If ChatGPT helped you outline, summarise prior work, or polish the language of a manuscript, the prose now reads like ChatGPT in places your reviewers will notice. TextSight runs a section-by-section scan against the same patterns Nature, Science, IEEE, ACS, Wiley, and Elsevier screeners look for, then helps you rewrite the flagged sentences in your own voice without touching citations, equations, or technical terms. Pre-submission sanity check and authentic-voice calibration, not a detector workaround.
A research paper is not one document. It is seven sections with seven different AI-tell profiles. Abstract and Discussion carry the highest risk because they are open prose. Methods carries the lowest because dense technical writing absorbs the template signal. The scan reflects that.
Submit the abstract alone. It is the highest-risk block in the manuscript: a desk editor and the assigned reviewers read it before anything else, and model-written abstracts follow a tight four-move template (background, gap, method, contribution) that classifiers weight heavily. Get the abstract reading in your own voice before you touch the body.
Introduction, Literature Review, Methods, Results, Discussion, Conclusion, each pasted on its own rather than as one manuscript. The highlights expose which paragraphs carry the assistant register. In most ChatGPT-assisted manuscripts the signal concentrates in two to four paragraphs, usually inside the Discussion and the literature framing, and the technical sections are already clean.
A manuscript is not edited at one setting. Light on Methods and Results, where a drifted number breaks reproducibility. Balanced on the Introduction and Discussion, where the hedging register lives. Keep Maximum off whole sections entirely; reserve it for the odd sentence that flags after a Balanced pass.
Re-submit the revised sections and confirm each one reads in the author voice. Then write the AI-use disclosure for the methods or acknowledgments your target journal requires. TextSight never touches a journal's submission pipeline and promises nothing about a specific screener; it reports its own reading of your prose and leaves the submit decision to you and your co-authors.
The AI rewriter was calibrated against a corpus of ChatGPT-assisted manuscripts across STEM, life sciences, and social sciences. The pattern that shows up in an Abstract is not the pattern that shows up in a Discussion. Knowing the profile helps you spend rewriting time where it matters.
ChatGPT abstracts follow a four-move template: background, gap, method, contribution. Each move is one sentence of 22 to 28 words, transitions are explicit. The fix is to compress background and gap into one sentence and lead with the finding, not the field. This is the single highest-yield rewrite in the manuscript.
The opening sentence is the biggest tell. "This paper presents," "This study investigates," "This paper proposes" appear in about 70 percent of generated introductions. Replace with the concrete problem or a finding that surprised you. The literature-context paragraph and the gap paragraph often read as separate template moves; merging them helps.
The highest over-flag section because it is citation-heavy and chronological. AI-generated lit reviews summarise one paper per sentence in citation order. Real reviews group three or four studies together by claim. Re-group by argument, keep citation tokens exact, and the section score usually moves 30 to 50 points without losing scholarly density.
The cleanest section by default. Dense technical prose with equations, variable names, and assay codes absorbs the template signal. Run Light mode only. If a sentence flags, rewrite it by hand rather than auto-rewriting, because precision-critical spans must survive the edit unchanged.
Results paragraphs that walk through tables read template by design, and that is fine; reviewers expect it. The flag risk is in the transitional sentences between table walks. The AI rewriter focuses on those and leaves the table-walk language alone.
The section that needs the most register attention. ChatGPT's hedging vocabulary ("Interestingly," "Notably," "These findings suggest," "Our results indicate") clusters here. The fix is to vary openings, anchor each paragraph in a specific number or a specific comparison to prior work, and name the limitation you actually worried about rather than a checkbox one.
Short and easy to rewrite from scratch if the score sits below 70. Drop "In conclusion," state the one finding that matters most, name the specific next experiment. The synthesis closer ("collectively underscore," "pave the way for") is one of the loudest tells in the manuscript and the easiest to remove.
Between 2024 and 2025, every major publisher updated its author guidelines on generative AI. The policies converge on the same line: AI assistance for outlining, summarising prior work, and language polishing is allowed if disclosed; AI-generated substantive content is not. A pre-submission scan catches sentences that cross that line before a reviewer does.
Disclosure required in methods or acknowledgments. LLMs may not be listed as authors. Internal classifier screening before peer review is documented at Nature and operates at several of the others without specific disclosure. A flag triggers an editor query about your AI use and can delay the review timeline by weeks.
AI-use statement required on every submission, naming the model and the sections it touched. ACM extends the policy to revisions, conference papers, and workshop submissions. Both bodies reserve the right to screen submissions for undisclosed AI use and to query authors when a section reads generated, which is the practical reason to run a pre-submission scan and clean up your own language first.
