03 · What You Need to Know
How AI Use Can Affect Confidentiality and Research Data Protection
Confidential Research Data Are Not Limited to Names and Contact Details
Researchers often associate confidential information with obvious identifiers such as names, student numbers, addresses, or medical record numbers. Yet research confidentiality can extend much further.
Interview transcripts may contain distinctive personal experiences, employment histories, institutional roles, locations, or combinations of circumstances that permit someone to recognize a participant. A dataset may also contain unpublished institutional information, proprietary methods, commercially sensitive records, or material subject to contractual restrictions.
Confidentiality concerns therefore extend beyond whether a file contains a column labeled "Name."
The relevant question is whether the information is subject to restrictions on access, use, processing, or disclosure. Those restrictions may arise from informed consent, research ethics approval, institutional policy, contractual agreements, or applicable law.
Uploading Data to AI May Introduce an Additional Processing Party
When researchers submit material to an externally hosted AI service, information may be processed through infrastructure operated by a provider or its authorized subprocessors.
The precise arrangement depends on the service, subscription, deployment model, and applicable agreement. Some services provide contractual data-processing protections and controls over retention or model training. Others may offer different arrangements.
Researchers should establish who receives or processes the information, for what purposes, under which contractual conditions, and with what technical protections.
Even where the provider does not use customer data to train its models, questions may remain about retention, access, logging, subprocessors, and permitted data transfers.
A promise not to train on submitted information addresses one potential concern. It does not, by itself, establish comprehensive confidentiality protection.
Why Removing Names May Not Be Enough
De-identification and anonymization are related but distinct concepts.
De-identification
Reducing or removing identifying information to lower the likelihood that individuals can be recognized, while acknowledging that residual identification risks may remain.
Anonymization
Processing information so that individuals are no longer identifiable under the applicable legal and technical standard, taking account of reasonably available means of identification.
Replacing names with participant codes may reduce direct identification, but distinctive combinations of attributes can still reveal identities.
For example, an interview excerpt referring to the only female dean in a small institution, a specific disciplinary incident, and a particular academic year may be recognizable even without a name.
Researchers should assess contextual identifiers, linkability with other information, and the possibility of reidentification. Pseudonymized data may remain personal data under applicable law.
Where effective anonymization cannot be established, the information should continue to be handled according to the protections required for identifiable or potentially identifiable research data.
Does Participant Consent Permit AI Processing?
Consent to participate in research does not necessarily authorize every subsequent form of data processing.
A participant may agree to have an interview recorded and analyzed by an approved research team without agreeing to have the transcript transferred to an external generative AI provider.
Whether additional consent, an ethics amendment, or another lawful authorization is required depends on the original consent materials, approved protocol, applicable law, and nature of the proposed processing.
In some circumstances, processing may rely on a lawful basis other than consent. Researchers should not assume, however, that an existing lawful basis automatically covers every new AI workflow.
When the proposed use materially differs from the approved arrangements, consult the responsible ethics or data protection authority before transferring the information.
What Does the Philippine Data Privacy Act Mean for AI-Assisted Research?
For research involving personal information subject to Philippine law, Republic Act No. 10173, the Data Privacy Act of 2012, and its implementing rules provide an important legal framework.
The Act establishes general data privacy principles, including transparency, legitimate purpose, and proportionality. It also addresses lawful processing, sensitive personal information, security measures, and accountability.
Research-related provisions and exceptions must be interpreted in their proper legal context. The fact that information is collected for academic research does not create a universal exemption from data protection obligations.
Researchers should identify the applicable lawful basis, determine the responsibilities of the relevant personal information controller and processor, and assess whether external AI processing is consistent with the authorized purpose and required safeguards.
The National Privacy Commission provides official guidance and issuances relevant to these obligations. Institutional data protection officers should be consulted where necessary.
For international projects, additional requirements may arise under other jurisdictions, including the European Union's General Data Protection Regulation where applicable.
What Protections Should Be Established Before Using AI?
The necessary safeguards depend on the sensitivity of the information and the proposed processing arrangement.
Potential measures include institutional approval, a suitable data-processing agreement, defined access permissions, encryption, retention controls, documented subprocessors, restrictions on secondary use, and procedures for responding to security incidents.
These measures should be evaluated together. Encryption during transmission does not eliminate risks associated with provider access or authorized processing after the information reaches the service.
Likewise, a private account does not necessarily provide the same protections as an institutionally managed environment governed by a negotiated agreement.
| Protection |
What it addresses |
What still needs checking |
| Institutional authorization |
Whether the proposed use is permitted |
Scope of approval and any conditions |
| Data-processing agreement |
Provider responsibilities and contractual controls |
Permitted purposes, subprocessors, retention, and transfers |
| Access controls |
Who can access research information |
Administrative privileges, authentication, and audit records |
| Encryption |
Protection of information in specified states |
Key management and access after decryption |
| Retention and deletion controls |
How long information is kept |
Logs, backups, exceptions, and contractual commitments |
| Data minimization |
Limiting information exposed to the system |
Whether remaining details are still identifying or restricted |
No single safeguard establishes that a workflow is appropriate. Researchers should evaluate whether the combined protections meet the project's actual obligations.
When Is a Locally Hosted AI Model Safer?
A locally deployed model may reduce some risks associated with transferring research information to an external provider, particularly when processing occurs within an approved institutional environment.
However, local deployment is not automatically secure.
The system may still contain vulnerabilities, generate logs, permit unauthorized access, or depend on external services for updates, monitoring, or other functions. Researchers must also consider model reliability, system maintenance, and the institution's capacity to manage the environment.
A properly governed local deployment may be preferable for some confidential tasks, but its suitability should be established through an actual security and research assessment.
Can Synthetic Data Be Used Instead?
Synthetic or fictional data can be useful for testing prompts, exploring code structures, and evaluating preliminary workflows without exposing actual participant records.
For example, researchers may create fictional interview excerpts to test whether an AI system can produce a desired output format before considering any authorized use with real data.
However, synthetic data generated from confidential datasets can sometimes retain sensitive patterns or information. Their privacy properties should not be assumed merely because they are labeled synthetic.
More importantly, successful performance on fictional examples does not establish validity on actual research data. Differences in language, context, complexity, and population characteristics may affect results.
Synthetic material is therefore useful for preliminary development but should not be treated as a complete substitute for task-specific validation.
What if the Researcher Has Already Uploaded Confidential Data?
If potentially unauthorized processing has already occurred, the researcher should preserve relevant facts and promptly follow the institution's applicable incident-reporting procedures.
Relevant information includes what was submitted, when it was submitted, which service was used, the account and contractual arrangement, and whether the information contained personal or otherwise restricted material.
Researchers should not assume that deleting a conversation necessarily removes every retained copy or eliminates an earlier disclosure.
The institution's data protection, information-security, ethics, or legal personnel may need to determine whether containment, provider communication, participant notification, or regulatory reporting is required.
These decisions depend on the actual incident and applicable obligations. Researchers should avoid independently promising that no breach occurred before an appropriate assessment has been completed.
