Opens in a new tab

Home / Resources / AI in Research

De-Identifying Research Data Before Using AI Tools:
Why Removing Names Is Not Enough

De-identifying research data for AI tools is more than deleting participant names. Dates, uncommon characteristics, combinations of demographic fields and open-text descriptions can also identify individuals. Removing direct identifiers may reduce risk, but it does not automatically authorize external processing or establish that data are anonymous (Rocher et al., 2019; HHS, 2012).

AI in Research
TERMSDirect, indirect IDs
NEEDSRe-identification review
REPORTApprovals and transfers
GUIDELINEHIPAA, GDPR
COMMON ERRORNames removed, done

The first question is whether a particular research dataset may be processed by a particular service under the study’s approvals and agreements. Technical de-identification and permission to upload are separate decisions.

Direct identifiers versus indirect identifiers

A data table with the name column removed but other columns still highlighting a possible identity.
Data element Example of the risk Possible governance response
Names, emails, record numbers Directly point to a person Remove or replace under an approved procedure
Exact dates and rare events Can uniquely identify a participant Assess date generalization or other approved transformation
Location and demographics Combinations may be distinctive Evaluate singling-out and linkage risk
Open-text interview material People may name workplaces, relatives or events Structured human review of text and context
Medical histories Rare diagnosis combinations may be distinctive Data minimization, access controls and expert risk assessment
Linking keys Reconnect coded records to identities Keep separately with appropriate restrictions

The examples are illustrative. The right transformation depends on analytic purpose, disclosure risk and applicable requirements. Excessive removal can also destroy information necessary for valid research.

What does published research show about re-identification?

Rocher, Hendrickx and de Montjoye (2019) studied re-identification in incomplete datasets and showed that apparently incomplete records can still carry meaningful identification risk. Their results challenge a simple “release it after removing names” model. That finding does not imply that every dataset is inevitably identifiable; it supports using an explicit risk assessment rather than relying on a single redaction step.

When assessing a proposed transfer, consider whether someone could link remaining variables to external information, whether the data are rare or highly sensitive, and whether an AI provider or subcontractor would obtain the records.

HIPAA de-identification is a specific US regulatory concept

For data subject to the HIPAA Privacy Rule, HHS recognizes two approaches to satisfying the de-identification standard: Safe Harbor and Expert Determination (HHS, 2012). Safe Harbor involves removing listed identifier categories and meeting the rule’s knowledge condition. Expert Determination involves a qualified expert applying accepted principles to determine that re-identification risk is very small under specified conditions.

These standards apply within HIPAA’s scope. They should not be presented as a universal permission to send all university research data to a commercial AI tool. A study may be subject to additional contracts, research consent, institutional policies or other legal obligations.

Why pseudonymization is not the same as anonymity

Replacing a participant’s name with a study ID creates pseudonymized or coded records when a key or other feasible means may still allow identification. Under GDPR, pseudonymized personal data may still be personal data. Whether data are truly anonymous depends on the realistic ability to identify people and the relevant legal framework. Seek qualified institutional privacy guidance when making that determination.

Even if records meet an applicable de-identification test, their further use must also respect research agreements and the limits of the approved purpose.

Practical questions before de-identifying research data for AI tools

  1. Who controls the data? Identify the sponsor, institution, data steward and any contractual restrictions.
  2. What did participants consent to? Review the research protocol and IRB/ethics materials for permitted processing.
  3. Which service and account? Check the exact provider, product tier, retention and onward-processing arrangements, not marketing claims about “AI” in general.
  4. What is actually necessary? Often AI can draft code from a variable dictionary or synthetic test data without receiving real records.
  5. How is disclosure risk evaluated? Confirm the method and decision-maker responsible for de-identification.
  6. How are processing and deletion documented? Keep the approved access, transfer and retention record where applicable.

This is a research-governance checklist, not individualized legal advice.

What can researchers do without uploading the dataset?

An analyst can describe a dataset’s structure, including fictional example rows and variable meanings, and ask AI to suggest a general script. The final script can be inspected and executed on the actual dataset in the research institution’s approved environment. For sensitive projects, this preserves much of the programming benefit without transferring individual-level records to a third-party chatbot.

Frequently asked questions

No. Indirect identifiers and combinations of variables can permit re-identification; the relevant standard and study context determine what is sufficient.

No. Meeting a de-identification standard, when applicable, is not the same as satisfying contracts, IRB requirements, institutional policies and specific vendor controls.

First check whether the transcript is genuinely de-identified and whether the exact tool is approved. Narrative details may disclose identities even without names.

Quantitative Data AnalysisWorking with participant or patient data? Project information is not published, and an NDA is available where required before sensitive materials are shared.

Request an Assessment

Quantitative Data Analysis

Quantitative Data Analysis at OnlineSPSS is carried out by in-house PhD-level statisticians. Project information is not published, and an NDA is available where required before sensitive materials are shared. Researcher names and institutions are not used publicly without permission.

References

Last reviewed November 2026 by the OnlineSPSS Statistical Team. Worked examples use hypothetical values and are not client data.

NEXT STEP

Need this done for
your study?

Share your research in about a minute and receive a complimentary assessment the same day: scope, timeline and one fixed fee.

No Commitment Required

Strictly Confidential

Fixed Project Fee

In-House PhD-Level Statisticians

Same-Day Response

Serving Researchers Since 2014