Generative AI can improve the efficiency and linguistic clarity of scholarly writing, but it cannot transfer scientific judgment, authorship, or accountability away from researchers. For authors targeting Q1 and Q2 journals, the relevant standard in 2026 is not whether a manuscript contains AI-assisted work it is whether that work is transparent, verifiable, appropriately bounded, and fully owned by the human authors.
This guide provides an operational framework for doctoral candidates, postdoctoral fellows, faculty members, and research-integrity officers. It synthesizes the shared direction of COPE, ICMJE, IEEE, and major publisher policies, then translates that direction into a defensible manuscript workflow: document use, preserve provenance, verify every claim and citation, audit textual overlap at source level, and disclose AI use in the exact form required by the target journal.
Where AI fits in scholarly writing today
AI is now part of the normal research-writing environment. It may assist with proofreading, translation, organization, code support, literature triage, and more controversially text generation. Yet the governing principle across major publishing policies is stable: an AI system is a tool, not an author. It cannot approve a submitted version, accept legal or ethical obligations, declare conflicts of interest, hold copyright, or respond to an allegation of misconduct. Those responsibilities remain with named human authors.
This distinction matters because manuscript integrity is broader than prose quality. A paper is a record of how a research question was framed, how data were produced and analysed, how prior work was interpreted, and who can defend each decision. AI can help a researcher express that record; it cannot credibly stand behind it.
The practical implication is straightforward. Basic, human-controlled assistance for spelling, punctuation, or readability may be treated differently from substantive generation or restructuring of text. However, policy thresholds are journal-specific. When AI changes the intellectual or rhetorical substance of a manuscript or contributes to data analysis, code, figures, images, or content generation - authors should presume that disclosure and documented oversight are required.
Why this matters for Q1 and Q2 submissions
Selective journals conduct more than a superficial language review. Editorial offices assess scope, originality, ethics declarations, authorship, conflicts of interest, data availability, citation quality, image integrity, and textual overlap. A manuscript can be technically competent yet invite avoidable scrutiny if it contains unsupported references, unexplained AI use, recycled passages, or generic language that obscures the authors' actual contribution.
The risk is not confined to a single detector score. Similarity software identifies matched text against its indexed corpus; AI-detection approaches may use stylometry and other linguistic signals to estimate whether individual passages resemble machine-generated or AI-paraphrased writing. Neither output independently determines misconduct. Both are evidence that requires context, source review, and human judgment. A sound pre-submission workflow therefore examines the passage, the source, the citation, the drafting history, and the journal's policy not merely a percentage.
Assistive and generative AI: the distinction that matters
The most useful distinction is functional rather than technological. Assistive AI helps improve human-authored material for example, correcting grammar, identifying typographical errors, checking consistency, or suggesting clearer phrasing. Generative AI produces new text, code, images, analyses, or structured content in response to prompts. A tool can perform either role depending on the task.
A grammar correction that changes "The results indicate" to "The results indicate that" is not equivalent to asking a model to draft a literature review, formulate a clinical interpretation, derive an analytic rationale, or rewrite a source passage to avoid textual overlap. The latter activities can affect authorship-relevant intellectual work, factual accuracy, originality, or attribution. They demand greater scrutiny and usually explicit disclosure.
Researchers should also distinguish raw AI output from AI-paraphrased writing. Raw output may exhibit generic transitions, overconfident claims, repetitive syntactic patterns, or broad assertions detached from the study's actual data. AI-paraphrased writing may be lexically altered but still retain the conceptual sequence, argument architecture, or distinctive expression of a source. The second case is especially important: changing words does not remove the obligation to cite the originator of an idea, finding, method, or interpretation.
Authorship as an accountability framework
COPE states that AI tools cannot be listed as authors because they cannot take responsibility for submitted work, manage conflicts of interest, or enter into copyright and licensing agreements. ICMJE likewise places responsibility for material produced with AI-assisted technologies on human authors. IEEE's authorship framework reserves credit for humans who make a significant intellectual contribution, participate in drafting or intellectual revision, and approve the accepted version.
These rules should not be read as merely formal restrictions. They protect the traceability of scholarly claims. When an editor asks who selected a dataset, who interpreted an outlier, who decided that a limitation was material, or who verified a contested reference, a named author must be able to answer. Listing an AI tool cannot resolve that obligation; it obscures it.
Disclosure is not a confession
Transparent disclosure is often misunderstood as an admission that a manuscript is less authentic. It is better understood as methods reporting for the writing and research process. A clear declaration enables editors and readers to assess whether a tool's role was appropriate and whether the authors retained intellectual control.
