If ChatGPT helped you outline, summarise prior work, or polish the prose on a federal R01 narrative, an NSF specific aims, a Wellcome programme application, or a MacArthur letter of inquiry, the proposal now reads like ChatGPT in the places reviewer panels learn to spot first. NIH, NSF, NIH-NIAID, DOE, Wellcome Trust, Ford, MacArthur, Open Philanthropy, and the Gates Foundation now require human-authorship attestation or AI disclosure on the cover form. TextSight runs a section-by-section scan against the six reviewer-facing patterns that flag, then helps you rewrite the lines that drift in your organisation's voice. Pre-submission sanity check and authentic-voice calibration, not a detector workaround.
Between 2024 and 2025, every major federal agency and most large private foundations updated their submission policy on generative AI. The policies converge on one line: AI assistance for outlining, literature synthesis, and language polishing is allowed when disclosed; AI-generated substantive content is not, and an AI-flagged proposal can be rejected outright.
Human-authorship attestation required on submitted narratives as of 2025. NIH-NIAID added explicit research-integrity language flagging AI-generated content. Specific aims and approach sections attract the closest reviewer review because templated approach paragraphs dilute the experimental hypothesis. Programme officers do not confirm automated detection on every R01, but reviewer panels increasingly note AI-shaped prose in their critiques, and several review sections are briefed on common AI tells before the meeting.
Wellcome Trust requires AI disclosure on the cover form for major research programmes. MacArthur asks applicants to attest to authorship on the application portal. Ford added AI disclosure to its grant submission system during 2025. All three reviewer panels include AI-shaped prose in their evaluation notes regardless of whether the formal policy text mentions it. The defensible posture is to pre-scan, rewrite reds, then disclose honestly on the cover form.
Open Philanthropy and the Gates Foundation added AI disclosure questions to their application portals during 2024 and 2025. Community foundations and smaller family foundations are following at varying speeds. Treat every funder as if AI disclosure matters and let the scan history be the evidence trail that supports an honest disclosure or a no-AI claim.
A flag is rarely fatal on its own, but reviewers who suspect AI use score the application more harshly on innovation, significance, and investigator capability. An integrity query from a programme officer can delay a review cycle by weeks, and a confirmed undisclosed-use finding ends the application and damages the relationship with the agency. For a six-figure ask, the pre-submission scan is the cheapest insurance you can buy.
These are the six patterns that recur across ChatGPT-drafted proposals and that reviewer panels learn to spot within the first page. Detectors catch the first four on their own. The last two are the ones experienced programme officers notice without any tool.
Outputs feeding into outcomes feeding into impact is templated by definition, and ChatGPT writes it with identical clause structure across three paragraphs. The risk is not the framework, which most funders ask for; it is the identical sentence rhythm across each link in the chain. Vary the construction, embed concrete actor language, and let the sentence map confirm the rewrite cleared the threshold.
Three or four consecutive sentences each opening with a statistic and a clause about underserved populations read as generated paragraphs rather than authored narrative. Reviewers spot the pattern on a scroll-through. Compress to one statistic per paragraph, anchor the rest in a specific community, partner organisation, or evaluation finding from your prior work.
"Comprehensive, evidence-based, community-driven." "Innovative, scalable, sustainable." "Rigorous, equitable, replicable." Tripled adjectives are an AI tic reviewers spot on a scroll-through. Each adjective does no work on the page. Replace with the concrete mechanism: who benefits, on what timeline, against which baseline.
"Systemic," "transformative," "scalable," "synergistic." ChatGPT defaults to these words whenever it needs to flag an initiative as serious. Reviewers parse them as evidence the team has not figured out the mechanism. Name the handoff: what data does Year 1 produce that Year 2 needs, on what timeline, with which partner.
Most grant narratives ship from three to seven contributors: lead PI, co-PIs, programme staff, evaluator, finance for budget narrative, and sometimes a contracted writer. Each contributor enters AI assistance at a different point and the combined draft reads uneven, which itself reads as AI. The sentence map surfaces the paragraphs where one author leaned on ChatGPT harder than the rest.
