AI detectors for grant and proposal reviewers

AIDI-U-10

Updated ·

An AI detector for a grant reviewer is judged first on whether the reviewer is allowed to use one at all, and second on whether the tool takes a 15-page proposal and says what it does with it. Two fields decide which cards appear in this view: a text card whose input_limits or plans state a per-check limit of 5,000 words or more or a document upload, and a retention field that states what happens to the submitted text, because the first question a funder asks is where the proposal went. Of the 72 cards, 18 state both on the pages read on 8 and 9 November 2026; QuillBot takes a document on Premium and states no retention terms on the pages read, so it is absent, and Writer's detector was withdrawn on 22 December 2025. Two federal funders have answered the first question in writing: the NIH notice NOT-OD-23-149 of 23 June 2023 prohibits NIH peer reviewers from using generative AI "for analyzing and formulating peer review critiques" and says that uploading content from an application to online generative AI tools "violates the NIH peer review confidentiality and integrity requirements" (grants.nih.gov, read 9 November 2026), and the NSF notice of 14 December 2023 states that "NSF reviewers are prohibited from uploading any content from proposals, review information and related records to non-approved generative AI tools" (nsf.gov, read 9 November 2026). Neither notice names AI detectors, and this page does not read them as covering detectors; it reads them as the confidentiality rule a reviewer is already under. This page is one of the views under AI detectors by use case; the researchers view reads the same document limits for an author checking a manuscript, the input limits ranking sorts every stated maximum, and the full index shows all 72 cards.

  1. Field 1Document route as stated
  2. Field 2Retention as stated
  3. Field 3Training on submissions as stated
  4. Field 4Source, read date

The document route comes first and is quoted in the vendor's unit, because a 6,000-word research strategy with a budget justification and biosketches is neither pasted in pieces nor checked as a web page. Retention is quoted as stated, because a proposal under review is confidential by the funder's rule and several vendor policies allow training on submissions unless the user opts out. The training column quotes the vendor's statement on whether submitted text trains its models, the clause the NIH notice's "no guarantee of where data are being sent, saved, viewed, or used in the future" is about. Field rules are in the methodology.

Which fields decide fit for a proposal review?