COPE recommends transparency when AI tools are used in manuscript writing, image production, or data collection and analysis. Elsevier requires a declaration for substantive use in manuscript preparation and distinguishes this from basic grammar, spelling, and punctuation checks. IEEE requires disclosure of AI-generated content in acknowledgements, including the system used, the sections affected, and the level of use; editing and grammar enhancement are generally outside the core requirement, though disclosure is recommended. ICMJE guidance similarly requires authors to describe AI use and preserve full human accountability.
The target journal's instructions always control. Before drafting a declaration, consult the journal's current author guide, submission system prompts, and any policy governing supplementary materials, visual abstracts, code, peer review, or use of proprietary data.
Plagiarism, textual overlap, and originality
Similarity is not plagiarism; it is an indicator of textual overlap. Common phrases, standard method descriptions, legally required statements, reference lists, and appropriately quoted passages can all produce matches. Conversely, a low similarity index does not establish originality if an author has reproduced another scholar's idea, structure, or reasoning without attribution.
The core question is therefore not "What is the overall percentage?" but "What is this sentence, where did it come from, and has it been cited and represented accurately?" Verbatim plagiarism, mosaic or patchwork plagiarism, close paraphrase without attribution, self-plagiarism, redundant publication, and idea plagiarism require contextual assessment. AI paraphrasing cannot repair any of them. Indeed, using a rewriter to conceal borrowed prose may add evidence of deliberate evasion.
A credible audit is source-grounded. The SciCampus AI Detector pairs a sentence-level stylometry assessment with a search grounded similarity index and direct DOI-linked sources. This lets the author move from a matched sentence to the underlying publication, confirm whether the source is correctly cited, and decide whether the text needs quotation, attribution, substantive rewriting, or removal. That is more defensible than relying on an unexplained aggregate score.
Stylometry and detection: useful but limited evidence
Stylometry examines measurable characteristics of writing: function-word use, vocabulary diversity, sentence length, punctuation, readability, and recurring syntactic patterns. These features can help assess whether a passage resembles a known writing style or a machine-generated distribution. But stylometry is not a provenance archive and cannot establish intent. Performance varies with discipline, language, document length, model evolution, editing, and the mixture of human and machine-authored text.
For this reason, authors should reject both extremes: "detectors are infallible" and "detectors are meaningless." A detector can identify passages worth reviewing, especially when combined with drafts, notes, sources, and an AI-use log. It should not be used as a single black-box verdict on an author's integrity.
Sentence-level probability bands support a more disciplined review. Rather than treating one manuscript-wide number as dispositive, researchers can inspect red bands (at or above 99%), orange bands (at or above 96%), and yellow bands (at or above 93%) sentence by sentence. SciCampus also distinguishes probable raw AI-generated text from AI-paraphrased writing. These categories should trigger targeted review: check whether the sentence is a generic generated claim, a paraphrase that needs a citation, or a valid human sentence that warrants supporting provenance if challenged.
How the major policies compare
Policy sourceAuthorship positionDisclosure expectationKey implication for authorsCOPEAI cannot be an author because it cannot carry accountabilityDisclose AI used for writing, images, or research processesHuman authors remain liable for accuracy and ethicsICMJEAI tools are not authorsDescribe how and where AI was usedPreserve responsibility for originality, plagiarism, and copyrightIEEEAI cannot be author or co-authorIdentify AI system, affected sections, and level of generationUse acknowledgements as directed; retain human approvalElsevierAI may support writing, not replace critical judgmentUse a dedicated declaration for substantive manuscript useVerify outputs; protect confidential inputs; disclose research-process use in Methods
Test the evidence, not a black-box score. SciCampus is free to start no card required. Review sentence-level stylometry, DOI-linked similarity evidence, and the passages that actually need editorial attention. Create a free account.
An eight-step workflow for AI-assisted manuscripts
Step 1: Set an AI-use protocol before drafting
At project launch, agree on permitted and prohibited uses. The protocol should specify whether tools may be used for proofreading, translation, outlining, code assistance, transcription, literature discovery, or data analysis. It should prohibit fabrication of data, images, citations, or participant information; uploading sensitive or confidential material to unapproved tools; and undisclosed generation of substantive scientific claims.
For multi-author papers, document who is responsible for each high-risk activity: source checking, reference management, statistical validation, figure production, AI logging, and final policy review. This prevents a common failure mode in which every author assumes another person verified the AI-assisted text.
