.
- Reopen triggers are on file: a new use, a new recipient, a smaller or different population, or added fields.
- The exported files were recounted after export, since joins and format conversions can bring back fields the review removed.
0 of 9 ticked. Tick each item as you go.
Where purpose-specific review meets a SourceX licensing deal
In a SourceX deal the supplier draws the first boundary itself. On SourceX's trust page, the supplier decides which systems, records and date ranges are in scope, and everything else stays out. The legal-framework page says SourceX looks at customer contracts, privacy notices and employee policies and leaves out any data the company has no right to share.
The preparation work follows the signed paperwork. Per the trust page, once SourceX gives the go-ahead, the data is "extracted, transferred and de-identified through approved processing routes", and the legal-framework page adds that the method used is recorded for each dataset. The same page says plainly that no method is perfect: de-identification reduces risk but cannot remove it, and re-identification may still be possible. A use-by-use review like the one above is the kind of record that makes a supplier's own scope decision easier to explain and defend.
The first step involves no records at all. The company assessment asks about systems and history without any data upload, and it is not an offer; it says nothing yet about what any buyer would pay.
Methods note: the generated database, the rules and what was left out
Corpus: the MySQL employees sample database in github.com/datacharmer/test_db at commit e324b56, read from a shallow clone on 2026-10-28. Its README says the data was generated, so no real person is in it. Row counts matched the README's expected records: 300,024 employees, 331,603 department rows, 443,308 title rows and 2,844,047 salary rows. No repository code was run; the script bin/purpose_fieldsets_testdb.py parses the dump files and prints counts only, never names or employee numbers.
Rules as applied: a current employee has a row dated 9999-01-01 in each of the salary, title and department tables, and the script checks that no one has two (240,124 people; none had two). The coworker-knowable fields are birth date, gender, hire date, title and department, an editorial list. A person is singled out when no other current employee shares their values of the knowable fields a use keeps; a small group has fewer than five people. Hire-year bands are 1985-1989, 1990-1994 and 1995-2000. The slice is every current employee whose employee number is divisible by 800, a fixed rule with no randomness; it contains no managers. The pay check is ordinary least squares of log salary on a female indicator with one fixed effect per title, department and hire-year cell (hire-year band for the generalized set, which also drops the nine managers), reported as a percent difference over the full workforce. The salaries are generated, so the check tests whether the estimate holds steady under treatment, not whether anyone is underpaid. Percentages are rounded once, half up.
The four purpose-built field lists are editorial necessity calls, not a legal standard. Not measured: real payroll or real companies, linkage to outside data (none was attempted), free-text fields, and how a regulator would treat any of these files. Rule-maker text comes from the ICO (UK), HHS (US), the EDPB (EU) and the ISO catalogue, each read through a search limited to its own domain on 2026-10-28.
Frequently asked questions
Is a dataset with names removed still personal data?
Often it is. Under the UK GDPR the ICO asks whether someone motivated to identify people could do it with means reasonably likely to be used, and in the generated HR file in this article 99.95 percent of current employees stayed alone on birth date, gender, hire date, title and department after names were dropped. Whether a real file is anonymous is a question for counsel and a privacy specialist.
Does a privacy review approved for one AI use cover another?
No. The necessity record and the group counts were made for a named use, recipient and population, and changing any of the three can change the result. A recipient that wants to add evaluation, fine-tuning or a new party should send a written request that reopens both the review and the license scope.
How is purpose-specific review different from data minimization?
Data minimization is the principle that a field set should be adequate, relevant and limited to what a purpose needs. Purpose-specific review applies it to one named use and then measures what the surviving fields reveal in the population delivered, so the output is a pair of records rather than a principle.
Can an aggregate table identify people?
Yes, when its cells are small. In the 297-person slice of the generated file, 218 of the 256 department-by-hire-month cells held exactly one person, so the table told any colleague who knew a joining month which department that person works in.
What group-size threshold should a reviewer use?
There is no single number. This article flags groups under five purely for illustration; the right value depends on the recipient, the controls on the copy and the harm a disclosure would cause, and it should be written down before anyone counts.
Sources
- The MySQL employees sample database, whose README says the data was generated; the source of every count in this article. Source
- The ICO reads adequate, relevant and limited as tests against a stated purpose under the UK GDPR. Source
- The ICO's motivated intruder test asks whether people could be identified with means reasonably likely to be used. Source
- HHS asks covered entities to limit protected health information to what the intended purpose needs. Source
- HHS limits a limited data set to research, public health or health care operations under a data use agreement. Source
- The EDPB asks supervisory authorities to check a documented necessity step when AI development relies on legitimate interest. Source
- ISO/IEC 27559:2022 is a framework for re-identification risk across the lifecycle of de-identified data. Source
- SourceX's trust page on supplier-set scope and data prepared through approved processing routes after the go-ahead. Source
- SourceX's legal-framework page on rights review, recorded de-identification methods and residual risk. Source
See if your company qualifies
A short company assessment. No data uploads are needed.