03 · What You Need to Know
Which Records Can Establish How AI Was Used During Research and Manuscript Preparation?
Why Is an AI-Detector Score Weaker Than Evidence of the Writing Process?
Most general-purpose AI-text detectors examine characteristics of the final manuscript rather than observing how it was created.
Depending on the system, they may analyze linguistic predictability, statistical features, or patterns learned from human-written and AI-generated examples.
Their classifications can be incorrect. Human-written material may be flagged as AI-generated, while genuinely generated text may escape detection.
By contrast, documentary evidence may record actual events during manuscript preparation: when a passage appeared, how it was revised, whether an AI tool produced relevant output, and what the researcher subsequently did with that output.
This does not make documentary evidence infallible. Records can be incomplete, misleading, altered, or taken out of context.
Nevertheless, records documenting the writing process generally have a more direct relationship to the allegation than a detector's inference from the finished prose.
The distinction underlies why AI-detection scores cannot independently establish authorship.
What Makes Evidence Strong in an AI-Use Investigation?
Evidence is stronger when it can support a specific factual proposition with relatively little unsupported inference.
Four considerations are particularly useful.
-
Authenticity: Can the record reasonably be established as genuine, and is its origin sufficiently documented?
-
Relevance: Does the record address the specific conduct being investigated?
-
Temporal connection: Was the record created during the relevant research or writing activity, rather than reconstructed afterward?
-
Corroboration: Is the account supported by other independent or mutually consistent information?
Completeness and reliability also matter. A genuine record may still provide only a partial account of what happened.
For example, a verified AI conversation establishes that an interaction occurred, but it may not establish that the resulting text appeared in the submitted manuscript.
Likewise, a manuscript revision history may show extensive human editing without proving that no AI-generated material was used at an earlier stage.
Investigators should therefore distinguish what a record directly demonstrates from what they infer from it.
How Useful Are Manuscript Drafts as Evidence of Human Authorship?
Earlier drafts can provide a chronology of manuscript development.
They may show how an argument emerged, how evidence was incorporated, and how sections changed through revision.
For example, a sequence of drafts might document the progression from an outline to an initial literature synthesis and then to a revised discussion incorporating supervisor feedback.
Such records can support a researcher's explanation of how the manuscript was written.
However, a draft's existence does not automatically establish human-only authorship. AI-generated material could have been incorporated before the earliest retained version, or copied into a draft without preserving its source.
The evidentiary value depends on the chronology, the content of the revisions, and whether the records are consistent with other information.
Investigators should avoid demanding a complete draft history unless the applicable requirements justify that expectation. Not every researcher uses software that automatically preserves every version.
Can Google Docs or Microsoft Word Version History Help?
Document revision histories may provide useful evidence about when text was added, deleted, or modified.
Google Docs version history can show earlier states of a document and, depending on the collaboration settings and available records, contributions associated with particular accounts.
Microsoft Word may preserve tracked changes, comments, and version information, depending on how the file was stored and edited.
These records can help reconstruct a writing chronology.
However, they have limitations. A large block of pasted text does not reveal whether its source was a human-authored draft, an AI response, or another document. An account associated with an edit does not necessarily establish who composed the inserted words.
Version histories are therefore useful corroborating records, not automatic authorship certificates.
Investigators should examine the underlying document history where authorized rather than relying solely on screenshots that may omit important context.
Are Saved AI Conversations Strong Evidence of AI Use?
Saved AI interactions may provide comparatively direct evidence that a particular AI tool was used for a particular purpose.
A conversation might contain a request to draft a literature review paragraph, generate research code, summarize findings, or revise manuscript language.
If the interaction can be authenticated and linked to the relevant researcher and period, it may establish that AI assistance was requested and that particular output was produced.
However, several distinctions remain important.
An AI conversation does not necessarily prove that the output was accepted, copied, or incorporated into the submitted work.
Researchers frequently reject suggestions, correct errors, or use AI responses only as preliminary material.
Conversely, the absence of saved conversations does not establish that AI was never used. Interactions may not have been retained, may have occurred through another account or service, or may have been deleted according to provider settings.
Evidence that AI was consulted
A credible record showing that a researcher interacted with an AI system for a particular task.
Evidence that AI output was used in the manuscript
Records connecting the generated material to the submitted work, such as authenticated output and corresponding manuscript revisions.
The second proposition generally requires more than the first.
