03 · What You Need to Know
How to Evaluate AI Data-Use Terms Before Using a Research Tool
Why Data-Use Terms Matter in Research
Research information is often collected or obtained under specific conditions. Participants may agree to particular uses of their information. Institutions may restrict access to unpublished datasets. Research collaborators may exchange materials under confidentiality agreements.
When a researcher submits such information to an external AI service, the provider's processing arrangements become relevant to whether those original obligations can still be satisfied.
Terms of service, privacy notices, data-processing agreements, security documentation, and product-specific settings may collectively describe what happens to submitted information.
These documents serve different purposes. A public privacy notice may describe general practices, while a contractual data-processing agreement may establish specific obligations for an institutional customer.
Researchers should therefore avoid relying on a single marketing statement when the intended activity involves restricted research material.
What Information Should the Provider Explain?
The necessary information depends on the sensitivity of the data and the intended processing. Nevertheless, several questions are commonly relevant.
| Data-use issue |
Question to ask |
Why it matters |
| Processing purpose |
For what purposes may the provider use submitted information? |
Processing must be compatible with the researcher's authorization and applicable obligations. |
| Model training |
May inputs or outputs be used to develop or improve models? |
Secondary use may exceed the permitted research purpose. |
| Retention |
How long are inputs, outputs, and related logs retained? |
Retention may conflict with confidentiality or data minimization requirements. |
| Human access |
Under what circumstances can provider personnel access submitted material? |
Access arrangements affect confidentiality and accountability. |
| Subprocessors |
Which other organizations may process the information? |
Additional processing parties may create contractual and security obligations. |
| Data location |
Where may information be stored or processed? |
Cross-border processing may engage additional legal or contractual requirements. |
| Deletion |
What happens when the user deletes a file or conversation? |
Interface deletion may differ from deletion of logs, backups, or other retained copies. |
| Security |
Which safeguards protect information during transmission, storage, and processing? |
Security controls must be appropriate to the information and intended use. |
| Contractual responsibility |
What obligations does the provider accept under the applicable agreement? |
Accountability cannot be established solely from technical features. |
The absence of a clear answer does not prove that the provider mishandles information. It does mean the researcher may lack sufficient evidence to justify transferring restricted material.
Why a Privacy Policy Is Not Always Enough
A privacy policy may describe how a company processes information about website visitors, account holders, or service users. It may not fully explain the contractual treatment of research content submitted through every product or subscription tier.
For example, a provider may offer different arrangements for individual accounts, institutional subscriptions, application programming interfaces, and privately deployed systems.
The relevant protections may depend on account configuration, an enterprise agreement, administrative settings, or a separately executed data-processing agreement.
Researchers should identify which documents govern their actual service rather than assuming that a protection advertised for one product applies to every account.
Where the relationship between the documents is unclear, institutional procurement, information-security, or legal personnel may need to obtain clarification.
Does a Promise Not to Train Models on Your Data Resolve the Problem?
No. Restrictions on model training address one possible secondary use of submitted information, but they do not establish every relevant aspect of data handling.
A provider may refrain from training models on customer content while retaining information for specified operational purposes, processing it through subprocessors, or permitting limited access under defined circumstances.
Those arrangements may be legitimate, but their suitability depends on the researcher's obligations and the provider's actual commitments.
No model training
A commitment concerning whether submitted information is used to train or improve specified AI models.
Confidential processing
A broader set of restrictions and safeguards governing access, permitted purposes, retention, disclosure, and protection of submitted information.
Researchers should not treat these concepts as interchangeable. A no-training commitment may be valuable, but it is not a complete data protection assessment.
Why Data Retention Deserves Particular Attention
Data retention refers to how long information remains stored or otherwise preserved within the provider's systems.
AI services may retain conversation histories, uploaded files, generated outputs, operational logs, or other records under different arrangements. Some services provide user controls or contractual retention options, while others may apply standard retention periods.
Researchers need to distinguish deletion from the visible interface from deletion across relevant systems.
For instance, removing a conversation from an account may not necessarily establish that every associated log or backup has been immediately erased. The applicable documentation should explain relevant exceptions and timelines.
Where research protocols or agreements impose retention limits, the provider's arrangements must be evaluated against those requirements.
If the necessary retention commitments cannot be established, the proposed service may be unsuitable for restricted information.