Step 2: Maintain an AI-use and provenance log
A minimal log should include the date, tool and version, purpose, input type, output used or rejected, manuscript section affected, reviewer, and disposition. Keep prompts when AI produced substantive text, code, analytical outputs, or images; basic spell-checking may be recorded at a higher level if policy permits.
The log is not bureaucracy for its own sake. It supports a truthful disclosure statement, helps co-authors reconstruct revisions, and provides evidence if an editor questions a flagged passage. Pair the log with version history from your word processor, reference manager, code repository, or electronic lab notebook.
Step 3: Write the scientific core as human work
Human authors should establish the research question, theoretical framework, inclusion and exclusion decisions, methods, analysis plan, results interpretation, limitations, and conclusions. An AI tool may assist with expression, but it must not decide what your data mean.
This is particularly important in the IMRaD structure. In the Introduction, ensure that the gap and contribution arise from genuine engagement with the literature. In Methods, report reproducible procedures rather than polished generalities. In Results, report what the analysis shows, not what a language model expects a results section to say. In Discussion, distinguish observed findings from speculation, and preserve limitations that an AI rewriter may otherwise smooth away.
Step 4: Use AI for editing under human control
A safe editing request is bounded: "Correct grammar and improve clarity. Do not add claims, citations, statistical interpretations, or changes in meaning." Review every revision against the source text, data, and cited literature. Never accept generated references without locating and reading the actual record, preferably through its DOI or publisher page.
Non-native English speakers should not be disadvantaged for seeking language support. Fluency assistance is legitimate when it clarifies human-authored scholarship without replacing the author's ideas or introducing unsupported content. The necessary safeguard is substantive review: the author must understand, endorse, and be able to defend every changed sentence.
Step 5: Audit attribution before paraphrasing
Do not paste a source passage into a paraphraser and treat the result as original prose. First determine what is being borrowed: a specific finding, a theoretical claim, a method, a definition, a data point, or distinctive wording. Cite the source. Use quotation marks for language that remains materially close. Then write the point in the logic of your own argument, adding your interpretation rather than merely reordering another author's sentences.
Improper AI-assisted paraphrase
Source: "Stylometric methods can distinguish AI-generated texts from human writing through lexical and syntactic features."
Manuscript: "Lexical and syntactic signals enable stylometry to separate machine-written text from writing produced by people."
Without a citation, this preserves the original proposition and structure while disguising its wording.
Appropriate, attributed use
Stylometric approaches examine lexical and syntactic distributions to estimate whether passages resemble AI-generated or human-authored writing; such estimates should be interpreted as probabilistic evidence rather than conclusive proof of provenance (Author, Year).
The appropriate version names the intellectual source, supplies the author's own interpretive boundary, and avoids claiming more than the evidence supports.
Step 6: Conduct a DOI-linked similarity review
Run a similarity audit before the final internal review, not five minutes before submission. In SciCampus, inspect the search-grounded similarity index, then open its DOI-linked matches. Work through each substantive match at sentence level and assign one of four dispositions: correctly quoted and cited; correctly paraphrased and cited; standard language requiring no change; or revision required.
For each revision-required match, record what changed and why. Check especially the abstract, introduction, methods boilerplate, figure legends, limitations, and any reused material from a thesis, preprint, grant, conference proceeding, or prior article. Do not assume self-authored prose is automatically reusable; journals may require disclosure, quotation, permission, or a more substantial rewrite.
Step 7: Review stylometry by sentence, not by score alone
Use the AI Detector as a triage instrument. Review red, orange, and yellow sentences in context. A red band may reflect copied raw AI output, but it might also flag a highly formulaic scientific sentence; an orange or yellow passage could represent heavy AI paraphrasing, conventional prose, or a false positive. The appropriate response is investigation, not cosmetic "humanizing."
Ask four questions for each flagged passage:
Is the scientific claim accurate and supported by the study data or a cited source?
Does the language reflect the author's genuine reasoning and disciplinary voice?
Does the passage resemble an external source in wording, sequence, or conceptual framing?
Can the author provide notes, drafts, source records, or an AI-use log that explains its origin?
If a sentence is generic, replace it with study-specific information. If it is unsupported, remove or substantiate it. If it is a close paraphrase, cite and genuinely reframe it. Do not rewrite solely to defeat a detector: the aim is provenance, clarity, and integrity.
Step 8: Draft a journal-specific declaration
Place the declaration where the journal directs often in Methods, Acknowledgements, a separate AI declaration, and/or the cover letter. It should be precise but proportionate.
