Can Generative AI Invent Manuscript Changes While Drafting a Response to Reviewers?

Can AI Describe Manuscript Revisions That Do Not Exist?

You ask generative AI to draft a point-by-point response to reviewer comments. You provide the reviewers' concerns and explain that the manuscript has undergone revision.

The AI produces a polished letter describing the changes supposedly made. It states that the Methods section now contains a more detailed sampling explanation, that Table 3 includes additional statistical estimates, and that the Discussion incorporates several newly cited studies.

The responses sound plausible. They even include page and line references.

However, when you examine the revised manuscript, you discover that some of the described changes do not exist. The table has not been updated, the citations were never added, and the page references point to unrelated passages.

The AI has generated a coherent narrative of manuscript revision without an accurate basis in the actual document.

How can researchers prevent these invented changes from entering the formal response to reviewers?

Yes, AI Can Fabricate Specific Manuscript Changes and Their Locations

How AI Can Hallucinate Manuscript Revisions and Why Those Errors Matter

What Does an Invented Manuscript Change Mean?

An invented manuscript change is a statement describing a revision that is not supported by the actual revised document or research record.

For example, AI might claim that the authors added a paragraph explaining missing-data handling when no such paragraph appears in the manuscript.

It might describe a new regression model that was never estimated or cite a table that contains different results.

The problem concerns the factual relationship between the response letter and the revised manuscript.

Even when the proposed revision would have been reasonable, describing it as completed without evidence is misleading.

Why Generative AI May Invent Detailed Revisions

Generative AI systems can produce plausible text based on the prompt and learned writing patterns. They do not necessarily verify generated statements against an authoritative manuscript version.

Response-to-reviewers letters contain recurring formulations such as "We have clarified," "We have added," and "The revised manuscript now includes."

When the model lacks a verified record of completed revisions, it may generate these conventional expressions and supply plausible details.

The risk can increase when authors ask for a complete response letter without specifying which changes were actually made.

Detailed output may appear especially credible because it resembles the documentation expected in journal resubmissions.

Which Manuscript Changes Can AI Invent?

Invented revisions may concern both scientific content and document structure.

These examples are hypothetical. Their purpose is to show how specific and convincing an unsupported revision claim can become.

Why Invented Revisions Are Different From Uncompleted Promises

An uncompleted promise describes an action that the authors say they will perform but have not yet completed.

An invented manuscript change describes an action as though it has already occurred, often with details about its content or location.

For example, "We will add a sensitivity analysis" is a future commitment. "We added a sensitivity analysis in Table 4, showing that the results remain unchanged" is a factual claim about completed work.

The second statement is more specific and may be especially misleading if neither the analysis nor the table exists.

The related issue of AI promising manuscript revisions that were never completed concerns commitments and revision status. Invented changes concern unsupported descriptions of the revised research itself.

AI Can Invent Results, Not Just Revision Locations

One of the most consequential risks is the generation of nonexistent analytical findings.

For example, a model might state that an additional regression confirmed the robustness of the original association.

If the analysis was never performed, the statement invents both a research activity and its outcome.

Even if an analysis was performed, AI may misreport its numerical results or exaggerate their implications.

Authors should verify all statistical claims using the actual analysis outputs and appropriate computational records.

A response letter should never contain fabricated findings merely because they provide a convenient answer to a reviewer.

Why Invented Page and Line Numbers Are Particularly Misleading

Many journals request precise references to revised passages so that reviewers can locate changes efficiently.

Generative AI may produce page and line numbers that look plausible but are not grounded in the document.

Such references can create a false appearance of traceability.

For example, a response may identify page 11, lines 245–260, even though those lines concern an unrelated section.

Authors should not allow AI to infer manuscript locations from typical article structures.

Page and line references should be inserted or verified after the revised document has reached a stable version.

AI May Describe a Change That Exists but Misrepresent Its Substance

Not every problematic statement refers to a completely nonexistent revision.

Sometimes the manuscript contains a genuine change, but the AI describes it inaccurately.

For example, the authors may have added a brief limitation acknowledging possible confounding. The AI might describe this as a comprehensive adjustment for confounding variables.

The first is a textual qualification. The second implies a substantive analytical procedure.

Similarly, adding a citation to the discussion does not mean the authors conducted a new literature review or systematically evaluated all relevant evidence.

Authors should verify both the existence and the actual meaning of every claimed revision.

Can AI Invent Citations That Were Supposedly Added?

Yes. A model may claim that the authors incorporated references that do not appear in the manuscript.

It may also suggest nonexistent publications or misrepresent real studies as supporting the revised argument.

Authors should check the revised reference list, in-text citations, and original sources.

Where a response states that a specific study was added, that citation should actually appear in the manuscript and support the relevant claim.