AI detectors for grant and proposal reviewers: the cards in this view and the fields that decide fit
DetectorDocument route as statedRetention as statedTraining on submissions as statedSource, read date
Winston AIUp to 200,000 characters per scan; .docx, .png and .jpg uploadText and images stored; deleted from active storage when the report or account is deletedNot used to train AI models without explicit consentgowinston.ai/privacy-policy, 8 November 2026
PangramPDF, DOCX, RTF; up to 100 files at a timeDashboard text stored while the account is active, deletable from the history tab; zero retention option for the enterprise APIDetection dataset does not include user submitted contentpangram.com/data-privacy, 8 November 2026
GPTZeroUp to 150,000 characters or a batch of 250 files with a free accountAPI inputs not stored; dashboard inputs stored and used in aggregateDashboard inputs used in aggregate to improve the servicegptzero.me/privacy-policy.html, 8 November 2026
CopyleaksUp to 2,000 pages per document; OCR on image filesAPI content not stored; platform data may be used to improve servicesModel training unless a direct customer opts out by contacting supportcopyleaks.com/privacy-policy, 8 November 2026
PhraslyNo character limit on the web detector; file uploadText deleted immediately after analysis (detector page); terms grant a perpetual licence over Contributions"we don't use it to train our models" (detector page)phrasly.ai/terms/privacy, 8 November 2026
SaplingUp to 100,000 characters per check on Pro and Enterprise; PDF or Word uploadContent stored and used to improve the Service; API results cached 3 days; zero_retention flag"used only to improve the Service"sapling.ai/privacy, 8 November 2026
JustDoneUp to 15,000 words in one scan; PDF, Word or TXTStored for as long as reasonably necessary for the purposes of the policyNot publishedjustdone.com/legal/privacy-policy, 8 November 2026
Originality.aidocx, pdf, doc or URL; no maximum publishedScan history retained 30 days (Pro) or 1 year (Enterprise)Scan history may train the models unless the user opts outoriginality.ai/privacy-policy, 8 November 2026
ZeroGPT15,000 to 350,000 characters by plan; .txt, .docx, text-based .pdfRetained only as long as needed to operate the service; no-retention modes for API and enterprise; hosted in GermanyNot used to train, retrain or fine-tune AI modelszerogpt.com/privacy-policy, 8 November 2026
Proofademic3,500 to 25,000 words per request by plan; .docx, .txt, PDF; Batch ScanRetained as long as necessary; no period for submitted text published; processed in the U.S.Not used for model training or resaleproofademic.ai/privacy-policy, 8 November 2026
Turnitin300 to 30,000 words of prose; .docx, .pdf, .txt, .rtf; institutional licenceAI results retained in the similarity report; institutions choose whether papers enter the repository; full deletion on institutional requestNot stated on the FAQ readguides.turnitin.com, 8 November 2026
AIDetectPlus.docx, .pdf, .txt, .html; separate detector for papers; no maximum publishedNo period stated; data may be stored outside India; deletion on request through the contact formNot publishedaidetectplus.com/privacy-policy, 9 November 2026
Undetectable AIUp to 10,000 words per web check; under 30,000 per API request; file upload"Your text is never stored or used to train models" (detector FAQ); hosted in the United StatesNot used to train modelsundetectable.ai/privacy, 9 November 2026
Smodin50,000 characters per text or document; .txt, .doc, .docx, .pdf"Your text is never stored or shared" (detector page); terms grant a perpetual licence over User ContentNot stated beyond the licence clausesmodin.io/ai-content-detector, 9 November 2026
CompilatioDocuments uploaded to Magister+ or Studium; 1 credit per 250 wordsAccount data including documents stored 5 years from the end of the contract, on servers in FranceNot used for commercial purposes or reused outside the agreementsupport.compilatio.net, 9 November 2026
QuetextBulk uploads of 20 or 100 files in .doc, .docx, .pdf, .txt on paid plansDetector FAQ: not stored; policy section 6.1: User Content may be included in AI training with an email opt-outTraining under section 6.1 unless the user opts outquetext.com/privacy-policy, 9 November 2026
TruthScanUnder 30,000 words per API request; Word, PDF and Google Docs processing for enterprise usersRetained 30 days by default; Zero Data Retention on Business and aboveNot used for training under ZDR; otherwise retention may continue for service improvementtruthscan.com/privacy, 9 November 2026
ScribbrPremium check bands up to 50,000+ words"we do not store or share your data"Submissions never added to the content databasescribbr.com/ai-detector, 8 November 2026

Values as each vendor states them, read 8 and 9 November 2026. QuillBot (document upload on Premium, no retention terms on the pages read), Walter Writes (750 to 2,000 words per request, no upload stated), Hive (text in 2,048-character chunks, no upload) and the three open-source tools (pasted text, no vendor policy) state one field or neither and are absent. Writer, which took up to 5,000 words per check, was withdrawn on 22 December 2025 (archived writer.com page, read 9 November 2026).

The self-publishing view, added 10 November 2026, applies the same two tests, a full document and a stated retention policy, to a book manuscript and adds the published price per 1,000 words and the KDP disclosure rule read live that day.

AI detectors for grant and proposal reviewers: register plate with the page title and its register ID
The view plate: register ID and title, drawn from the index data.

Why each card is in this view

  1. AIDI-D-0005Winston AI

    200,000 characters per scan with .docx upload and a policy that stores text until the report is deleted (gowinston.ai/privacy-policy, read 8 November 2026); 0 false accusations in 18 human documents in Weber-Wulff 2023, 0.55% at threshold 0.5 in RAID 2024.

  2. AIDI-D-0006Pangram

    up to 100 files at a time and a zero-retention option for the enterprise API (pangram.com/data-privacy, read 8 November 2026); 0.0000 of 1,992 human passages at threshold 0.5 in Jabarian and Imas 2025, 0 of 40 human papers in Van Vlasselaer 2026, and 45.5% of 1,163 real theses flagged when run alone in the same study.