The possible role of AI conversations as part of a research audit trail depends on how they are preserved, authenticated, and interpreted.
Can Comparing AI Output With Manuscript Text Establish Undisclosed Use?
A close correspondence between authenticated AI output and a manuscript passage may provide relevant evidence that the generated material contributed to the manuscript.
Its strength depends on the circumstances.
A long, distinctive passage appearing in an authenticated AI interaction before its insertion into a manuscript may be more informative than a match involving a short conventional academic expression.
Investigators should examine the chronology and whether the passage could reasonably have originated elsewhere.
Similarity alone does not establish who copied the text, whether the AI response was generated from an earlier human-authored draft, or whether the relevant assistance was prohibited.
For example, a researcher may have pasted an existing paragraph into an AI tool to request grammar feedback. A later AI conversation containing that paragraph would not establish that the AI originally wrote it.
Direction of information flow matters.
Where possible, investigators should compare dated drafts, prompts, outputs, and subsequent manuscript revisions rather than assuming that textual correspondence proves generation.
What Can Research Notes and Literature Records Demonstrate?
Research notes can document the intellectual development of a manuscript.
Annotated articles, literature matrices, coding notes, analytical memos, and conceptual outlines may show how the researcher interpreted evidence and developed an argument.
These materials are especially useful when an allegation concerns whether a manuscript section was meaningfully developed by the researcher.
For example, a literature synthesis matrix may show the researcher's comparison of findings across studies before the corresponding discussion was drafted.
Such records support the existence of independent scholarly work, but they do not necessarily establish that every sentence was written without AI assistance.
Investigators should also recognize disciplinary differences. Experimental laboratory records, qualitative analytical memos, and theoretical research notes may look very different while serving comparable documentary purposes.
Can Correspondence With Supervisors and Coauthors Help?
Relevant correspondence can help establish how a manuscript developed and who contributed to particular revisions.
Emails discussing an argument, supervisor comments on an early draft, or coauthor exchanges about analysis may provide contemporaneous evidence of scholarly decision-making.
For example, a supervisor's comments requesting a revised theoretical framework, followed by a dated manuscript revision addressing those comments, may support the researcher's account of the writing process.
However, correspondence should not be treated as a substitute for direct evidence when the allegation concerns a specific AI interaction.
Its value depends on what it actually documents.
Investigators should also respect the confidentiality of correspondence and avoid collecting unrelated personal communications merely because they are technically accessible.
What About AI-Use Logs Maintained by Researchers?
An AI-use log is a record of how a researcher used AI tools during a project.
Depending on the research activity, it may document the tool, date, purpose, inputs, outputs, verification steps, and decisions about whether suggestions were incorporated.
A contemporaneously maintained log may help clarify the extent of assistance and whether it falls within the relevant disclosure policy.
For example, a log might show that an AI tool was used to identify grammatical inconsistencies but not to generate substantive manuscript arguments.
Such a record is useful, although its accuracy and completeness still require consideration.
An investigator should not assume that every researcher was required to maintain an AI-use log unless an applicable policy imposed that requirement.
Researchers who choose to maintain one should ensure that its contents are accurate and that sensitive information is handled appropriately.
The practical details of what an AI-use log should contain depend on the activities being documented and the purpose for which the record may later be needed.
Can AI Tool Account Records or Technical Logs Establish Use?
In some circumstances, authorized account records, institutional system logs, or platform records may provide information about AI-tool access and activity.
Such records may document account access, timestamps, or interactions, depending on the system and its retention practices.
However, an access log generally cannot establish what substantive content was generated unless the relevant content is also available.
Similarly, evidence that a researcher accessed an AI platform does not establish that the platform was used to prepare the manuscript under investigation.
Technical records should be interpreted according to what they actually capture.
Investigators must also respect applicable privacy, employment, institutional, and data-protection requirements. The possibility of obtaining technical information does not automatically justify collecting it.
Can Fabricated References or Factual Errors Prove AI Use?
No. Fabricated citations, nonexistent references, and unsupported claims may warrant investigation because they can compromise the reliability of research.
Generative AI systems are known to produce plausible but inaccurate information, including nonexistent references.
However, humans can also make citation errors, misattribute sources, or fabricate references without AI assistance.
The presence of a nonexistent reference therefore does not establish that AI generated the manuscript.
It may establish a separate problem requiring correction or investigation, depending on its seriousness and circumstances.
