Is It Safe to Enter Research Information Into an AI Tool?

When Is Research Information Safe to Share With an AI Tool?

Asking an AI tool to improve the wording of a generic paragraph is one thing. Uploading interview transcripts, participant data, an unpublished manuscript, reviewer comments, a grant proposal, or a collaborator's draft is quite another.

The difficulty is that the same interface may accept all of them. A text box does not tell you whether the information you are about to paste is confidential, whether you have authority to disclose it, how the provider will process it, or whether your institution permits that use.

So the useful question is not simply, "Is this AI tool safe?" It is: Is it appropriate to provide this particular information to this particular AI service under these particular conditions?

There Is No Universal Yes or No

Safety Depends on the Information, the Tool, and the Rules Around Your Research

Entering information into AI is a form of data disclosure and processing

When you type a prompt, paste text, upload a file, attach an image, or otherwise provide material to an externally operated AI service, information leaves the immediate environment in which you were working and is processed by another system. That matters even when the AI produces an excellent answer and even when you delete the conversation afterward.

The European Data Protection Supervisor notes that personal data may be processed at several points in the lifecycle of a generative AI system, including through the inputs submitted to the system and the outputs it produces. Its guidance also emphasizes data minimization: personal data should be adequate, relevant, and limited to what is necessary for the purpose.

For researchers, the first safety question therefore comes before prompting: What exactly am I disclosing by submitting this?

Not all research information carries the same risk

"Research information" is a very broad category. A publicly available abstract, a fabricated dataset created for teaching, an unpublished theory, a participant's medical history, and a confidential peer-review report are all research-related information, but they do not carry equivalent confidentiality, privacy, contractual, ethical, or intellectual-property risks.

This classification is deliberately cautious rather than absolute. Something that looks harmless in isolation may become sensitive when combined with other details. Later in this part, the distinction between removing obvious identifiers and achieving meaningful de-identification is examined in more depth through the limits of removing names before using AI.

The product matters, not merely the name of the AI company

A common mistake is to ask whether a provider is "safe" as though every service, account type, configuration, and contract from that provider handled data identically. They may not.

For example, OpenAI currently states that inputs and outputs from its business offerings, including ChatGPT Business, Enterprise, Edu, and its API platform, are not used to train its models by default. Its consumer ChatGPT products have separate data controls through which users can control whether eligible conversations are used to improve models. Google likewise documents settings governing Gemini Apps activity and model improvement, while warning users not to submit confidential information they would not want human reviewers to see or Google to use to improve its services.

These examples are not endorsements or permanent guarantees. They illustrate why researchers should inspect the terms and controls of the specific service they are actually using. Provider practices, account types, settings, contractual arrangements, retention options, and features can change.

The separate questions of whether providers can store research prompts and whether prompts may be used to train or improve models therefore need to be evaluated independently.

Your right to possess information is not necessarily a right to disclose it

You may legitimately have access to a file and still lack authority to send it to an external AI service. A researcher might possess participant transcripts under an ethics-approved protocol, receive an unpublished manuscript as a reviewer, hold a collaborator's draft for discussion, or access restricted institutional data for an approved project. None of those circumstances automatically authorizes onward disclosure.

This is particularly important for material belonging partly or entirely to someone else. Before entering such information into AI, consider who supplied it, under what conditions you received it, and whether those conditions permit processing by the AI service. The issue becomes especially clear when dealing with a collaborator's unpublished work.

Research obligations can follow the information into the AI workflow

Using AI does not make existing research obligations disappear. Depending on the project and jurisdiction, your use may still be governed by research ethics approval, participant consent, privacy and data-protection requirements, data-management plans, confidentiality agreements, contracts, intellectual-property arrangements, funder conditions, publisher rules, or institutional information-security policies.

If your ethics documentation says that identifiable data will be stored and processed only in specified systems, for example, sending those data to another service may require further assessment. If a confidentiality agreement limits disclosure to named parties, an AI provider does not become an authorized recipient merely because the disclosure occurs through a prompt box.

Removing names does not automatically remove the risk

Deleting names, email addresses, or participant numbers may reduce risk, but it does not necessarily make information anonymous. Combinations of occupation, location, age, rare experiences, quotations, dates, or contextual details can sometimes permit identification. NIST identifies unauthorized disclosure and de-anonymization of personally identifiable or sensitive information among the data-privacy risks associated with generative AI.

The EDPS similarly cautions that claims of anonymity require attention to whether people can reasonably be identified or personal data extracted. This is why apparently anonymous research information can remain sensitive or identifiable.

Think beyond participant data

Researchers often associate privacy risk only with human-subject data. Yet AI-related disclosure can create other problems even when no participant is identifiable. A research file might contain a patentable idea, confidential methods, commercially sensitive findings, embargoed results, privileged correspondence, unpublished arguments, proprietary code, restricted geographic information, reviewer identities, or material entrusted to you by another researcher.

Privacy is therefore only one dimension of research information safety. Confidentiality, research integrity, intellectual property, contractual obligations, cybersecurity, and professional duties may also matter.

Use the minimum information necessary

A useful practical principle is data minimization. If an AI task can be completed without providing the original sensitive material, do not provide it merely because uploading the entire file is convenient.

You might be able to replace real names and details with genuinely non-identifying placeholders, extract only the non-sensitive passage required for the task, describe a methodological problem abstractly, or create synthetic examples that preserve the analytical structure without exposing the underlying records. Whether those measures are sufficient depends on the information and the research context.

Sometimes minimization will not solve the problem. If the task fundamentally requires the AI system to process information that you are not permitted to disclose, the safer answer is not a more ingenious prompt. It is to use an appropriately approved environment or not submit the information.

Before Uploading an Interview Transcript for AI Analysis

The important point is that replacing a participant's name was not enough to answer the safety question. The decision depended on the content, the researcher's obligations, and the characteristics of the specific AI service.

Common Mistakes When Judging Whether AI Is Safe for Research Information

Use a Disclosure Test Before You Paste, Type, or Upload

You do not need to conduct a full security audit before asking an AI tool to explain a public statistical concept. The level of scrutiny should correspond to the sensitivity and obligations attached to the information. What you do need is a habit of checking before disclosure rather than after it.

For genuinely confidential or unpublished material, a conservative default is usually sensible. The later discussion of the safest default for confidential or unpublished research develops that principle more fully.

One final check is surprisingly useful: imagine that the information is leaving your computer and being processed by an external service rather than imagine that you are merely "asking AI a question." That framing makes the disclosure decision much easier to see.

Check These Before Entering Research Information Into AI

Questions Researchers Ask About AI and Research Information

Do Not Treat an AI Prompt Box as a Confidential Workspace by Default

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.