01 · The Question
Should Unpublished Research Go Into a Public AI Tool?
You have a manuscript that has not yet been submitted, preliminary results that nobody outside the project has seen, a promising theoretical argument, or a research design you are still developing. An AI tool could help critique the logic, improve the writing, generate alternatives, or identify weaknesses.
Should you paste the material into a publicly available AI service?
The answer requires more nuance than "never upload unpublished work" or "it is fine because AI is just a tool." Unpublished research varies enormously. So do AI services. The key issue is whether you are authorized to disclose the particular material to the particular service under its current data-handling arrangements.
03 · What You Need to Know
Unpublished Is a Status, Not a Complete Risk Classification
Unpublished research is not one category of information
"Unpublished" simply tells you that the work has not yet been formally published. It does not tell you whether the information is confidential, identifiable, proprietary, restricted, patentable, embargoed, or harmless to disclose.
Consider how different these examples are:
| Unpublished material |
Possible concern |
AI-use implication |
| A generic paragraph explaining a well-known method |
Relatively little confidential information may be present |
May be suitable after ordinary provider and institutional checks |
| A draft containing novel results |
Premature disclosure, intellectual property, collaborator interests |
Requires greater caution |
| An interview transcript |
Participant privacy, consent, ethics requirements, re-identification |
Do not submit merely because names have been removed |
| A patentable technical method |
Intellectual property and disclosure consequences |
Seek appropriate institutional or legal guidance before external disclosure |
| A collaborator's unpublished chapter |
Ownership, confidentiality, permission |
Your access to the file does not establish authority to upload it |
| An embargoed or contractually restricted dataset |
Contractual and access restrictions |
Use only in environments authorized under the governing terms |
The publication status is therefore only the beginning of the assessment.
A public AI tool should be treated as an external service
The familiar conversational interface can make an AI system feel like software running privately on your computer. A publicly available cloud AI service is better understood, for disclosure purposes, as an external system to which you are transmitting information.
The provider must process what you submit to generate an answer. Depending on the product, account, settings, and terms, interactions may also be retained, logged, reviewed, or eligible for model or service improvement.
For example, OpenAI currently states that eligible content submitted through its services for individuals may be used to improve model performance depending on user settings, whereas its Business, Enterprise, Edu, Healthcare, and API offerings exclude customer inputs and outputs from model training by default. Google advises Gemini Apps users not to enter confidential information they would not want human reviewers to see or Google to use to improve its services.
Those examples do not mean that every public AI service handles information identically. They show why the safety of entering research information into AI has to be evaluated at the product level.
Unpublished does not necessarily mean secret
Researchers sometimes overcorrect and treat every sentence of an unpublished manuscript as though it were classified information. That is not a useful rule either.
A literature-review paragraph based entirely on public sources, a generic explanation of a standard procedure, or a hypothetical example may contain little that is confidential. Conversely, one sentence describing an undisclosed result or one quotation from a participant could be highly consequential.
Assess the information itself rather than relying solely on the label "unpublished."
Novelty and confidentiality are different questions
A novel idea may deserve protection even when it contains no personal data. An unpublished manuscript might reveal a new theoretical framework, software architecture, intervention, device, dataset, commercial application, or potentially patentable method.
Whether disclosure to an AI provider has legal consequences for intellectual-property protection depends on the circumstances and jurisdiction. Researchers working with potentially patentable or commercially sensitive material should not improvise based on generic internet advice. Institutional technology-transfer, intellectual-property, research, or legal offices may need to assess the proposed disclosure.
Watch Out
Do not use a public AI service as the place where you first discover whether your unpublished invention, dataset, or commercially sensitive finding was supposed to remain confidential. Check before disclosure.
Co-authorship changes the decision
A manuscript may be saved on your computer and still not be yours alone to disclose. Co-authors and collaborators may have contributed text, data, ideas, code, analyses, or intellectual property. Agreements with external partners may impose additional restrictions.
Uploading the entire document to an AI service can therefore affect interests beyond your own. This is why using a collaborator's unpublished work with AI requires attention to permission rather than mere possession of the file.
Participant information deserves separate scrutiny
Unpublished human-subject research may contain information governed by participant consent, research ethics approval, privacy requirements, institutional policy, and data-management commitments. Those obligations do not disappear when AI is used for analysis, summarization, translation, transcription, coding, or editing.