Investigators should avoid confusing evidence of an inaccurate manuscript with evidence of a particular mechanism that produced the inaccuracy.
The appropriate response is to examine the accuracy of the cited material and determine whether there is sufficient evidence of a policy violation.
Can a Researcher's Explanation Be Considered Evidence?
Yes. A researcher's account of the writing process is relevant evidence, although it should be evaluated alongside available records and other information.
An explanation may clarify why a passage was revised, how a tool was used, or why particular records are unavailable.
Investigators should assess whether the account is internally consistent and compatible with independently established facts.
They should not assume that an explanation is false merely because it conflicts with an AI-detector score.
Nor should they automatically accept an explanation that is contradicted by credible documentary evidence.
A fair investigation considers both supporting and contradictory information without treating the initial allegation as established.
How Should Different Types of Evidence Be Compared?
There is no universal hierarchy in which one document type is always stronger than another.
Evidence must be evaluated against the specific proposition being investigated.
| Evidence Type |
What It May Support |
Important Limitation |
| AI-detector report |
A particular system classified the text as resembling AI-generated writing. |
Does not independently establish actual AI use. |
| Dated manuscript drafts |
Chronology of writing and revision. |
May not reveal the original source of inserted text. |
| Document version history |
Changes to a document over time. |
Does not necessarily identify who composed the text. |
| Authenticated AI interaction |
An AI tool was used to request or generate particular content. |
Does not necessarily establish incorporation into the manuscript. |
| AI-use log |
Recorded purpose and extent of assistance. |
May be incomplete or depend on self-reporting. |
| Research notes and analysis records |
Development of ideas, methods, and interpretation. |
Do not establish human-only wording. |
| Relevant correspondence |
Contemporaneous discussion of manuscript development. |
May document only part of the writing process. |
| Verified correspondence between AI output and manuscript |
Possible incorporation of specific generated material. |
Requires authentication, chronology, and consideration of alternative sources. |
Evidence becomes more persuasive when several relevant records support a coherent account and plausible alternative explanations have been considered.
Why Does the Applicable AI-Disclosure Policy Matter?
Even convincing evidence that AI was used does not automatically establish misconduct.
The investigator must determine whether the activity required disclosure or was otherwise restricted under the policy applicable when the work was prepared or submitted.
The International Committee of Medical Journal Editors recommends disclosure of AI assistance in submitted work and emphasizes human responsibility for its accuracy and integrity.
COPE likewise states that authors should transparently disclose relevant AI use in manuscript preparation, images, and research activities.
However, journals and institutions may differ in how they classify grammar correction, substantive drafting, analysis, and other uses.
The investigator must therefore establish what assistance occurred, what policy applied, and whether the conduct satisfied the elements of the alleged violation.
Questions about which AI activities require disclosure should not be answered by assuming that every interaction with an AI system constitutes prohibited conduct.
What If the Researcher Cannot Produce the Requested Records?
The absence of records can limit the ability to reconstruct a writing process. It does not automatically establish that AI assistance occurred.
Researchers use different writing tools and document-management practices. Some software preserves extensive revision histories, while other systems retain only the current version.
AI providers may also delete interactions or change retention arrangements.
Investigators should distinguish between records that were required to be retained, records that would ordinarily exist, and records that were never created.
Where an applicable record-retention requirement existed, failure to comply may raise a separate question. However, that does not by itself prove a particular allegation of undisclosed AI use.
The institution should apply its governing burden and standard of proof rather than treating missing documentation as an automatic admission.
How Should Evidence Be Preserved During an Investigation?
Once an investigation is reasonably anticipated or initiated, relevant records should be preserved according to the applicable institutional and legal requirements.
Preservation may involve retaining original files, documenting their source, restricting unnecessary modification, and maintaining a record of how evidence was collected and handled.
For digital evidence, investigators may need to consider metadata, access permissions, file integrity, and whether exported records preserve relevant context.
These practices are particularly important when an AI conversation or document history may later be disputed.
The updated US Public Health Service research misconduct regulations under 42 CFR Part 93, applicable to allegations received on or after January 1, 2026, provide a formal framework for covered research misconduct proceedings and associated records. They should not be assumed to govern every university or journal investigation worldwide.
Investigators should follow the rules that apply to their particular institution and allegation.
Researchers may also benefit from understanding how to preserve AI interactions when providers may later change or delete them.