Who Can Access Research Data Submitted to AI?
Access is not necessarily limited to the researcher who enters the information.
Depending on the system, authorized provider personnel, administrators, service operators, or subprocessors may have defined forms of access for support, security, maintenance, or other permitted functions.
The existence of such access does not automatically mean the service is inappropriate. The question is whether access is limited, governed, documented, and compatible with the confidentiality obligations attached to the information.
Researchers should examine role-based permissions, administrative controls, authentication requirements, access logging, and relevant contractual restrictions.
A personal account used on an institution-owned computer does not automatically become an institutionally governed processing environment.
What Are Subprocessors and Why Do They Matter?
A subprocessor is another organization engaged by a service provider to process information on its behalf under a relevant processing arrangement.
AI providers may depend on other organizations for infrastructure, storage, security, support, or additional services.
Where personal data are involved, applicable data protection frameworks may impose requirements concerning processor arrangements and the engagement of subprocessors.
Under the European Union's General Data Protection Regulation, Article 28 establishes requirements for certain controller-processor relationships, including conditions relating to subprocessors.
Researchers should determine whether the provider's subprocessor arrangements are disclosed and whether the applicable agreement provides the protections required for the project.
The presence of subprocessors is not inherently disqualifying. Undocumented or unauthorized processing arrangements, however, may prevent a researcher from establishing compliance.
How Do Cross-Border Data Transfers Affect the Decision?
Externally hosted AI services may process information in jurisdictions different from the researcher's location or the location where data were collected.
Cross-border processing can create additional obligations under applicable privacy law, contractual agreements, or institutional policies.
For example, the GDPR contains provisions governing transfers of personal data to third countries and international organizations. Philippine researchers must also consider applicable requirements under the Data Privacy Act of 2012 and relevant National Privacy Commission guidance.
Researchers should not assume that all international processing is prohibited. Nor should they assume that a provider's global availability establishes permission to transfer restricted information.
The appropriate assessment depends on the relevant jurisdictions, legal relationships, safeguards, and specific information involved.
What if the Provider Can Change Its Terms?
Online services may update privacy notices, terms of service, product configurations, or data-processing arrangements.
Researchers should determine which terms apply when the data are submitted and whether the agreement provides notice, contractual protections, or other mechanisms for managing material changes.
For a long-term project, it may be useful to retain dated copies of relevant terms and document the approved configuration.
If a later change materially affects the processing arrangement, the research team may need to reassess whether continued use remains authorized.
This is particularly relevant where AI processing forms part of a continuing data collection or analytical workflow rather than a single isolated task.
Can Researchers Rely on a Provider's Security Certifications?
Independent security assessments and certifications can provide useful evidence about specified controls.
However, a certification does not necessarily establish that every service, configuration, data type, or research use is covered by the assessed controls.
Researchers should examine the scope of any assurance and whether it addresses the risks relevant to the proposed task.
For example, evidence about information-security management may support a security assessment without independently establishing that participant consent authorizes the processing.
Technical assurance, contractual authorization, and research ethics approval address different questions. They should be considered together rather than treated as substitutes.
What if the Tool's Terms Are Unclear but the Data Are Public?
Unclear data-use terms do not necessarily prohibit all use of a service.
A researcher may be able to submit publicly available information for a low-risk task, provided the activity is lawful, consistent with applicable restrictions, and scientifically appropriate.
However, public availability does not automatically eliminate copyright, licensing, contractual, or ethical obligations.
For example, a publicly accessible article may still be protected by copyright, and access to a research database may be governed by terms restricting automated processing or redistribution.
Researchers should also avoid entering confidential instructions, unpublished analyses, or restricted contextual information alongside otherwise public material.
The appropriate response is to evaluate the actual information and processing purpose rather than treating all public data as unrestricted.
When Should Researchers Stop Evaluating and Choose Another Tool?
Clarifying a provider's terms can require time, particularly when the service lacks accessible documentation or institutional support.
Researchers should not assume that continued investigation is always worthwhile.
If the provider cannot establish the protections necessary for the proposed information, an approved alternative may be preferable.
This is especially true when the task can be completed using existing institutional software, a validated conventional method, or a locally governed processing environment.
The objective is not to obtain the most elaborate AI service. It is to select a workflow whose scientific and data-handling obligations can be adequately justified.
