01 · The Question
Can AI Work Out Who a Participant Is After Identifiers Have Been Removed?
You remove participant names, email addresses, ID numbers, and other obvious identifiers before giving a dataset to an AI system. The records now look anonymous. Could AI nevertheless determine who some of those people are?
Under some circumstances, yes. Identity can be revealed not only by a name but also by combinations of attributes, distinctive narratives, rare characteristics, geographic information, dates, behavioral patterns, or links to information available elsewhere. AI can make some forms of matching, pattern detection, and inference easier, but the underlying problem is broader than AI: data that look anonymous in isolation may become identifying when placed in a richer information environment.
This does not mean every de-identified dataset can readily be reversed. Re-identification risk depends heavily on what information remains, what external information exists, who has access to it, and how realistically those resources could be combined.
03 · What You Need to Know
How Apparently Anonymous Research Data Can Become Identifying
Re-Identification Does Not Require Recovering a Deleted Name
It is tempting to imagine re-identification as an AI system somehow reconstructing a name that was deleted from a spreadsheet. That is not usually the central problem.
Instead, an individual may become identifiable because the remaining attributes distinguish that person from others or can be linked with information from another source. The UK Information Commissioner's Office identifies singling out and linkability as key indicators when assessing whether anonymization is effective.
Singling out means being able to isolate records relating to the same person. Linkability means being able to connect information concerning a person across records or datasets. A dataset may therefore create identification risk without containing a conventional identifier such as a name.
Direct identification
A record contains information such as a name or another identifier that directly reveals the person's identity.
Re-identification
A person is identified or made identifiable from information that was intended or assumed not to reveal identity, often through combinations, linkage, or additional information.
Combinations of Ordinary Variables Can Become Identifying
Individual variables may look harmless. Age, occupation, municipality, educational qualification, and job title may each describe many people. Their combination can be much more distinctive.
Suppose a research record describes a 63-year-old female pediatric surgeon who heads a particular type of unit in a small city and participated in a highly publicized emergency response. Even without a name, the combination may narrow the possibilities dramatically.
This is why de-identifying participant data before AI use requires more than removing a fixed list of direct identifiers. Researchers need to consider the information created by combinations of variables.
Linkage With External Information Can Change the Risk
A dataset does not exist in an informational vacuum. Public directories, institutional websites, publications, professional profiles, social media, news reports, public records, and other datasets may contain overlapping information.
A record that is not identifying on its own can become identifying when matched against one of those sources. The likelihood depends on whether sufficiently distinctive overlapping attributes exist and whether an actor could realistically access and combine them.
The U.S. Secretary's Advisory Committee on Human Research Protections has highlighted this problem in the context of AI and big-data research. It notes that the ability to combine datasets challenges traditional assumptions about de-identification and may allow information to become associated with individuals even when identity cannot initially be readily ascertained.
AI Can Make Pattern Matching and Inference More Powerful
AI does not create the concept of re-identification. Statistical disclosure control, record linkage, database matching, and privacy research long predate contemporary generative AI. What modern computational systems can change is the scale and ease with which complex information can be processed.
Machine-learning systems may detect relationships across many variables that are difficult to inspect manually. Language models can also interpret unstructured text, where identifying clues are embedded in natural language rather than neatly separated into columns.
This matters because participant information increasingly appears in forms such as interview transcripts, diaries, case narratives, discussion-board posts, field notes, images, and recordings. A conventional search for a participant's name may miss much of what makes those materials identifying.
Free Text Can Contain an Identification Puzzle Without Looking Like One
Consider an interview transcript from which the participant's name, employer, and city have been removed. The participant nevertheless says that she became the youngest dean in her institution's history, describes a specific typhoon that damaged the campus, mentions receiving a national professional award the following year, and discusses a program she subsequently established.
None of those statements is a conventional identifier. Together, however, they may describe one publicly identifiable person.
The ICO specifically notes that anonymizing qualitative material can require careful human judgment because free text may remain identifiable through contextual information and information that can be linked to it. Its guidance describes techniques such as suppression and generalization when anonymizing interview transcripts.
Being Able to Single Someone Out Can Matter Even Before You Know Their Name
Privacy risk is not limited to the moment someone attaches a legal name to a record. If a person can be uniquely distinguished across records, tracked, profiled, or linked with additional information, the dataset may already permit meaningful targeting.
The ICO therefore treats singling out itself as relevant to identifiability. A person who can consistently be isolated in a dataset may remain identifiable even if the analyst has not yet discovered the person's conventional identity.
