01 · The Question
Can You Give an Unpublished Paper to AI Before Submission?
You have a complete manuscript on your computer. An AI tool could proofread it, shorten the abstract, critique the discussion, check the logical flow, or suggest clearer wording. The simplest approach seems obvious: upload the file and ask.
But the manuscript has not been published. Perhaps it contains new findings, participant quotations, confidential project information, or contributions from several co-authors. Perhaps you intend to submit it to a journal with its own AI policy.
Uploading an unpublished manuscript is therefore not merely a writing decision. It can also be a confidentiality, privacy, authorship, intellectual-property, research-integrity, and publication-policy decision.
03 · What You Need to Know
Your Own Manuscript Can Still Contain Information You Cannot Freely Disclose
An unpublished manuscript is not automatically confidential
A manuscript that has not yet appeared in a journal is unpublished, but that fact alone does not establish its confidentiality status. A draft literature review based entirely on public information may present relatively little confidentiality concern. A manuscript describing undisclosed clinical results, proprietary technology, participant experiences, or an industry collaboration can present considerably more.
This distinction follows the broader principle for using unpublished research with public AI tools: classify the actual information rather than treating "unpublished" as a complete risk category.
The complete manuscript contains much more than the sentence you want edited
Uploading a document can expose the title, abstract, methods, results, tables, figures, supplementary text, acknowledgments, funding information, affiliations, author identities, participant quotations, references, and embedded metadata even when the AI task concerns only one paragraph.
That makes full-document upload a poor default when a smaller excerpt would accomplish the same task.
| What you want AI to do |
What may actually be necessary |
What may be unnecessary to disclose |
| Improve one paragraph's clarity |
The paragraph and minimal surrounding context |
Results tables, author details, complete methods, other sections |
| Shorten the abstract |
The abstract |
The entire manuscript file |
| Suggest alternative section headings |
Current headings and a non-sensitive description of the paper |
Raw data and confidential findings |
| Critique the argument |
Relevant argument and enough context to understand it |
Information unrelated to the reasoning being assessed |
| Analyze participant quotations |
Potentially sensitive research data |
This may require an approved research-data environment rather than a public writing tool |
Co-authors have interests in the manuscript too
A multi-author manuscript is not simply one researcher's private text file. It contains intellectual contributions from other people who may reasonably have expectations about where unpublished work is disclosed and how AI is used in preparing the paper.
Some projects also have formal collaboration agreements, sponsor restrictions, intellectual-property arrangements, or institutional policies. Before uploading substantial portions of a jointly authored manuscript, establish whether the proposed use is consistent with those arrangements.
The underlying principle is the same as for a collaborator's unpublished work: access to material is not necessarily equivalent to permission to disclose it to an external service.
Participant information can remain embedded in the manuscript
A paper may no longer contain a spreadsheet of participant names yet still reveal personal information. Case descriptions, verbatim quotations, photographs, geographic details, rare occupations, demographic combinations, dates, or unusual events can sometimes identify participants.
Researchers should therefore not assume that manuscript preparation falls outside research-data governance merely because the analysis is finished. Participant consent, ethics approval, privacy requirements, and institutional policies may continue to govern information appearing in the draft.
Removing obvious names does not automatically solve this problem. Apparently anonymous research information may still be identifiable or sensitive.
Potentially patentable or commercially sensitive results deserve special care
Some manuscripts describe inventions, software, processes, discoveries, or other intellectual property before protection strategies have been finalized. Researchers should not assume that submitting such material to an external AI service is inconsequential.
The legal significance of a disclosure can depend on jurisdiction, contractual arrangements, the nature of the service, and the intellectual property involved. If a manuscript contains potentially patentable or commercially sensitive material, consult the appropriate technology-transfer, intellectual-property, research, or legal office before external disclosure where necessary.
Publisher policies increasingly address author use of AI
The fact that you are the author does not mean the intended journal has no interest in how AI was used during manuscript preparation.
ICMJE recommends transparency when AI tools are used during manuscript preparation and states that humans remain responsible for the accuracy and integrity of AI-assisted content. It also notes that AI tools cannot qualify as authors.
Elsevier's current journal policy permits authors to use generative AI and AI-assisted technologies during manuscript preparation under specified conditions, while requiring authors to review the tool's terms so that privacy and confidentiality of inputs, including unpublished manuscripts and personally identifiable information, are maintained. Elsevier also requires disclosure of qualifying generative AI use in manuscript preparation, while basic spelling, grammar, and punctuation tools are excluded from that disclosure requirement.
