03 · What You Need to Know
How Privacy and Security Requirements Apply to AI-Assisted Research
Why AI Introduces Additional Information-Handling Questions
Research information may pass through several systems before an AI output is generated.
A researcher may submit a prompt through a web interface, upload a file, connect an external database, or use an application programming interface.
Depending on the service and contractual arrangement, information may be processed, logged, retained, or accessed by different parties.
These arrangements can differ substantially across consumer, enterprise, institutional, and locally deployed systems.
Researchers should therefore evaluate the actual service and configuration being used rather than assume that every product offered by the same provider has identical data-handling terms.
Privacy Review and Information-Security Review Are Not the Same
Privacy review
Examines whether personal information processing is appropriate and authorized, including purpose, lawful basis, transparency, data minimization, disclosure, retention, and individual rights where applicable.
Information-security review
Examines the confidentiality, integrity, and availability of information and systems, including access controls, encryption, vulnerabilities, monitoring, and incident response.
A technically secure system may still process information for an unauthorized purpose.
Conversely, a processing activity may have a lawful basis but lack adequate technical safeguards.
Responsible AI research may require both assessments, depending on the information and proposed use.
Which AI Research Activities Commonly Warrant Review?
| AI-assisted activity |
Privacy or security concern |
Review consideration |
| Uploading identifiable interview transcripts |
Disclosure of personal or sensitive information |
Authorization, provider terms, access, retention, and safeguards |
| AI transcription of recorded interviews |
Identifiable voices and potentially sensitive speech |
Permitted processing, storage, and confidentiality protections |
| Analyzing institutional student records |
Restricted educational information |
Institutional authorization, data minimization, and security controls |
| Connecting AI to research databases |
Broad or unintended data access |
Permissions, integration security, logging, and access limitations |
| Using external AI APIs |
Transfer to another processing environment |
Contractual terms, data flows, authentication, and retention |
| AI processing of unpublished manuscripts |
Confidentiality or intellectual property obligations |
Permission to disclose and applicable contractual restrictions |
| AI analysis of genuinely public, non-sensitive data |
Potentially limited information protection concerns |
Confirm data status, permitted use, and system requirements |
| Locally deployed AI processing |
Local access, infrastructure, and configuration vulnerabilities |
Security configuration, permissions, maintenance, and institutional controls |
These activities do not necessarily require identical formal assessments. The appropriate review depends on the information, processing arrangement, and governing requirements.
Why Confidential Research Data Require Special Attention
Confidential research information may include interview transcripts, unpublished records, institutional datasets, identifiable observations, and sensitive participant disclosures.
Researchers may have promised participants that access will be limited to specified individuals or authorized research personnel.
Uploading the information to an external AI provider may introduce additional parties or processing activities not contemplated by those commitments.
Even where the provider maintains strong security controls, the proposed transfer must still be authorized.
Researchers should examine the original consent arrangements, research protocol, data management plan, and relevant institutional obligations.
Scientific convenience does not independently justify changing confidentiality commitments.
Does De-Identification Make AI Processing Safe?
De-identification can reduce some privacy risks, but it does not necessarily eliminate them.
Removing names, student numbers, or other direct identifiers may leave contextual details that permit reidentification.
For example, an interview excerpt describing a unique institutional role, location, and personal experience may remain identifiable even after the participant's name is removed.
Researchers should assess the likelihood of reidentification in the intended processing environment.
The applicable legal distinction between anonymous, de-identified, and pseudonymized information should also be considered.
A dataset should not be treated as unrestricted merely because its direct identifiers have been replaced with codes.
What Should Researchers Examine in AI Provider Terms?
AI providers may offer different contractual and technical arrangements across their services.
Researchers should determine which terms apply to the exact product, account type, and processing method they intend to use.
Relevant questions include whether submitted information may be used for model improvement, how long it is retained, who may access it, and whether subprocessors are involved.
Researchers should also examine deletion arrangements, available audit records, security commitments, and restrictions on processing particular information.
A statement that data are not used for model training addresses one concern but does not necessarily resolve retention, access, confidentiality, or transfer obligations.
Why Data Retention Matters
Research data may be subject to retention schedules established by institutional policy, ethics approval, funding agreements, or applicable law.
An AI provider may retain prompts, uploaded files, outputs, or technical logs under separate terms.
Researchers should establish whether these arrangements are compatible with the project's obligations.
For example, a research team may be required to delete identifiable audio recordings after a specified period while retaining de-identified analytical records for longer.
The AI processing arrangement should not create uncontrolled copies that undermine those requirements.
Researchers should verify the available deletion and retention controls rather than assume that deleting a conversation from an interface removes every associated record.
How Should Researchers Assess External Data Transfers?
Using a cloud-based AI system may involve transferring information outside the researcher's institution or jurisdiction.
Depending on the applicable legal framework, international transfers may introduce additional requirements.
Researchers should identify where processing occurs to the extent relevant and available, which entities receive the information, and what contractual or organizational safeguards apply.