Example: substantive language assistance
The authors used [tool name, version] to identify grammar and clarity revisions in portions of the manuscript. All suggested changes were reviewed, edited, and approved by the authors, who take full responsibility for the accuracy, originality, and integrity of the final text.
Example: research-process use
[Tool name, version] was used to assist with [specified analytic or coding task]. The workflow, inputs, validation procedures, and human oversight are reported in the Methods section. The authors verified all outputs and retain responsibility for the analysis and conclusions.
Never use a generic declaration to conceal material use. A transparent statement is shorter, more credible, and easier for an editor to assess.
Turn review into a defensible workflow. Start free on SciCampus no card required to map sentence-level AI signals, investigate DOI-linked matches, and create evidence before the manuscript reaches an editor. Run your first analysis.
Five misconceptions that get manuscripts flagged
Mistake 1: Treating a low similarity score as clearance
There is no universally "safe" percentage. A similarity index depends on corpus coverage, exclusions, quotations, reference lists, methods language, and editorial thresholds. More importantly, it cannot evaluate whether an idea was properly attributed. Review the passages and sources; do not optimize for a number.
Mistake 2: Treating AI disclosure as a substitute for verification
Disclosure does not excuse fabricated references, flawed calculations, copied prose, biased interpretation, or undisclosed conflicts. An author who discloses AI use remains responsible for every statement. Verification requires reading cited sources, reproducing key calculations, checking tables and figures against raw outputs, and confirming that AI edits did not alter certainty, sample size, units, effect estimates, or causal language.
Mistake 3: "Humanizing" text to evade detection
Deliberately varying sentence length, inserting errors, or using rewriting tools to escape a detector is not an integrity strategy. It can create awkward prose, erase methodological precision, and suggest an attempt to conceal provenance. Revise because a sentence is inaccurate, generic, poorly supported, improperly attributed, or not genuinely yours not because a colour band is inconvenient.
Mistake 4: Assuming a false positive needs no response
A false positive does not establish misconduct, but dismissing it without a review is unwise. Preserve drafts, timestamps, tracked changes, source notes, and AI-use records. If an editor asks, respond calmly with a concise explanation of your process and targeted documentation. Request a contextual human review rather than arguing that any detector is universally invalid.
Mistake 5: Diffusing accountability in large teams
A shared document can conceal who generated or approved a passage. Each author must review work relevant to their contribution, and the corresponding author should lead the final integrity review. Agree in advance who runs the pre-submission audit, and keep the resulting reports with the project record so a lab can trace what was screened and when.
Preparing a case file for an editor or committee
If a manuscript is flagged, do not submit a mass of unstructured screenshots. Prepare a concise, dated case file containing:
A cover note identifying the questioned passages and your response.
Version history showing how the manuscript developed.
The AI-use log, including tool, purpose, dates, human reviewer, and final disposition.
Source notes and DOI records for cited or matched material.
A sentence-level similarity and stylometry review annotated with corrective actions.
The final disclosure statement and relevant correspondence.
A downloadable highlighted PDF evidence report can serve as the audit exhibit: it preserves sentence highlights, probability information, and linked similarity evidence in a shareable format. It is not proof that an author did or did not use AI. It is a defensible case file demonstrating that the team reviewed the relevant signals, located the underlying sources, and made documented editorial decisions.
Pre-submission checklist
Use this checklist as a final gate for manuscripts, dissertations, or major revisions.
A. AI use and data protection
The team has defined permitted and prohibited AI uses for this project.
All substantive AI use is logged with tool, version, purpose, affected sections, reviewer, and outcome.
No restricted data, confidential manuscript text, peer-review material, personal information, or unlicensed content was sent to an unapproved AI service.
AI was not used to fabricate, manipulate, or selectively omit data, images, code, or citations.
Every AI-assisted claim, calculation, reference, and figure has been independently verified by an author.
B. Authorship and contribution integrity
Every listed author meets the target journal's authorship criteria and approved the final version.
No AI tool appears in the author line, CRediT taxonomy, or contributor list.
Human contributors' roles are accurately recorded.
The corresponding author has completed a final authorship, conflict-of-interest, and disclosure review.
C. Similarity and attribution audit
A complete similarity review has been conducted before submission.
Each substantive matched passage was examined against its DOI-linked source or an equivalent authoritative record.
Quotations are marked and cited; paraphrases accurately attribute the original source.
Overlap with theses, preprints, conference papers, protocols, grants, and prior publications has been disclosed or revised as appropriate.
The manuscript has not been mechanically paraphrased to reduce textual overlap.