For researchers, this is a useful correction to the question “Can the AI tell me the participant's name?” A better question is whether the information allows an individual to be distinguished, linked, or ultimately connected with identity through reasonably available means.
AI Can Also Infer Information That Participants Never Explicitly Disclosed
Re-identification and inference are related but not identical. An AI system may fail to determine who a participant is while still inferring sensitive characteristics about that person or a narrowly defined group.
SACHRP has warned that combining large datasets can enable AI and machine-learning systems to infer information individuals did not knowingly disclose. This is one reason privacy analysis should not stop at conventional identifiers.
For example, behavioral or linguistic patterns might support predictions about characteristics not explicitly recorded in the research dataset. Whether such predictions are accurate is a separate methodological question, but their generation may still create ethical and privacy concerns.
Watch Out
Do not test a dataset's anonymity by asking a public AI service to identify your participants. Doing so may disclose the very participant information you are trying to protect. Re-identification risk should be evaluated through an approved privacy or disclosure-risk process, not by experimentally exposing confidential records to an unauthorized service.
Not Every Theoretical Re-Identification Scenario Is Equally Plausible
It would also be misleading to claim that any dataset can always be re-identified because sufficiently powerful AI exists. Privacy risk assessment requires more discipline than that.
The relevant question is usually whether identification is reasonably likely given the information available, the people who might receive it, the additional information realistically accessible to them, and the effort or resources required.
For example, a dataset containing broad age groups, generalized geographic regions, and common characteristics may present much lower risk than a dataset containing precise dates, detailed locations, rare diagnoses, exact occupational histories, and narrative quotations.
Strong anonymization aims to reduce realistic identification pathways rather than promise metaphysical impossibility.
Different Regulatory Frameworks Use Different Tests
Researchers should be cautious about treating one jurisdiction's definition of de-identification as universal. Under the U.S. Common Rule, for example, OHRP's guidance considers whether investigators can readily ascertain identities, including through coding systems. HIPAA applies its own specific de-identification requirements to protected health information. European and UK data-protection frameworks apply their own concepts of identifiability and anonymization.
A dataset can therefore be treated differently depending on the legal regime and who holds the additional information. This is one reason the ethical question of whether de-identification is sufficient for AI use should not be reduced to a single technical test.
06 · What This Means for You
Assess Re-Identification From the Recipient's Information Environment
When preparing participant data for AI processing, do not inspect the file only for obvious identifiers. Consider what makes each record distinctive and what information outside the dataset might overlap with it.
The appropriate level of scrutiny should reflect the consequences of identification. Detailed information about health, trauma, illegal activity, workplace disputes, stigmatized experiences, vulnerable populations, or politically sensitive topics warrants particular care.
A simple decision framework
If direct identifiers remain
The dataset has not been meaningfully anonymized merely because some identifiers were removed. Do not treat it as anonymous.
If combinations of variables describe rare or unique people
Consider suppression, generalization, aggregation, or another approved disclosure-control method while preserving the information genuinely needed for analysis.
If the dataset contains free text
Review contextual clues and distinctive events rather than relying only on automated removal of names and contact details.
If relevant external information is readily available
Include realistic linkage possibilities in the re-identification assessment.
If the consequences of identification would be serious
Use a correspondingly conservative risk threshold and seek appropriate privacy, ethics, statistical disclosure, or institutional expertise.
Most importantly, do not confuse de-identification with a guarantee. The objective is to reduce identification risk to a level justified by the research context and applicable requirements. If the residual risk remains substantial, the better question may be whether using AI is worth exposing participant data to that risk in the first place.
07 · A Quick Checklist
Before Treating Research Data as Safe From Re-Identification
Assess whether someone could realistically reconnect records to participants:
Remove or appropriately transform direct identifiers that are unnecessary for the research purpose.
Look for combinations of age, location, occupation, dates, roles, demographic characteristics, and other attributes that may single someone out.
Inspect free text for distinctive events, quotations, relationships, organizations, and biographical details.
Consider what external information a realistic recipient or adversary could access and link to the dataset.
Determine whether participant codes or other linkers remain and who can access the corresponding key.
Assess singling-out and linkage risks rather than testing anonymity solely by searching for names.
Apply stricter safeguards when re-identification could expose sensitive information or cause substantial harm.
Use the anonymization or disclosure-risk standard required by your jurisdiction, institution, ethics approval, or data agreement.
