Privacy and preparation
Dataset privacy claims language: describing de-identification without promising zero risk
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
A supplier should describe de-identification by what was done, not what the data now is: which fields and record types were treated, by what method, how the result was tested and on what sample, the recipient and use assumed, the risk that remains and the controls managing it. Avoid anonymous, untraceable and zero risk; regulators treat remaining risk as low and assessed, never absent.
Redaction
Dataset privacy claims
Key takeaways
- Describe the method, the test and the remaining risk; never promise that records cannot be re-identified.
- NIST, HHS and the ICO all treat remaining risk as something to reduce and document, not to eliminate.
- Hashing alone does not make data anonymous, so a card should not say it does.
- U.S. tests also look at public commitments and recipient contracts, so the card and the license must agree.
- Every permitted phrase needs an evidence record a second person can open.
- When the evidence does not exist, stop and bring in a privacy specialist instead of softening the wording.
How should a supplier describe de-identification and the privacy risk that remains?
Describe what was done rather than what the data now is. A defensible privacy statement names the fields and record types treated, the method used on each, how the result was tested and on what sample, the recipient and use the assessment assumed, the risk that remains and the controls that manage it. It avoids words that promise an outcome nobody can prove, such as anonymous, untraceable or zero risk.
The difference matters because a privacy statement in a data card, a sales sheet or a license schedule is a representation. Buyers repeat it to their own reviewers, and regulators compare representations against practice. A sentence that says records cannot be re-identified is a claim about every possible recipient with every possible outside dataset. A sentence that says direct identifiers were removed from named fields, quasi-identifiers were generalized, and a documented test found no re-identification on a stated sample is a claim the supplier can back.
This page gives a reviewed claims checklist for that wording and an evidence table for each element of the statement. It is general information, not legal advice; the legal standard that applies depends on the records, the jurisdictions involved and the contract, and counsel should review the final language.
Why can no dataset description promise zero re-identification risk?
Because the regulators and standards bodies that define de-identification do not treat the remaining risk as zero, and a description that does so overstates what any method delivers. The bar they set is a low, assessed and documented risk, not its absence.
NIST SP 800-188, the federal guidance on de-identifying government datasets, says inference risk cannot be eliminated if there is any information in the de-identified data, though it can be managed or reduced to an acceptable level. The HIPAA expert determination method, as HHS describes it, asks a qualified expert to determine that the risk is very small that an anticipated recipient could identify an individual. The UK ICO describes identifiability as a spectrum and frames effective anonymisation as making the risk of identification sufficiently remote, judged against the means reasonably likely to be used.
The FTC's technologists made the same point about a common shortcut. In a 2024 post they wrote that data is only anonymous when it can never be associated back to a person, and that hashing does not achieve that, because a hash of the same input is always the same output and can still be used to recognize a user. A data card that calls hashed customer IDs anonymous is making exactly the claim that post describes as flawed.
What do the legal tests expect a de-identification description to show?
Several U.S. tests ask for more than a technical step; they also look at public commitments and contracts. Wording that describes only the technique can therefore be incomplete even when the technique is sound.
California's definition of deidentified information in Civil Code section 1798.140 requires that the information cannot reasonably be used to infer information about or be linked to a particular consumer, and that the business holding it takes reasonable measures to keep it that way, publicly commits to keep and use it in deidentified form without attempting re-identification, and contractually obligates recipients to comply. The FTC's 2012 privacy report set out a similar three-part approach: reasonable de-identification measures, a public commitment not to re-identify, and a contractual prohibition on re-identification by downstream recipients.
For a supplier, this means the privacy statement and the license should match. If the card says recipients may not attempt re-identification, the agreement should say so. If the card describes a de-identification method, that description is part of what a regulator could compare against practice. Which test applies to a given dataset is a question for counsel.
The reviewed claims checklist: wording to replace and the evidence each version needs
Run every privacy sentence in the card, the sales material and the license schedule through this table before release. The left column holds wording that tends to appear in first drafts; the right two columns hold what a reviewer can sign off and the record that must exist for it to stay true.
The replacements are longer on purpose. Each one says what was done, to what and under which assumption, which is what lets a reader judge whether the statement fits their own use. Keep the qualifiers even when marketing asks for something shorter; a short privacy sentence is almost always a broader one.
Privacy wording rarely lives in one place. It shows up in the data card, in a one-page summary sent to prospective buyers, in answers to a buyer's security and privacy questionnaire, in the license schedule and sometimes in a press release. Each is written by a different person at a different time, and the least careful one is the one a buyer quotes back.
Pick one controlling text, normally the privacy section of the data card for the release, and have every other document quote it or point to it rather than paraphrase it. Questionnaire answers are the usual weak point: they are written under time pressure, often by sales or customer success staff, and tend to round up. Route them through the same reviewer who signed off the card.
