03 · What You Need to Know
Privacy Risk Should Buy You Something Methodologically Useful
“AI Would Be Faster” Is a Benefit, but Not Always a Strong One
Efficiency matters in research. Automated transcription can save days. AI-assisted coding can help researchers explore large qualitative datasets. Automated extraction can make thousands of clinical records tractable.
But convenience should be weighed against what the workflow requires from participants' data.
If AI saves ten minutes while requiring identifiable interview transcripts to leave an approved environment, the trade-off looks different from a carefully governed AI system that makes analysis of a dataset possible when manual processing would be practically infeasible.
There is no universal numerical threshold. The principle is proportionality: greater exposure should require a correspondingly meaningful research justification.
Ask Whether Participant Data Are Necessary for the AI Task at All
Many researchers disclose real data when the AI does not actually need them.
If you want help writing R or Python code, provide variable names and synthetic data. If you want feedback on a coding framework, create fictional excerpts. If you need assistance interpreting a statistical method, describe the design rather than uploading the participant dataset. If you want help drafting a table shell, provide the structure rather than the results file.
This is often the easiest privacy intervention because it eliminates the disclosure rather than trying to secure it.
AI requires the participant data
The research task depends on information contained in actual participant records, and a meaningful analytical benefit cannot reasonably be achieved with a lower-risk representation.
AI requires information about the research
The researcher needs methodological, coding, writing, or technical assistance but can obtain it without transmitting actual participant records.
Confusing those two situations causes a surprising amount of unnecessary exposure.
Use the Least Sensitive Representation That Can Answer the Question
Even when actual research information is needed, the full raw dataset may not be.
A task might work with:
- aggregate statistics instead of participant-level records;
- selected variables instead of the entire dataset;
- de-identified excerpts instead of complete transcripts;
- transcripts instead of original audio recordings;
- cropped or transformed images instead of identifiable originals;
- broad categories instead of precise locations or ages; or
- a small authorized sample rather than every participant record.
This is the practical logic behind data minimization. Under the Philippine Data Privacy Act, for example, proportionality requires personal information to be adequate and not excessive in relation to the purposes for which it is processed. Comparable minimization principles appear in other data-protection frameworks.
Sensitive Data Raise the Price of a Small Benefit
The consequences of disclosure should influence how much benefit is required to justify AI processing.
A dataset containing broad, non-sensitive questionnaire responses may permit a relatively low-risk AI workflow after appropriate safeguards. Identifiable trauma interviews, genomic information, medical histories, illegal activity, sexual information, facial images, or recordings from vulnerable participants create a different risk profile.
The more serious the potential consequences of unauthorized disclosure or re-identification, the less persuasive “it will make analysis easier” becomes.
This is particularly important with clinical research data and other records in which both identifiability and sensitivity can be high.
Some Data Are Difficult to De-Identify Without Destroying Their Research Value
De-identification works well for some datasets and poorly for others.
Detailed qualitative narratives derive much of their value from context. Video research may depend on faces, gestures, movement, and interaction. Voice research may require acoustic characteristics. Rare-disease research may involve small populations whose characteristics are inherently distinctive.
If removing identifying information destroys the phenomenon you need to analyze, do not pretend the transformed dataset solves the problem. Either use an environment authorized for identifiable data or reconsider whether AI belongs in the workflow.
An Approved Alternative With Comparable Value Weakens the Case for Riskier AI
Suppose an institution already provides a secure transcription service approved for confidential research recordings. A researcher nevertheless prefers a public generative AI service because its interface is slightly faster.
If the scientific outcome is essentially the same, the additional privacy exposure of the less controlled service may have little justification.
Risk-benefit reasoning should compare realistic alternatives, not compare AI with doing nothing.
Local or Institution-Controlled AI Can Change the Trade-Off
Sometimes the problem is not AI itself but the architecture through which it is accessed.
A model running locally or inside an institutionally controlled environment may allow researchers to obtain AI assistance without transmitting participant information to an external provider. Depending on the implementation, this can reduce concerns about third-party processing, international transfer, retention, and provider reuse.
Local deployment still requires security, access controls, ethics compatibility, and methodological validation. It simply changes several components of the privacy equation.
De-Identification Should Reduce Risk, Not Serve as a Ritual
If participant data can be effectively de-identified while retaining the information needed for analysis, AI processing may become substantially easier to justify.
But de-identified data are not automatically anonymous, and removing names from a rich dataset may leave considerable residual identification risk.
Researchers should therefore ask how much the transformation actually reduces realistic risk rather than whether a de-identification step appears somewhere in the workflow.
Unclear Provider Practices Are Themselves a Reason to Stop
If you cannot determine whether the provider retains participant information, uses it for model training, permits deletion, employs subprocessors, or processes data in other jurisdictions, you cannot meaningfully assess the exposure.
Uncertainty should not be silently converted into an assumption that the provider probably handles data safely.
For low-risk public information, uncertainty may be tolerable. For confidential participant data, the inability to establish basic processing conditions can itself make the service unsuitable.
International Processing Can Add Burden Without Adding Scientific Value
If using an AI service introduces cross-border transfer assessments, contractual safeguards, data-residency review, and additional institutional approvals, ask whether the research actually benefits enough to justify those steps.
There may be good reasons to proceed. There may also be an institutionally approved alternative that avoids the issue entirely.
Where location matters, determine where the AI provider actually processes participant data before deciding whether the governance burden is worthwhile.
AI's Methodological Weaknesses Belong in the Risk-Benefit Calculation Too
Privacy exposure might be justified if AI produces a meaningful analytical advantage. That justification becomes weaker if the method itself is poorly validated.
A generative model that inconsistently codes interviews, invents clinical details, misclassifies images, or produces results that cannot be adequately validated may create both privacy risk and questionable evidence.
The appropriate comparison is therefore not:
AI benefit versus privacy risk.
It is closer to:
validated research benefit from this AI workflow versus the privacy, confidentiality, methodological, governance, and opportunity costs it introduces.
That formulation is less catchy. Research methods occasionally have this unfortunate relationship with accuracy.
Participant Expectations Can Matter Even When Formal Permission Exists
A technically permissible AI use may still be difficult to defend if it departs sharply from the context in which participants contributed highly sensitive information.
SACHRP's recommendations on AI research emphasize gaps that can arise between regulatory compliance and adequate protection of participants, particularly with de-identified data, secondary uses, inferred information, bias, and group harms.
That is why legal or technical de-identification does not settle the entire ethical question.
“Can We?” and “Should We?” Need Separate Answers
An institution may approve an AI service. The provider may offer strong contractual safeguards. Participants may have consented appropriately. The data may be adequately protected.
You can still decide that AI contributes too little to justify introducing it.
Good research governance is not a contest to find the maximum amount of technically permissible data processing. Sometimes the best privacy control is simply not creating the additional data flow.
Watch Out
Do not treat institutional approval of an AI platform as a recommendation to upload participant data. Approval means a service may be available under specified conditions. You still need a research reason for using participant information with it.
