How Do Researchers Decide Which Information Must Never Be Entered Into an AI System?

What Research Information Should Never Be Pasted Into an AI Chatbot?

A researcher wants to summarize interview transcripts using generative AI. Before uploading the files, the researcher removes participants' names and replaces their student identification numbers with anonymous-looking codes.

The transcripts still describe unusual academic experiences, particular university departments, family circumstances, and incidents involving identifiable individuals.

The researcher assumes that removing direct identifiers makes the information safe to submit to a commercial AI service.

But is the information genuinely anonymous? Does the researcher's authority to access a dataset also include permission to transfer it to an external provider? Would the answer change if the institution offered an approved AI environment with contractual data protection safeguards?

Researchers need to distinguish information that must not be entered into a particular AI system from information that may be processed under an appropriately authorized and protected arrangement.

Do Not Enter Research Information Unless Its Processing Is Authorized and Adequately Protected

How to Determine Which Research Information Is Unsafe or Unauthorized for AI Processing

Why the Question Is About Authorization, Not Just Sensitivity

Researchers often classify information as public, internal, confidential, or highly restricted. These categories help identify protection requirements, but sensitivity alone does not determine whether AI processing is permitted.

A dataset may contain information that is not especially sensitive yet remains subject to a contractual restriction against sharing with third parties.

Conversely, sensitive information may be processed within an appropriately authorized research environment when the necessary legal, ethical, organizational, and technical safeguards are established.

The central question is whether the researcher is permitted to submit the particular information to the particular system for the intended purpose.

Permission to access information for research does not automatically include permission to disclose it to another organization or processing service.

Passwords, Access Tokens, and Authentication Secrets

Researchers should not paste live passwords, private cryptographic keys, authentication tokens, recovery codes, or similar secrets into general-purpose AI prompts.

These credentials may provide access to institutional systems, research databases, cloud storage, or participant records.

Disclosure can create risks extending beyond the immediate conversation, particularly when a credential remains valid or grants broad permissions.

If AI assistance is needed to troubleshoot authentication code, researchers can generally use synthetic examples or redact actual credentials.

Where a credential has been exposed to an unauthorized system, the researcher should follow institutional security procedures, which may include revocation, rotation, and incident assessment.

Secrets should be managed through authorized credential-management mechanisms rather than treated as ordinary research text.

Identifiable Participant Records

Participant information may contain names, contact details, identification numbers, photographs, audio recordings, and other direct identifiers.

Submitting these records to an external AI service may introduce processing that participants and the institution did not authorize.

For example, an interview recording may reveal a participant's identity through their voice even when the accompanying transcript uses a pseudonym.

Researchers should establish the applicable consent conditions, approved protocol, data protection requirements, and permitted processing arrangements before using AI.

When authorization is absent or unclear, identifiable participant records should not be entered into the proposed system.

Highly Sensitive Participant Information

Some research records concern health, financial circumstances, intimate experiences, disciplinary matters, or other information whose exposure could cause significant harm.

The consequences may include distress, stigma, discrimination, or loss of confidentiality.

Researchers should assess both the sensitivity of individual fields and the risks created by combining them.

For example, a dataset containing demographic characteristics and a sensitive interview narrative may permit identification even after names are removed.

Processing such information may require stronger access controls, data minimization, institutional authorization, and additional legal or ethical safeguards.

The fact that AI can summarize the material accurately does not establish that it should receive it.

Why Pseudonymized Data Are Not Necessarily Anonymous

Pseudonymization replaces direct identifiers with codes or other references while preserving the possibility of linking records to individuals through additional information.

Anonymization aims to prevent identification under the applicable standard and relevant circumstances.

These concepts are not interchangeable.

For example, replacing a student's name with Participant 07 does not remove references to a unique academic position, unusual incident, or identifiable location.

Researchers should evaluate reidentification risks in relation to the information available to the intended recipient and other reasonably accessible data.

They should not assume that deleting names automatically makes a dataset unrestricted.

Unpublished Manuscripts and Confidential Peer Review Materials

Unpublished research may be subject to confidentiality obligations even when it contains no personal information.

Manuscripts received for peer review, confidential grant applications, embargoed findings, and proprietary research proposals may have restrictions on disclosure to external services.

A reviewer who receives a manuscript under confidentiality obligations should not assume that uploading it to a generative AI system is permissible.

Researchers should examine the relevant journal, funder, institutional, or contractual policy.

Where AI processing is not authorized, the material should not be submitted.

Using synthetic excerpts or general methodological questions may provide assistance without exposing the protected document.

Restricted Institutional and Educational Records

Educational research frequently involves grades, attendance records, learning management system activity, assessment results, and institutional administrative information.

These records may be governed by institutional policies, consent arrangements, and applicable data protection law.

For example, an instructor may be authorized to access student grades for educational administration but not to upload those grades to an external AI platform for unrelated research analysis.

Researchers should establish the purpose for which access was granted and whether the proposed AI processing is covered.

Authorization should be assessed for the actual research task rather than inferred from the researcher's institutional position.

Proprietary Datasets, Source Code, and Contractually Restricted Information

Research collaborations may involve datasets, software, technical specifications, or business information governed by nondisclosure agreements or licensing terms.

Such information may not be suitable for submission to external AI services without the relevant permission.

Even when the material contains no personal information, uploading it may constitute a disclosure prohibited by an agreement.

Researchers should examine data-use agreements, intellectual property obligations, and relevant institutional policies.

When restrictions are uncertain, the appropriate data owner or authorized institutional office should clarify the permitted processing.

Which Information Categories Require Particular Caution?