Policies differ among publishers and journals and can change. Check the intended journal's current instructions rather than assume that one publisher's policy represents scholarly publishing as a whole.
AI assistance does not transfer responsibility for the manuscript
An AI system can generate fluent prose that contains factual errors, distorted interpretations, invented citations, inappropriate generalizations, or subtle changes in scientific meaning. Authors remain responsible for what they submit.
ICMJE emphasizes human responsibility for checking AI-generated content. Elsevier similarly requires authors to review and edit AI-assisted content and remain accountable for the final work.
This matters especially when AI rewrites technical sections. A stylistically smoother sentence is not necessarily a scientifically equivalent sentence. The model has no standing at your viva when someone asks why the confidence interval suddenly changed meaning.
AI use during research may need different reporting from AI used for writing
There is an important difference between using AI to improve prose and using AI as part of the research method.
Elsevier, for example, distinguishes generative AI used in manuscript preparation from AI used as part of formal research design or methods. Methodological use should be described in the Methods section when relevant, while qualifying use during writing requires a separate declaration under its journal policy.
Researchers should therefore determine what role AI actually played rather than treating every use as "editing."
Manuscripts received for peer review are a different category
Your own unpublished draft and another researcher's manuscript sent to you for confidential peer review should not be treated as equivalent.
ICMJE describes submitted manuscripts as authors' private, confidential property and states that reviewers should not upload them to AI systems where confidentiality cannot be assured unless such use is explicitly permitted. Elsevier's current reviewer policy is stricter: reviewers must not upload submitted manuscripts, or any part of them, into AI tools because of confidentiality, intellectual-property, and possible privacy concerns.
That separate obligation is addressed more directly in the use of confidential peer-review material with AI.
The AI service's own terms still matter
Even when your institution and journal permit AI-assisted writing, you still need to know what happens to the manuscript after upload.
Depending on the service and configuration, prompts and files may be stored, retained, reviewed, or eligible for model improvement. Business and institutional services may provide materially different contractual and data-handling arrangements from consumer accounts.
Checking what happens to research information after it enters an AI system is therefore part of manuscript risk assessment, not a separate technical curiosity.
Watch Out
A publisher allowing authors to use generative AI for manuscript preparation does not mean that every AI tool is approved for every manuscript. The author may still be responsible for choosing a service whose terms preserve confidentiality, privacy, intellectual-property rights, and the ability to publish the work.
04 · A Practical Example
Proofreading Does Not Necessarily Require Full-Manuscript Disclosure
Hypothetical Example
A researcher wants AI help with the discussion section
A four-author manuscript contains unpublished findings from a funded study, several participant quotations, and a potentially commercially useful analytical method. One author wants an AI assistant to improve the readability of three paragraphs in the discussion and considers uploading the complete Word document.
Define the actual task
The researcher needs language assistance with three paragraphs, not processing of the entire manuscript.
Check the manuscript
The complete document contains information irrelevant to the editing task, including participant material and a potentially sensitive method.
Check permission and policy
The researcher reviews co-author expectations, institutional policy, the intended journal's AI instructions, and the AI service's current terms.
Minimize disclosure
If permitted, only the necessary passages and minimal contextual information are provided rather than the complete file.
Verify the result
The researcher checks every change for scientific accuracy, restores any nuance altered by the model, and follows applicable disclosure requirements.
The AI task remains useful, but the workflow no longer treats an unpublished manuscript as an all-or-nothing upload.
07 · A Quick Checklist
Before Uploading an Unpublished Manuscript to AI
Before submitting manuscript content, check:
Identify exactly what you want the AI to do and whether it genuinely requires the complete manuscript.
Check the manuscript for participant information, confidential findings, proprietary material, potentially patentable content, or other restricted information.
Confirm that co-authors, collaborators, institutions, partners, and other rights holders have not imposed restrictions inconsistent with the proposed AI use.
Review the intended journal or publisher's current policy on generative AI and manuscript preparation.
Check the AI service's current privacy, retention, model-improvement, file-handling, intellectual-property, and deletion terms.
Use the smallest non-sensitive excerpt that can accomplish the task when full-document access is unnecessary.
Review every AI-generated revision for factual accuracy, scientific meaning, appropriate attribution, and fabricated references.
Document and disclose qualifying AI use when required by the journal, publisher, institution, funder, or applicable reporting policy.