Invented references create a separate research integrity concern beyond inaccurate revision reporting.

Does Providing the Revised Manuscript Eliminate Hallucinations?

No. Access to the revised manuscript may improve grounding, but it does not guarantee accurate comparison or interpretation.

A model may overlook information, confuse document versions, misread tables, or infer changes that were never made.

Furthermore, a revised manuscript alone may not reveal which passages changed from the original submission.

Reliable change identification may require the original version, revised version, tracked changes, and verified revision notes.

Even when these materials are available through an authorized environment, authors should independently check the generated response.

Why Version Confusion Can Produce False Revision Claims

Manuscript revision often involves multiple files, including drafts, coauthor versions, tracked-change documents, and clean copies.

AI may describe changes found in an intermediate draft that were subsequently removed from the final manuscript.

It may also compare the wrong versions or assume that an earlier proposed change remains present.

Authors should identify which manuscript version is authoritative for the resubmission.

Response claims should be checked against that exact version rather than an outdated working draft.

How Can Authors Establish a Reliable Revision Record?

A revision record should connect each reviewer comment with the authors' decision, the completed action, and the evidence supporting that action.

For textual changes, the evidence may be an identifiable passage in the revised manuscript.

For additional analyses, it may include code, model specifications, outputs, and the corresponding reported results.

For a declined recommendation, the record should contain the scientific justification rather than a fabricated manuscript change.

AI may help convert these verified records into clear prose, but it should not supply missing evidence.

Can AI Be Asked to Compare Original and Revised Manuscripts?

Potentially, when the journal permits the intended processing and the tool provides appropriate confidentiality safeguards.

AI may identify candidate differences between document versions or organize tracked changes into a preliminary summary.

However, it may overlook subtle revisions or misinterpret whether a change addresses a particular reviewer concern.

Authors should verify each identified change directly in the relevant documents.

Conventional document-comparison tools may be preferable for establishing exact textual differences, while scientific interpretation still requires author judgment.

How Should Authors Respond When a Claimed Revision Cannot Be Found?

The claim should not remain in the response letter without resolution.

Authors should determine whether the revision was omitted, removed during editing, described inaccurately, or never performed.

If the change is scientifically appropriate and still required, it may need to be completed before resubmission.

If the request was not adopted, the response should explain the decision accurately.

Authors should not invent a page reference or adjust the wording to imply that an absent change exists.

What If an Invented Change Has Already Been Reported to the Journal?

Authors should correct the inaccurate statement promptly through the appropriate editorial channel.

The correction may involve a revised response letter, clarification of the actual manuscript changes, or submission of a corrected manuscript.

The appropriate procedure depends on the journal's instructions and the stage of editorial assessment.

Authors should avoid attempting to conceal the discrepancy or fabricate research outputs to match the earlier claim.

An accidental drafting error and deliberate misrepresentation are not equivalent, but both require accurate handling of the scholarly record.

Confidentiality and AI Disclosure Remain Relevant

Original and revised manuscripts, reviewer comments, and analytical records may contain unpublished or restricted information.

Authors should not upload these materials to external AI services without checking the journal's policies and the relevant data-handling safeguards.

Where AI assistance is permitted, the authors should use authorized processing arrangements and follow applicable disclosure requirements.

Verification of generated content does not retrospectively authorize an inappropriate disclosure of confidential material.

The authors remain responsible for the accuracy of the final response letter and the integrity of the revised manuscript.

When AI Invents a New Results Table and Statistical Findings

The corrected response is appropriate only if the authors actually made the described changes to the Discussion and interpretation.

The example illustrates why specific AI-generated details should not be accepted without documentary evidence. A fabricated analysis accompanied by a table number and page reference remains fabricated.

Misconceptions About AI-Invented Manuscript Changes

How to Verify AI-Generated Manuscript Changes Before Resubmission

Authors should treat every statement about a completed revision as a factual claim requiring evidence.

The final manuscript and supporting research records should determine what the response letter says, not the other way around.

Use a Claim-by-Claim Verification Process

A practical approach is to examine every sentence that asserts a completed action.

Statements beginning with "We added," "We revised," "We conducted," or "The results now show" should be connected to identifiable manuscript evidence.

For analytical changes, authors should verify not only that the procedure occurred but also that the reported results and interpretations are accurate.

Where evidence is missing, the statement should be removed or revised to reflect the actual situation.

A response letter should function as a verifiable account of revision, not a plausible reconstruction of what the authors might have done.

Checklist for Detecting Invented Manuscript Changes in AI Responses

Frequently Asked Questions About AI-Invented Manuscript Revisions

Specific Revision Details Are Not Proof That the Changes Exist

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.