01 · The Question
Does De-Identification Make Participant Data Safe to Use With AI?
You know that uploading identifiable participant data to a public AI service may create privacy and confidentiality problems. So you remove the names, email addresses, participant IDs, and other obvious identifiers first. Can you now upload the dataset?
Possibly, but “de-identified” can describe very different levels of protection. A spreadsheet with names replaced by codes is not equivalent to a dataset from which participants can no longer reasonably be identified. The distinction matters because the remaining data may still reveal who particular participants are, especially when records contain unusual combinations of characteristics, detailed narratives, locations, dates, or other distinctive information.
Before treating de-identification as permission to use an AI tool, you need to determine what your de-identification process actually accomplished.
03 · What You Need to Know
What De-Identification Actually Changes
“De-Identified” Does Not Have One Universal Meaning
Researchers often use de-identified as a broad description for data from which identifying information has been removed or obscured. The term is useful, but it can conceal important differences.
In one dataset, de-identification might mean replacing participant names with codes while retaining a separate key that connects those codes to identities. In another, it might involve removing or generalizing enough information that participants are no longer reasonably identifiable. Those two datasets do not present the same privacy risk.
Legal terminology also varies by jurisdiction. Under UK data-protection guidance, for example, replacing direct identifiers while retaining separately held information that can restore the connection is generally pseudonymisation. The resulting information remains personal data. By contrast, anonymisation aims to make people no longer identifiable, in which case the resulting anonymous information is not personal data under that regime.
Pseudonymized data
Direct identifiers are replaced, removed, or separated, but the information can still be attributed to individuals using additional information. Depending on the applicable legal framework, it may remain personal data.
Anonymous data
Individuals are no longer identifiable under the applicable standard, taking account of relevant contextual information and realistic means of identification.
Do not assume that your discipline's informal use of “anonymous” or “de-identified” matches the definition used by your ethics committee, institution, data-protection authority, or applicable law.
Removing Names Is Only the Beginning
A name is an obvious identifier, but it is far from the only way a person can be recognized. Identifiability can arise from combinations of information.
Imagine a dataset containing no names but recording each participant's age, occupation, institution, municipality, gender, and a detailed description of an unusual event. If only one person fits that combination, removing the name may accomplish very little.
The UK Information Commissioner's Office advises organizations to consider identifiability broadly rather than focusing only on obvious identifiers. A person may be identifiable even when their name is unknown, and whether particular information identifies someone depends on context.
The U.S. Department of Health and Human Services takes a similarly contextual approach within HIPAA's specific health-data framework. Its de-identification guidance recognizes that information can potentially be linked with other data and that even properly de-identified health information retains some identification risk. HIPAA provides formal methods for satisfying its own de-identification standard, but those rules should not be generalized as a universal de-identification method for every dataset, jurisdiction, or research context.
Free Text Can Be Much Harder to De-Identify Than a Spreadsheet
Structured data make many identifiers relatively easy to locate. A researcher can remove columns containing names, email addresses, telephone numbers, or participant numbers. Unstructured data are harder.
An interview transcript might contain no participant name yet describe a person's workplace, family relationships, rare medical condition, position, neighborhood, or distinctive experience. A survey's open-ended response can create the same problem.
HHS specifically notes in its HIPAA guidance that identifiers can occur in free-text records and that information-rich clinical narratives may provide context allowing a patient to be identified.
This is one reason the permissibility of uploading interview transcripts to generative AI deserves separate scrutiny from uploading a clean numerical dataset.
A Participant Code Does Not Necessarily Make Data Anonymous
Suppose you replace “Maria Santos” with “P017” and keep a separate file showing that P017 is Maria Santos. That can be a valuable security measure because the analytical dataset no longer contains the direct identifier. It does not necessarily mean the underlying participant information has become anonymous.
If the research team can reconnect P017 to Maria using the key, the data remain linkable within the research environment. Under frameworks such as the UK GDPR, that is pseudonymisation rather than anonymisation, and pseudonymized information remains personal data for a party able to use the additional information.
Whether an external recipient that does not possess the key would regard the same dataset as anonymous may require a context-specific assessment. This is precisely why researchers should avoid treating data categories as intrinsic properties of files. Identifiability can depend partly on who receives the information and what additional information is realistically available to them.
