Is de-identification the same thing?

Usage varies. Specify what was removed or transformed and what identifiability assessment was performed rather than relying on the term alone.

Is synthetic data automatically anonymous?

No. Its generation process and possible disclosure of source information still need assessment.

Can aggregation reveal information?

Yes. Small groups or comparisons between overlapping outputs can expose details, so the release must be considered as a whole.

AI Glossary / Governance, ethics and safety / Advanced

Anonymization

Definition

Anonymization is the process of changing information about people so that they are no longer identifiable within the relevant context and applicable standard. It addresses indirect identification as well as obvious identifiers such as names. Whether information is effectively anonymous depends on the data, other available information and the circumstances of access or release, not simply on the label attached to a file.

What it means

Removing a name may leave a distinctive combination of details that identifies the person. A precise location, unusual job and dated event can reveal an identity when combined with information elsewhere. Anonymization therefore concerns what someone can reasonably learn from the information as a whole.
Pseudonymization is different: replacing an identifier with a code can reduce exposure while retaining a way to reconnect the records to people. It should not automatically be described as anonymization. In UK guidance, the ICO treats pseudonymous information as personal data; legal interpretation and thresholds must be checked for the relevant jurisdiction.
In AI work, anonymization can support analysis and data sharing, but training a model does not itself make the source information anonymous.

How it works

The process begins by defining the intended use and audience. A public download creates different exposure from analysis in a controlled environment. The team identifies direct identifiers, potentially identifying combinations and sensitive information that might be inferred.
Possible transformations include suppressing fields, replacing precise values with broader categories and releasing suitable aggregates instead of individual records. Each choice changes both disclosure risk and usefulness. Removing too much detail can make the intended analysis unreliable, while retaining unusual combinations can leave people exposed.
Assessment considers linkage to other information and what a recipient could infer. Technical checks need documented assumptions, appropriate expertise and review as circumstances change. Encryption and access controls protect information, but they do not by themselves establish that the underlying records are anonymous. The release decision should reflect the full process, not just successful completion of a masking tool.

Three practical examples

Service reporting: An organization publishes regional usage totals rather than customer-level records. It reviews small groups and overlapping reports that could expose individual contributions.
Research preparation: Analysts replace exact ages with bands and detailed locations with broader areas. They then assess remaining combinations instead of assuming that generalization guarantees anonymity.
AI evaluation: A team prepares customer-support examples for testing. Removing names is insufficient if a narrative still describes a uniquely recognizable incident, account detail or quoted message.

Think of it this way

Taking a name badge off someone does not make them unrecognizable to everyone who knows their face, workplace or story. Anonymization asks whether the remaining clues can still lead back to the person. Identifiability depends partly on others' knowledge.

Worked example

Consider a fictional table of twelve people. Names have been removed, but exact age and district remain. For the chosen combination of age and district, eight records are unique and four form two pairs.
The proportion unique on those selected fields is 8 / 12, approximately 66.7%. That is a warning about this representation; it is not a measured probability that each person will be identified.
Suppose ages are grouped into broader bands and districts into regions. The twelve records now form three groups of four with identical values on those selected fields. The minimum group size is four, satisfying a k-anonymity condition with k = 4 for that specified set of fields.
This does not prove effective anonymization. Another retained field could separate the records. If all four people in a group share the same sensitive outcome, that outcome might be inferred even without identifying a particular row. Four is an illustrative group size, not a universal safe threshold or legal standard.

Illustrative case study

Fictional company Cedar Surveys plans to share feedback for analysis. Staff remove names and email addresses, then discover that several free-text comments mention exact appointment dates and unusual events. A colleague could recognize the authors.
The team revisits the purpose and decides that the external analyst needs themes and broad service categories rather than the original narratives. It prepares a more limited extract, checks rare combinations and documents who will receive the information and under what conditions.
Specialist review assesses whether the proposed output is appropriately anonymous for that use. The case illustrates why a name-removal step cannot justify a blanket claim. It does not claim that replacing comments with themes automatically resolves every disclosure risk.

Common misconceptions and mistakes

“No names means anonymous.” Indirect clues and combinations can still identify people.
“A random ID makes records anonymous.” A retained mapping or other linkable information may make them pseudonymous instead.
“Encryption is anonymization.” Authorized decryption restores the original information; encryption addresses access rather than eliminating identifiability.
“A privacy metric settles legal status.” A metric measures particular properties under assumptions, while legal and contextual assessment may require more.
“Anonymous forever.” New external information, new releases or changing capabilities can affect the earlier assessment. Review is part of responsible management.

Business use

Clarify whether the objective requires individual records at all. A carefully designed statistical output may answer the business question with less exposure. Document the transformation, intended recipients, relevant external information and limitations of the assessment.
For AI datasets, examine free text and unusual combinations alongside structured identifiers. Check that transformations preserve the features needed for valid evaluation. Obtain appropriate privacy and legal review before relying on a claim of anonymity for a real disclosure; this glossary entry is an explanation, not a determination about a particular dataset.

Check your understanding

What does the 66.7% figure measure?

Answer: Uniqueness on the selected fields, not an observed re-identification rate.

Why is k = 4 insufficient as a blanket guarantee?

Answer: Other information and attribute disclosure may remain relevant.

Why consider the recipient?

Answer: Available knowledge and access conditions affect what the released information can reveal.

Related terms

De-identification
k-anonymity
Differential privacy
Pseudonymization
quasi-identifier
re-identification
data minimization
disclosure risk

Learn next

Compare identity disclosure with inference of sensitive attributes. Then study the assumptions behind k-anonymity and differential privacy, keeping technical guarantees separate from legal conclusions.

Sources

NIST SP 800-188, De-Identifying Government Datasets: Techniques and Governance. https://csrc.nist.gov/pubs/sp/800/188/final ICO, Introduction to Anonymisation, UK-specific distinctions. https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/anonymisation/introduction-to-anonymisation/ ICO, How Do We Ensure Anonymisation Is Effective? https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/anonymisation/how-do-we-ensure-anonymisation-is-effective/ ICO guidance was marked under review when checked in October 2026. The table and Cedar Surveys case are original teaching illustrations.

Previous: Anomaly detection Next: Ansari-Bradley test