Removing direct names may reduce risk without producing genuine anonymity. Free-text responses and qualitative transcripts are particularly capable of containing contextual details that identify people indirectly. The later distinction between de-identification and simply removing names before uploading research information is therefore important.
Peer review can create an additional layer of confidentiality
Researchers sometimes use "unpublished research" to include manuscripts they are reviewing for journals. That situation is materially different from using your own draft.
The ICMJE describes submitted manuscripts as privileged communications and warns that using AI systems to process or evaluate manuscripts may violate confidentiality where confidentiality cannot be assured without authors' explicit permission. Reviewers and editors are therefore dealing not merely with unpublished information but with material entrusted to them under an editorial relationship.
The appropriate handling of confidential peer-review material in AI systems should be evaluated under those specific duties rather than under a general rule for your own drafts.
A no-training setting does not settle the decision
Suppose your AI account is configured so that prompts are not used to improve the provider's models. That removes one possible concern. It does not answer whether the information is stored, who may access it under authorized circumstances, whether third-party services receive it, or whether you had authority to disclose it.
Researchers should therefore avoid reducing the decision to whether research prompts are used for model training. That is an important question, but it is not the only one.
You often do not need to upload the whole thing
AI interfaces make full-document upload convenient, but convenience is not a data-minimization principle.
If you want help improving a sentence, submit only the necessary non-sensitive passage. If you need advice about a methodological problem, describe the problem abstractly rather than providing the dataset. If you want to test an analysis workflow, a synthetic example may sometimes reproduce the relevant structure without exposing real research information.
These approaches do not solve every confidentiality problem. They do, however, force a useful question: What is the minimum information the AI actually needs to perform this task?
An approved institutional environment may materially change the risk
A university or research organization may provide an enterprise AI environment governed by institutional agreements, administrative controls, security requirements, retention policies, and no-training commitments that differ from a personal consumer account.
That can make an institutionally approved system more appropriate for certain research tasks. It still does not mean every category of research information is automatically authorized. Project-specific restrictions, participant consent, contractual obligations, ethics requirements, and collaborator agreements can remain controlling.
04 · A Practical Example
Do You Need to Upload the Whole Unpublished Manuscript?
Hypothetical Example
A researcher wants feedback on the clarity of a discussion section
A research team has completed a manuscript but has not yet submitted it. The paper contains novel findings, tables of unpublished results, acknowledgments naming project partners, and a discussion section that one author thinks is difficult to follow. The author considers uploading the complete manuscript to a personal public AI account.
Identify the actual task
The researcher wants help with the clarity of several paragraphs, not analysis of the entire study.
Identify unnecessary disclosure
The results tables, partner information, methods, and most of the manuscript are unnecessary for that writing task.
Check authorization
Before sharing even an excerpt, the researcher checks the team's expectations, institutional AI policy, and the data practices of the service.
Minimize the material
If permitted, the researcher supplies only the passage needed for the task and removes information that is not necessary for meaningful editorial feedback.
Use a different environment if needed
If the material remains confidential or project rules prohibit the public service, the researcher uses an approved institutional environment or performs the editing without that AI service.
The useful decision was not simply "AI or no AI." It was whether the proposed disclosure was necessary, authorized, and proportionate to the task.
07 · A Quick Checklist
Before Putting Unpublished Research Into a Public AI Tool
Before submitting unpublished material, check:
Identify exactly what unpublished information the AI needs for the task and remove material it does not need.
Check whether the material contains participant information, confidential content, restricted data, proprietary information, or potentially valuable intellectual property.
Confirm that collaborators, data providers, funders, partners, or other rights holders have not imposed restrictions on disclosure.
Verify relevant ethics approval, participant consent, data-management, contractual, institutional, and publication requirements.
Check the exact AI service's current storage, retention, model-improvement, human-review, and deletion arrangements.
Consider whether a limited excerpt, abstract description, or synthetic example can accomplish the task instead of uploading the complete research material.
Prefer an institutionally approved AI environment when the information requires stronger contractual, security, or governance safeguards.
If authorization is unclear, keep the unpublished material out of the public AI tool until the uncertainty is resolved.