Policies tightened in early 2025. ACS prohibits AI use for "creating or altering scientific content" and screens with both internal and third-party tools. Elsevier, Wiley, and Springer require disclosure across their journal portfolios. PLoS journals require a specific statement about whether AI tools contributed to text, images, or analysis.
The AI scan covers one risk; similarity screening covers a different one. Most journals run iThenticate or a Crossref Similarity Check report on submissions, which compares your manuscript against published literature and detects plagiarism or self-plagiarism. Pre-flighting both before submission is the sober move; the two reports rarely overlap and together they cover most of what desk review actually checks.
For academic prose the mode choice matters more than for any other content type. Maximum can flatten the formal voice journal reviewers expect, so the default we suggest is conservative and section-specific. Different sections want different modes within the same manuscript.
Light protects what a reviewer checks for reproducibility: reagent quantities, instrument settings, statistical tests, p-values, sample sizes, and the exact phrasing of a hypothesis. It edits only the connective prose between those spans. Run it on Methods and Results, where a paraphrase that drifts a single number is a worse outcome than a sentence that still reads slightly templated.
Balanced reworks the hedging register that clusters in the Discussion and the Introduction gap paragraph. It varies the openings ChatGPT repeats ("Interestingly", "These findings suggest"), re-paces the prose, and pushes vocabulary off the model default while leaving your citation tokens and prior-work comparisons untouched. This is the section where reviewers form their read of whether the author actually reasoned through the results.
Maximum paraphrases hard enough to risk the formal academic register a desk editor and a peer reviewer expect, and it can blur a carefully bounded claim into a vaguer one. The cost of a softened claim in a manuscript is higher than in any other format, because a reviewer reads imprecision as either sloppiness or overreach. Limit Maximum to one or two stubborn red sentences after a Balanced pass, then re-read each for whether the bounded claim survived.
The sequence that works across a full manuscript: Light on Methods and Results, Balanced on Introduction and Discussion, Light then a targeted Balanced on the Abstract. The Conclusion is short enough that rewriting it by hand from your own findings is usually faster, and safer, than running any mode on it.
All three modes available on every paid plan. Full details on the pricing page.
Billed $89.88/year — Save $30
Billed $179.88/year — Save $60
Billed $359.88/year — Save $120
Yearly billing saves 25%. View full pricing →
Research papers are the use case where the line between legitimate AI-assisted writing and academic dishonesty matters most, because the reputational and disciplinary stakes are highest. We want to be explicit about which side of that line we are on.
It belongs at the language-polish step of a manuscript you and your co-authors designed, ran, and analysed, where a model helped outline a section, summarise prior work, or smooth non-native phrasing inside your journal's disclosure policy. The experiments are yours, the data are yours, the interpretation is yours. The AI rewriter only catches sentences where the assistant register settled into the connective prose, so the submission reads in the author's voice rather than the model's. Think language-editing service, not ghostwriter.
We promise nothing about getting a specific submission past a publisher's screener, and we say so plainly. If a section is mostly model output thinly edited, the scan reports that, and no mode can supply analysis that was never performed. The score and the highlights are a diagnostic on your own prose, not a way to disguise work you did not do. Reviewers test claims against data; a smoother sentence does not change whether the underlying result holds.
Whatever the model touched, state it in the methods or acknowledgments exactly as your target journal requires, and never list an LLM as an author. Undisclosed use is the serious integrity breach; disclosed-and-cleaned-up language assistance is routine. Running the AI rewriter does not replace that statement. It is the step you run before you write the statement, not a reason to skip it.
Generating the substance of a paper with a model, attaching author names, and submitting it. That violates the policies of every major publisher no matter which rewriter processed the text, and it puts your co-authors and your record at risk. If that is the situation, the honest path is to use the detector to see which paragraphs read generated, then rebuild them from the experiments and reasoning you genuinely carried out.
Vetting TextSight for a group: treat it like the language-editing services labs already pay for, scoped to disclosed-use polish on the group's own work, not to disguising generated substance. The shared scan history on the Business tier gives a corresponding author an audit trail of which drafts went through the tool, which is the accountability layer a supervising PI usually wants.
The detector workflow tuned for thesis chapters, journal submissions, and grant prose.
Open the detector →The shorter-form coursework version of this workflow, calibrated for student essays.
Open essays page →The flagship AI rewriter page covering all source content. Three modes, closed-loop calibration.
Open AI rewriter →How the score is computed and what threshold to aim for before journal submission.
Read the guide →Free to try, no card. Section-by-section workflow, three modes, citations and equations preserved, ESL-aware calibration.
Same Authenticity Score, tuned for other formats and surfaces.