D. Sentence-level stylometry review
Detector results have been reviewed at sentence level rather than interpreted as a single manuscript-wide verdict.
Red (≥99%), orange (≥96%), and yellow (≥93%) bands have been investigated in context.
Flagged passages have been checked for accuracy, study specificity, citation, provenance, and authorial reasoning.
Revisions addressed substantive problems rather than merely attempting to change detector output.
Evidence reports, version history, and the final AI-use log are retained with the project record.
E. Journal fit and declarations
The target journal's current AI, authorship, image, data, and peer-review policies have been reviewed.
The manuscript follows the journal's declaration location and wording requirements.
The cover letter accurately identifies relevant AI use when requested.
References resolve correctly, contain verified DOI information where available, and have been checked against original sources.
The manuscript's scope, reporting guideline, data-availability statement, ethics approval, and supplementary files meet the target journal's instructions.
The final manuscript audit
A strong final audit works from claim to evidence, not from score to score. Read the manuscript as a skeptical reviewer would. Mark every major assertion in the abstract, introduction, discussion, and conclusion. Can you locate the supporting data, analysis, or cited source? Does the wording accurately reflect its strength? Did an AI edit convert an association into a causal claim, omit a limitation, or introduce polished but empty generalization?
Then inspect the document's provenance. Use DOI-linked similarity findings to examine meaningful overlap. Use sentence-level stylometry bands to target passages requiring human review. Confirm that drafts and logs tell a coherent story of authorship and oversight. Finally, compare the disclosure statement with what actually occurred. Consistency across these records is the foundation of an editor-ready submission.
Building a lab-level standard
Research groups benefit from converting this workflow into a standing standard operating procedure. Establish an approved-tool list, a data-classification rule for AI inputs, a minimal logging template, and an internal review gate for manuscripts. Train early-career researchers to differentiate citation from paraphrase, similarity from plagiarism, detector probability from proof, and editing from ghostwriting.
Preserve run history, DOI-linked source checks, and downloadable PDF reports as part of the manuscript's internal record, so that a screening decision made months earlier can still be explained.
Frequently asked questions
Can I list an AI tool as a co-author?
No. COPE, ICMJE, and IEEE all exclude AI systems from authorship because a tool cannot take responsibility for the work, manage conflicts of interest, hold copyright, or answer an allegation of misconduct. Credit AI use in a disclosure statement or acknowledgement instead, in the location your target journal specifies.
Do I need to disclose using AI for grammar and spelling?
Usually not, but it depends on the journal. Elsevier and IEEE both distinguish basic language correction from substantive generation, and generally do not require a declaration for the former. Because thresholds differ between publishers and change over time, check the current author guide of your target journal before you submit.
What similarity percentage is acceptable?
There is no universal threshold. A similarity index reflects corpus coverage, quotation handling, reference lists, and standard methods language not misconduct. A 25% index composed of correctly cited quotations may be fine, while a 5% index containing one uncited close paraphrase is not. Review the matched passages and their sources rather than optimizing a number.
Can an AI detector prove that I used AI?
No. Detection tools estimate whether a passage statistically resembles machine-generated writing. They are probabilistic evidence, not a provenance record, and their accuracy varies with discipline, language, document length, and editing. Treat a flagged sentence as a prompt for review, and keep drafts, version history, and an AI-use log as your own evidence.
Is it safe to use a paraphrasing tool to lower my similarity score?
No. Rewriting borrowed text without citing it does not resolve plagiarism; it conceals it, and may itself be read as evidence of deliberate evasion. If a passage matches a source, the correct response is to quote, cite, genuinely reframe, or remove it not to reword it until the match disappears.
What should I do if my manuscript is flagged unfairly?
Respond with documentation rather than argument. Prepare a dated case file containing the questioned passages, your version history, your AI-use log, source notes with DOI records, and an annotated sentence-level review. Ask for a contextual human assessment instead of disputing the validity of detection tools in general.
The standard worth meeting
Responsible AI use is not achieved by avoiding every tool or by trusting every tool. It is achieved through bounded assistance, transparent disclosure, source-level verification, documented human oversight, and an editorial record that can withstand informed scrutiny. That is the standard that protects researchers, strengthens peer-review confidence, and gives high-quality work the best chance to be assessed on its scientific merits.
Publish with evidence, not uncertainty. Start free on SciCampus no card required and run a sentence-level, DOI-grounded integrity audit before your next submission. Get started.
Editorial note: Always follow the current instructions of your target journal and institution, which may be more restrictive than general publisher guidance. Policy pages change; verify requirements immediately before submission.