Version the controlling text with the dataset. When release 2.0 changes the method or adds a modality, the old wording must not survive in a sales deck or an onboarding email, because it now describes a dataset that no longer exists.
| Tempting wording | Why it fails review | Defensible replacement | Evidence it needs |
|---|---|---|---|
| Fully anonymized | States an outcome for every recipient and every outside dataset | Direct identifiers removed from the named fields; quasi-identifiers generalized as listed | Field inventory with the treatment applied to each field |
| Cannot be re-identified | Absolute claim no test can establish | A re-identification test on a stated sample found no confirmed matches under the stated assumptions | Test design, sample size, attacker assumptions, results and date |
| Hashed, so anonymous | A consistent hash still links records and can identify users | Customer IDs replaced with keyed tokens; the key is held separately and not delivered | Tokenization method, key custody record, delivery manifest |
| Zero privacy risk | No recognized standard treats remaining risk as zero | Residual risk assessed as low for the stated recipient and use, with these controls | Risk assessment naming recipient, use, controls and reviewer |
| Compliant with all privacy laws | A legal conclusion that depends on facts and jurisdictions | Reviewed by counsel against the laws listed for the stated use | Counsel's review reference and the scope it covered |
| No personal data | Usually untrue for free text, images or audio | Personal data in free text, attachments and images was treated as described; residual items found in review are listed | Modality review log and sampled findings |
| Safe for any use | Risk depends on the recipient and purpose | Assessed for model training by the licensee under the agreement's restrictions | License terms that restrict recipients and prohibit re-identification |
| Permanently de-identified | Linkage risk changes as new outside data appears | Assessment dated and scheduled for review on each release or new data source | Review schedule and the latest dated assessment |
What belongs in the evidence table behind the privacy statement?
Each element of the privacy statement should map to a record a second person can open. The table below is the evidence side of the checklist; a statement whose evidence row is empty is cut or narrowed before release.
| Statement element | Evidence record | Reviewer | Reopen when |
|---|---|---|---|
| Fields and record types covered | Field inventory marking direct identifiers, quasi-identifiers and free text | Data owner and privacy reviewer | A field is added or a source system changes |
| Method per field | Transformation specification and the pipeline run that applied it | Engineering lead | The rules or the pipeline version change |
| Testing | Test design, sample, attacker assumptions and results | Privacy specialist | Method, sample frame or outside data changes |
| Cross-system linkage | List of other datasets the recipient may hold and the join keys checked | Privacy specialist | The dataset is combined with another delivery |
| Other modalities | Review log for attachments, images, audio and embedded metadata | Modality reviewer | A new modality is included |
| Recipient and use | License schedule naming permitted use and restrictions | Counsel | The buyer, use or territory changes |
| Residual risk and controls | Dated risk assessment with accepted risks and controls | Privacy specialist and data owner | Any of the rows above changes |
| Incident handling | Procedure for a reported re-identification and the contact who owns it | Security lead | Contacts or contract terms change |
What identifying combination or modality could remain after the transformation?
Ask this before writing a single sentence about the result, because the answer decides what the statement may say. Removing names and account numbers addresses direct identifiers; it says nothing about combinations of ordinary fields or about content that lives outside the structured columns.
The ICO ties the assessment to context: what the information is, who may get or want access to it, and what other information they could use to identify someone. Data can be identifiable in one recipient's hands and not in another's. Operational records are full of the details that make that assessment hard.
- Combinations of date, location and job type that single out one customer or one event.
- Rare events, such as an unusual failure or a large claim, that are visible to people outside the company.
- Free-text notes that mention names, addresses, phone numbers or family details the field-level rules never saw.
- Attachments and images with faces, license plates, signatures or embedded file metadata.
- Audio with voices and spoken names, which text redaction does not reach.
- Consistent tokens or hashes that let a recipient link records across deliveries or to their own data.
- Precise timestamps that match public or customer-held logs.
Worked example (illustrative): rewriting the privacy paragraph in a field service data card
A hypothetical field service company drafted this paragraph for its work order dataset: the data is fully anonymized, contains no personal information and carries zero risk of re-identification. Each clause failed the checklist. Customer IDs had been hashed, technician notes were free text, and site photos were included.
After review the paragraph read as follows. Customer names, phone numbers and street addresses were removed from structured fields. Customer IDs were replaced with keyed tokens; the key is held by the company and is not delivered. Street addresses were generalized to the first three digits of the ZIP code. Technician notes were processed by a named-entity redaction step, and a second reviewer read a sample of 500 notes after redaction; residual items found are listed in the appendix and were removed. Site photos were excluded from this release. A re-identification test under stated assumptions found no confirmed matches. The licensee may not attempt re-identification or combine the data with other sources to identify individuals. The assessment is dated and will be repeated for each release. All counts are illustrative.