De-Identification Is About the Dataset the AI Provider Receives
If you intend to send data to an external AI provider, assess the information from the perspective of that disclosure. Ask what information the provider receives, what other data may accompany the upload, and whether unnecessary contextual details remain.
This means inspecting more than the visible cells in a spreadsheet. Depending on the file type and workflow, relevant information may also occur in free-text fields, filenames, labels, annotations, images, audio, document properties, or other accompanying content.
A useful principle is data minimization: if the AI task does not require a variable or contextual detail, consider removing or generalizing it before transmission. Under the Philippine Data Privacy Act, for example, personal information processing is subject to proportionality principles requiring information to be adequate and not excessive for the purpose for which it is processed.
Successful De-Identification Can Change the Legal Analysis
If information has genuinely been rendered anonymous under the applicable legal standard, some privacy-law obligations that apply specifically to personal data may no longer apply to that information. For example, the UK ICO states that anonymous information is not personal data, while the Philippine National Privacy Commission has likewise explained in an advisory opinion that truly anonymized information falls outside the application of the Data Privacy Act's data-protection principles.
That distinction can be consequential. But it should not be reversed into the assumption that any dataset called “de-identified” has automatically crossed that threshold.
Watch Out
Do not rename a dataset “anonymous” and then reason from the label. Determine whether the transformation actually satisfies the relevant standard for anonymization or de-identification. A coded participant file may be much safer than the original while still remaining identifiable data.
De-Identification Does Not Automatically Authorize the AI Use
Even a strong de-identification process answers only part of the question. Research data may remain subject to commitments made to participants, ethics conditions, data-use agreements, funder requirements, institutional rules, contractual restrictions, or discipline-specific standards.
For example, an approved protocol might restrict research files to designated institutional systems. A data-use agreement might prohibit transmission to unapproved services. A consent process may have described specific secondary uses or confidentiality arrangements.
That is why the broader question of whether participant data can be uploaded to an AI tool requires more than determining whether identifiers have been removed.
06 · What This Means for You
Treat De-Identification as a Risk-Control Process, Not a Checkbox
Before using de-identified participant data with an AI tool, document what you removed, transformed, generalized, or retained. Then assess what someone receiving the resulting dataset could realistically learn about individual participants.
This assessment should be proportionate to the data. A table containing broad age bands and aggregated questionnaire scores presents a different problem from detailed life-history interviews, genomic data, photographs, voice recordings, or records describing rare conditions.
A simple decision framework
If you only removed names and obvious contact details
Do not assume the dataset is anonymous. Check indirect identifiers, combinations of variables, and free-text information.
If participants are represented by codes linked to a separate key
Treat the dataset as potentially pseudonymized or otherwise linkable rather than automatically anonymous.
If the dataset contains detailed narratives or rare characteristics
Assess contextual identifiability and whether unnecessary details can be removed or generalized before AI processing.
If the data appear genuinely anonymous
Verify that conclusion under the relevant institutional and legal standard, then separately check whether other research restrictions still govern the proposed AI use.
If you are uncertain whether individuals can be re-identified
Do not resolve the uncertainty merely by calling the dataset de-identified. Seek appropriate privacy, ethics, statistical disclosure, or institutional expertise.
The next question is equally important: even when the probability of identifying participants has been substantially reduced, whether de-identification makes the AI use ethically acceptable depends on more than identifiability alone.
07 · A Quick Checklist
Before Using De-Identified Participant Data With AI
Before uploading de-identified data, check:
Identify which direct identifiers were removed, replaced, masked, or retained.
Check for indirect identifiers such as precise locations, exact dates, occupations, institutional roles, rare characteristics, and unusual combinations of variables.
Determine whether a code, key, or other separately held information can reconnect records to participants.
Inspect free-text responses and narratives rather than assuming de-identification of structured fields is sufficient.
Remove or generalize information the AI task does not require.
Assess whether other reasonably available information could make participants identifiable.
Verify the applicable definition of anonymized, pseudonymized, or de-identified data rather than relying on informal terminology.
Confirm that consent, ethics approval, institutional policy, data-use agreements, and provider conditions permit the proposed AI processing.