The table is a practical screening aid, not a universal legal classification. Restrictions depend on the actual information and governing arrangements.

Does Public Availability Mean Information Can Always Be Entered Into AI?

Public availability may reduce some confidentiality concerns, but it does not necessarily remove licensing, copyright, ethical, or contractual restrictions.

For example, publicly accessible online posts may still contain personal information and may have been published in a context different from the proposed research use.

Researchers should consider whether the collection and processing are appropriate for the research question and permitted under applicable requirements.

Publicly accessible information should not automatically be treated as ethically unrestricted.

What Should Researchers Examine in AI Provider Terms?

AI services differ in their retention, access, training, and data-processing arrangements.

Researchers should examine the terms applying to the specific service, account type, and configuration.

Relevant questions include whether prompts and uploaded files are retained, whether they may be used for model improvement, who can access them, and what deletion mechanisms are available.

External integrations and subprocessors may introduce additional processing relationships.

A statement that submitted information is not used for model training does not necessarily mean that it is never stored or accessed for other authorized purposes.

Researchers should avoid relying on broad marketing claims when determining whether a service meets research obligations.

Why a Paid or Enterprise Account Is Not Automatic Authorization

Enterprise AI services may offer contractual commitments, administrative controls, and technical safeguards that differ from consumer services.

These features can support appropriate research processing.

However, they do not independently establish that a particular dataset may be submitted.

The institution may approve a service for ordinary administrative information while prohibiting its use with identifiable participant records.

Researchers should confirm both the system's approved configuration and the information categories permitted for the intended task.

Tool approval and data-processing authorization are related but separate decisions.

Can Researchers Use Synthetic or Aggregated Information Instead?

Replacing real records with synthetic examples can sometimes allow researchers to obtain programming or methodological assistance without exposing protected information.

For example, a researcher may ask AI to generate a script using fictional variable names and a small synthetic dataset, then execute the validated script within an authorized research environment.

Aggregated statistics may also be suitable for certain tasks, provided the aggregation does not reveal protected information.

However, synthetic examples may fail to represent unusual structures in the original data, and small aggregate groups can sometimes create disclosure risks.

Researchers should verify both the privacy suitability and technical adequacy of the substitute information.

What Is Data Minimization?

Data minimization involves limiting personal information processing to what is necessary for a legitimate and authorized purpose.

For AI-assisted research, the principle encourages researchers to examine whether the system needs the complete dataset or only a restricted subset of information.

For example, AI-generated statistical code may require knowledge of variable names and data types but not actual participant identities.

Reducing unnecessary information can lower exposure risks.

Nevertheless, minimization does not make otherwise unauthorized processing permissible.

The remaining information and processing arrangement must still satisfy applicable requirements.

What if an AI System Has Connected Access to Research Files?

Some AI applications can retrieve information from connected document repositories, databases, or institutional services.

The information available to the system may therefore extend beyond what the researcher explicitly types into a prompt.

Researchers should examine the scope of connected permissions and whether the application can access unrelated confidential material.

Least-privilege access, appropriate authentication, and institutional review may be necessary.

Connecting a repository should not be treated as harmless merely because the researcher does not manually upload individual files.

When Does AI Use Require Privacy or Security Review?

Privacy and information-security review may be necessary when AI introduces a new provider, external transfer, storage arrangement, access permission, or processing purpose involving protected information.

The applicable requirements depend on institutional policies, contractual commitments, and law.

Researchers should establish whether the proposed AI processing requires privacy or information-security review before submitting restricted material.

Where participant procedures or approved research activities change, ethics review may also be relevant.

What if Protected Information Has Already Been Entered?

If a researcher discovers that restricted information was submitted to an unauthorized AI service, the appropriate response is not simply to delete the conversation and assume the problem has disappeared.

The researcher should follow applicable institutional incident procedures promptly.

Depending on the circumstances, this may involve securing relevant records, notifying authorized privacy or security personnel, assessing the information exposed, and taking required corrective action.

Notification obligations depend on the jurisdiction, information, and nature of the incident.

Researchers should not conceal a potentially consequential exposure because the AI output was useful or because no immediate harm is visible.

When Should Researchers Decide Not to Use AI?

Researchers should avoid a proposed AI system when they cannot establish permission to process the necessary information or cannot provide adequate protection for it.

They may consider an authorized institutional environment, a locally controlled implementation, synthetic examples, or an established non-AI procedure.

Local deployment may reduce some external transfer risks, but it still requires appropriate security controls.

The decision should be based on whether the complete processing arrangement is defensible, not merely whether the AI model is technically capable.

Deciding Whether Student Interview Transcripts Can Be Uploaded to AI

The example illustrates why removing direct identifiers is not enough to establish that external AI processing is permitted.

It also shows that declining to upload information to a particular service does not necessarily prohibit all AI assistance. The task may be redesigned around an appropriately authorized environment or a less sensitive input.

Misconceptions About Information That Should Not Be Entered Into AI

Decide What May Be Shared Before Opening the AI System

Begin by identifying the information the AI task actually requires. Determine whether that information may be processed by the selected system under the applicable research, institutional, and contractual arrangements.

Where provider terms are uncertain, consider whether unclear data-use conditions make the proposed AI service unsuitable.

The decision should form part of a documented AI-assisted research workflow and reflect the risks associated with the particular task.

Where confidentiality cannot be adequately protected, researchers should evaluate whether the proposed AI use should be avoided altogether.

The practical standard is not whether the information can be pasted into a prompt. It is whether the researcher is entitled to process it through that system.

Before Entering Research Information Into AI

Frequently Asked Questions About Sharing Research Information With AI

Never Submit Research Information Without Establishing Permission and Protection

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.