The rewrite is longer and less comforting, and every sentence in it points to a record. The exception is worth noting: the team removed the photos instead of describing them, because no review method it had could support a sentence about faces and plates in thousands of images.
Should privacy statements and usefulness claims be reviewed together?
Yes. The same transformation that lowers risk changes what the data can teach, and a card that describes each in isolation can mislead on both. If ZIP codes were truncated, a claim that the dataset supports local demand forecasting needs to be narrowed, and the card should say why.
Put the two side by side for each treated field: what was done for privacy, what analytic use it affects, and whether the claimed use was retested after the change. Reviewers then see the trade-off the supplier made rather than discovering it during a pilot. Where a privacy step removed too much for a claimed use, the honest options are to drop the use claim or to change the method and reassess, never to soften the privacy wording.
A short paired note per field is enough. For the illustrative work order dataset, it would say that ZIP codes were truncated to three digits for privacy, that this limits analysis below the regional level, and that the regional demand use claimed in the card was retested on the truncated data. A buyer reading that note knows both what was protected and what was given up.
Signals that the wording problem is really a privacy problem
Some drafts cannot be fixed by an editor because the evidence the defensible wording needs does not exist. When that happens, stop the release and bring in a privacy specialist and counsel rather than searching for vaguer language.
Vaguer language is the usual temptation, and it does not help. Phrases such as industry-standard de-identification or privacy-safe sound modest but make an unbounded promise, because they invite the reader to assume whatever standard they had in mind. A specialist can usually say quickly whether the gap is a missing test, a missing contract term or a dataset that should be narrowed before anything is said about it.
- No field inventory exists, so nobody can say which fields held identifiers.
- The dataset includes health, financial, children's or biometric information, which carry their own legal tests.
- Free text, images or audio are included and no sample was reviewed after treatment.
- The buyer already holds data that shares keys, dates or locations with this dataset.
- The draft relies on hashing or a single masking step as the whole method.
- The license does not restrict re-identification or onward sharing.
- Someone asks for wording that says anonymous because a buyer requested it.
How privacy preparation is described when SourceX prepares the dataset
Where SourceX prepares the dataset, de-identification is carried out by SourceX through approved processing routes, so the method description can be written alongside the work instead of reconstructed later by the supplier. The company approves the scope and terms and signs the agreement before delivery, which is the moment the privacy wording, recipient restrictions and record scope should line up. Checking eligibility takes a short assessment and no data upload.
Frequently asked questions
Can a dataset be described as anonymous?
Only with great care, and usually it should not be. The FTC's technologists describe data as anonymous only when it can never be associated back to a person, a standard that operational records rarely meet. Describe the methods used, the tests run and the remaining risk instead.
Is hashed data de-identified?
Not by itself. A consistent hash of the same input always produces the same output, so hashed IDs can still link records and identify users. Keyed tokens with the key held back, plus other measures, are a stronger basis, and the statement should describe them specifically.
What does very small risk mean under HIPAA expert determination?
HHS guidance describes the expert determination method as a qualified expert finding that the risk is very small that an anticipated recipient could identify an individual. It is an assessed, documented standard tied to the recipient, not a promise of zero risk.
Should the data license repeat the privacy statement?
The two should agree. California's deidentified definition and the FTC's 2012 framework both look at contractual commitments by recipients, so a statement that recipients may not re-identify the data should be backed by a matching clause. Counsel should review both together.
How often should a de-identification statement be updated?
Whenever the data, the method, the recipient or the outside data that could be linked to it changes, and at least for each release. Date the assessment and say when it will be reviewed, so a reader can tell whether it still applies.
Sources
- The FTC Office of Technology wrote in July 2024 that data is only anonymous when it can never be associated back to a person, and that hashing does not make data anonymous. Source
- The FTC's 2012 privacy report treats data as not reasonably linkable when a company takes reasonable de-identification measures, publicly commits not to re-identify, and contractually prohibits downstream recipients from re-identifying it. Source
- California Civil Code 1798.140 defines deidentified information and requires reasonable measures, a public commitment not to re-identify, and contractual obligations on recipients. Source
- HHS describes HIPAA expert determination as a qualified expert finding the risk very small that an anticipated recipient could identify an individual. Source
- NIST SP 800-188 says inference risk cannot be eliminated if there is any information in de-identified data, but can be managed or reduced to an acceptable level. Source
- The ICO describes identifiability as a spectrum and effective anonymisation as reducing identification risk to sufficiently remote, considering who may access the data. Source
See if your company qualifies
A short company assessment. No data uploads are needed.