Stock language about governance, oversight, and organisational capacity recycled from board materials, then polished by ChatGPT, reads as boilerplate to programme officers who read fifty applications a season. The fix is to replace governance abstractions with one specific decision your board made about this work, and one specific way your evaluator will report against it.
A grant narrative is not one document. It is seven sections with seven different AI-tell profiles. Need Statement and Project Description carry the highest rhetorical risk. Methods absorbs the template signal because of dense procedural detail. The scan reflects that and recommends a mode per section.
The highest-yield single rewrite in the proposal. ChatGPT need statements stack three or four statistics with underserved-population clauses; reviewer panels spot the pattern within the first paragraph. Aim for an Authenticity Score above 75 on the need statement before you move on. Compress to one statistic per paragraph and anchor the rest in your organisation's lived work with the population.
The narrative spine of the proposal. ChatGPT's tripled adjectives and theory-of-change abstractions cluster here because the section invites broad claims about impact. The fix is to name the concrete mechanism in each paragraph, vary openings across the section, and let the sentence map confirm. Aim for above 75 across the section.
The cleanest section by default. Dense procedural prose with assay codes, equipment models, statistical tests, and partner-protocol numbers 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. Aim 55 to 70 here and rely on the sentence map, not the headline score.
Templated by definition because outputs-outcomes-impact is the framework most funders ask for. The risk is identical clause structure across three paragraphs, which reads as generated even when each link in the chain is real. Run Light mode, vary construction, embed concrete actor language. Aim 60 to 75.
Evaluators are usually less practised at sentence-level voice than the lead PI, and the section carries domain vocabulary that overlaps with templated AI phrasing: formative, summative, mixed-methods, developmental. Run Balanced mode on the narrative paragraphs and Light mode on the instrument descriptions. Aim above 70.
Notoriously stiff by default, and ChatGPT cleans it into something even stiffer. Run Light mode to vary rhythm in the connective prose while keeping dollar figures, percent effort, calendar-month allocations, personnel names, and IRB protocol numbers as quoted spans. Pull one example: instead of "personnel costs include the PI and a graduate student," say "0.6 calendar-month PI effort plus one 12-month graduate student stipend at the institutional rate."
The section PIs hate writing and therefore the section most heavily AI-edited in a final draft. Stock language about diversified funding, long-term partnerships, and institutional commitment recycles word for word across ChatGPT drafts. Run Balanced mode, then add one specific sentence about what happens to the resource once funding ends and which partner takes on which line item.
Most grant packages clear three internal checks before they reach the funder: the writer's own draft pass, the IRB or development office compliance review, and the institutional grant office sign-off. The TextSight pre-submission scan fits before the first internal review, so the version your IRB and grant office read is already AI-clean.
Write in your usual editor: Word, Docs, or the funder portal. Using ChatGPT for an outline pass, a literature synthesis, or to break writer's block on a difficult section is the realistic 2026 default and inside the disclosure regime most funders allow. Compose the prose itself in your own voice or the organisation's voice from your notes, prior funded proposals, and conversations with programme staff.
Open app.textsight.ai, paste each narrative section, and scan. Free tier handles 5,000 characters in one paste. Pro handles 10,000, which fits a typical narrative subsection like a need statement or a logic-model paragraph. For a 12-page R01 Research Strategy, split by section: specific aims, significance, innovation, approach, broader impacts. The scan returns in well under a minute with an Authenticity Score and a sentence-by-sentence colour map.
Need statement, project description, sustainability: above 75, submit; below 75, rewrite reds in organisational voice. Methods, logic model, evaluation instruments: above 55, the sentence map is the signal, not the score. Budget narrative: above 70, rewrite reds rather than humanising stock phrasings around line items.
Send the consolidated package to your IRB, development director, or institutional grant office for compliance review. Pro history keeps every section scan for 90 days. PDF export gives you a contemporaneous record useful if a programme officer asks downstream how the narrative was produced. A typical narrative subsection round-trips in about fifteen minutes; a full 12-page narrative in about ninety.
For grant narrative the mode choice matters more than for any other content type. Maximum can flatten the organisational voice reviewer panels score on, so the default is conservative and section-specific. Different sections want different modes within the same proposal.