The appropriate legal basis and transfer mechanisms depend on the jurisdiction and circumstances.
Researchers should not assume that a provider's international availability establishes compliance with every country's data protection requirements.
What Is a Data Protection Impact Assessment?
A data protection impact assessment, or DPIA, is a structured process used under certain privacy frameworks to evaluate risks associated with personal data processing and identify appropriate safeguards.
Under the European Union General Data Protection Regulation, Article 35 addresses DPIAs for processing likely to result in a high risk to individuals' rights and freedoms.
Not every AI-assisted research activity automatically requires a DPIA.
The requirement depends on the applicable law and the nature, scope, context, and purposes of processing.
Researchers should consult the responsible privacy or data protection authority within their institution to determine whether a formal assessment is necessary.
What Does the Philippine Data Privacy Act Require Researchers to Consider?
In the Philippines, Republic Act No. 10173 establishes the Data Privacy Act of 2012 and provides the legal framework for personal information processing.
Relevant considerations include lawful processing, transparency, legitimate purpose, proportionality, and appropriate protection of personal information.
Researchers should also consider applicable National Privacy Commission issuances and institutional policies.
The law does not establish a universal requirement that every use of generative AI undergo a separate research-specific privacy assessment.
However, AI processing involving personal or sensitive personal information may require careful evaluation of the lawful basis, data protection obligations, and security arrangements.
Institutions should seek guidance from their data protection officer or qualified legal personnel when the applicable requirements are uncertain.
Why Access Controls Matter in AI Research
AI systems may be connected to shared drives, databases, document repositories, or other institutional resources.
These integrations can introduce risks when permissions are broader than necessary for the research task.
For example, an AI application intended to summarize one research dataset should not automatically receive access to unrelated confidential records.
Researchers should follow the principle of least privilege, granting only the access necessary for the authorized operation.
Permissions should be reviewed when users, tasks, or connected systems change.
Access control is particularly important when AI tools can retrieve information or perform actions across multiple systems.
What Security Controls Should Be Evaluated?
Relevant controls depend on the deployment environment and information sensitivity.
They may include encryption in transit and at rest, authentication, role-based access, audit logging, vulnerability management, and incident response procedures.
Researchers should also consider the security of application programming interfaces, connected repositories, and local devices used to access AI services.
A security review should examine whether the controls are appropriate for the actual research workflow.
General claims of enterprise-grade security are not a substitute for assessing the relevant configuration and institutional requirements.
Can Locally Deployed AI Avoid Privacy and Security Review?
Local deployment may reduce the need to transfer information to an external provider, but it does not automatically eliminate privacy or security risks.
Local systems may still have vulnerabilities, inappropriate access permissions, insecure storage, or inadequate maintenance.
Researchers should evaluate who can access the system, where information is stored, whether external connections are enabled, and how updates and logs are managed.
Local deployment may be advantageous for some confidential research tasks, but its suitability depends on the actual safeguards.
Security should be assessed according to the system's implementation rather than its physical location alone.
How Do AI Integrations Create Additional Risks?
Some AI systems can retrieve documents, query databases, or interact with connected applications.
These capabilities may expand the information available to the system beyond the content explicitly entered into a prompt.
Researchers should examine the scope of connected permissions and whether the integration could expose unrelated or restricted records.
They should also consider risks associated with untrusted retrieved content, unintended actions, and insecure configuration.
Where the AI system can take actions rather than merely generate text, additional operational safeguards may be necessary.
The review should address the complete integration, not only the underlying language model.
How Should Researchers Handle Security Incidents?
Research teams should know how to respond if confidential information is uploaded to an unauthorized service, exposed through an integration, or accessed inappropriately.
The response should follow applicable institutional incident procedures and legal obligations.
Depending on the circumstances, this may involve containing the exposure, preserving relevant records, notifying responsible security or privacy personnel, and assessing whether additional reporting is required.
Researchers should not assume that deleting a prompt or uploaded file necessarily resolves an unauthorized disclosure.
Incident response should be planned before consequential AI processing begins.
How Does Privacy Review Differ From Research Ethics Approval?
Research ethics review and privacy assessment may both address participant information, but their responsibilities are not identical.
An ethics committee may evaluate confidentiality safeguards as part of a broader assessment of participant protection.
A privacy or security review may examine technical, legal, and organizational details that require specialist expertise.
Approval from one authority does not necessarily satisfy the requirements of the other.
Researchers should identify which institutional offices are responsible for the proposed processing and coordinate their determinations where necessary.
When Should Researchers Avoid an AI Service?
Researchers should not submit restricted information to an AI system when they cannot establish that the processing is authorized and adequately protected.
Unclear provider terms, incompatible retention practices, insufficient access controls, or unresolved contractual restrictions may prevent responsible use.
The appropriate response may involve obtaining additional information, negotiating an approved processing arrangement, using an institutionally controlled environment, or selecting another method.
The scientific value of an AI output does not independently justify exposing information that researchers are obligated to protect.