Light makes mild edits and preserves precision: assay codes, equipment models, dollar figures, percent effort, calendar-month allocations, personnel names, IRB protocol numbers, named partner institutions. Score gains per pass are smaller, but the output still reads like a proposal you would send to NIH. This is the starting mode for Methods, Logic Model, and Budget Narrative.
Balanced runs moderate rewrites and shifts vocabulary and rhythm more aggressively than Light without flattening voice. It is the right choice for Need Statement, Project Description, Evaluation Plan narrative paragraphs, and Sustainability. The places where ChatGPT's hedging register and tripled adjectives cluster are exactly the places where Balanced helps most.
Maximum runs the most aggressive rewrite. The caveat is real on grant narrative: aggressive rewrites can flatten the organisational voice reviewer panels score on, replacing your distinctive phrasing with generic patterns that read flat for an experienced programme officer. Use Maximum on isolated red sentences after a Balanced pass has already done the work, never on a whole section, and never on Specific Aims.
The recommended sequence for a full proposal: Light on Methods, Logic Model, and Budget Narrative; Balanced on Need Statement, Project Description, and Sustainability; scan Specific Aims with Light and rewrite by hand because the PI voice on that one page is part of the scoring rubric.
All three modes available on every paid plan. Pro at $19.99 a month standard, $14.99 a month on yearly billing, fits solo grant writers and consultants pushing 20 to 50 applications a year. Business at $39.99 a month, or $29.99 a month on yearly, fits nonprofit development offices and research administration. Full details on the pricing page.
Billed $89.88/year — Save $30
Billed $179.88/year — Save $60
Billed $359.88/year — Save $120
One avoided rejection on a six-figure ask covers either tier. View full pricing →
Grant proposals are the use case where the line between disclosed AI-assisted writing and breaching funder attestation matters most, because federal agencies and major foundations treat undisclosed AI authorship as a research-integrity issue. We want to be explicit about which side of that line we are on.
Proposals you authored, where ChatGPT was used as an outline assistant, a literature summariser, or a language polisher inside your funder's disclosure regime. The research is yours, the budget is yours, the partnerships are yours, the analysis is yours. The AI rewriter helps you catch sentences where the assistant register leaked into the prose so the submitted narrative reads in your organisation's voice rather than the ChatGPT voice. This is closer to a careful proofread than to anything else.
We make no promise that TextSight will get any specific proposal past an NIH integrity screen, an NSF programme officer's read, or a Wellcome reviewer panel. We report our own score honestly and explain what it means. If a section is mostly ChatGPT and only lightly edited by you, our scan will tell you that and the AI rewriter will not magically fix it; it cannot put authentic analysis that was not there. The score and the highlights are diagnostic, not laundering.
Even after authenticity, if you used ChatGPT for outlining, lit-review summarising, or language polishing, disclose it on the cover form or in the methods as your funder's policy requires. NIH, NSF, Wellcome, MacArthur, Ford, Open Philanthropy, and the Gates Foundation all have a place for this disclosure on the application portal as of 2025. Detection of undisclosed use is a far bigger problem than disclosed-and-cleaned-up use. The AI rewriter is the polish step you run before the disclosure statement, not a substitute for it.
Generating substantive research content with ChatGPT, attaching your name to it, and submitting to NIH or Wellcome. That breaches the human-authorship attestation every major funder now requires, regardless of which AI rewriter you run the output through. We will not pretend otherwise. If that is the situation you are in, we would rather you used the detector to understand which paragraphs read AI and then rewrote them with the work you actually did.
If you are advising on whether TextSight is appropriate for your development office, the framing is: same scope as a grammar checker or a professional editing service. Legitimate as a self-check on disclosed-use language polish, not legitimate as a way to disguise generated substantive content. Business tier with 5 seats and shared scan history lets your team see the same Authenticity Score and flagged sentences across multi-author drafts, so the threshold is a team agreement rather than a per-author judgment.
The detector workflow tuned for NIH, NSF, Wellcome, Ford, and MacArthur narratives.
Open the detector →The pre-journal-submission version of this workflow, calibrated for manuscripts.
Open papers 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 funder submission.
Read the guide →Free to try, no card. Section-by-section workflow, three modes, numbers and named partners preserved, IRB-and-grant-office pre-flight ready.