  3. AIDI-D-0001GPTZero

    a batch of 250 files and an API that stores nothing, while dashboard inputs are stored (gptzero.me/privacy-policy.html, read 8 November 2026); 0.0071 of 1,992 human passages in Jabarian and Imas 2025, 9 of 18 in Weber-Wulff 2023. The two routes carry different retention and a reviewer would need to know which one a tool is on.

  4. AIDI-D-0003Copyleaks

    documents to 2,000 pages, an API that stores nothing and a platform policy that trains unless a direct customer opts out (copyleaks.com/privacy-policy, read 8 November 2026); 0 of 42 human papers in Walters 2023.

  5. AIDI-D-0012Phrasly

    no character limit and deletion on analysis on the detector page, beside terms that grant a perpetual licence over Contributions (phrasly.ai/terms/privacy, read 8 November 2026); no study record.

  6. AIDI-D-0010Sapling

    100,000 characters per check on Pro with PDF or Word upload, a 3-day API cache and a zero_retention flag (sapling.ai/privacy, read 8 November 2026); 0 of 42 in Walters 2023 with 31% uncertain, 1 of the 7 detectors in Liang 2023.

  7. AIDI-D-0013JustDone

    15,000 words per scan with PDF, Word or TXT upload, and a policy that stores data as long as reasonably necessary (justdone.com/legal/privacy-policy, read 8 November 2026); the vendor claims "98% accuracy on academic and scientific texts" with no study record (JustDone legit question).

  8. AIDI-D-0004Originality.ai

    docx, pdf and doc upload, scan history for 30 days or a year, and training on scan history unless the user opts out (originality.ai/privacy-policy, read 8 November 2026); 0.0011 of 1,992 in Jabarian and Imas 2025, 0 of 42 in Walters 2023.

  9. AIDI-D-0011ZeroGPT

    15,000 to 350,000 characters by plan, no-retention modes for API and enterprise, and hosting in Germany (zerogpt.com/privacy-policy, read 8 November 2026); 0 of 18 in Weber-Wulff 2023, 1.42% at threshold 0.5 in RAID 2024.

  10. AIDI-D-0014Proofademic

    3,500 to 25,000 words per request with Batch Scan and a statement that submitted content is not used for training or resale (proofademic.ai/privacy-policy, read 8 November 2026); no study record.

  11. AIDI-D-0002Turnitin

    300 to 30,000 words through an institutional licence, with AI results kept in the similarity report and deletion of whole submissions on institutional request (guides.turnitin.com, read 8 November 2026); 0 of 42 in Walters 2023, 0 of 18 in Weber-Wulff 2023, 0 of 40 in Van Vlasselaer 2026. A reviewer without an institutional account cannot run it.

  12. AIDI-D-0018AIDetectPlus

    a separate detector for papers and deletion on request through a contact form, with no period stated (aidetectplus.com/privacy-policy, read 9 November 2026); no study record.

  13. AIDI-D-0024Undetectable AI

    10,000 words per web check with file upload and a detector FAQ stating that text is never stored or used to train models (undetectable.ai/privacy, read 9 November 2026); 1 study record, Kar 2024, on a single ChatGPT article.

  14. AIDI-D-0029Smodin

    50,000 characters per document with four upload formats and a never-stored statement beside a perpetual licence in the terms (smodin.io/ai-content-detector, read 9 November 2026); no study record.

  15. AIDI-D-0032Compilatio

    documents through Magister+ or Studium, stored 5 years after the contract ends on servers in France (support.compilatio.net, read 9 November 2026); 0 of 18 in Weber-Wulff 2023.

  16. AIDI-D-0033Quetext

    bulk uploads of 20 or 100 files and a detector FAQ saying scans are not stored, against a policy section that allows AI training with an email opt-out (quetext.com/privacy-policy, read 9 November 2026); no study record.

  17. AIDI-D-0034TruthScan

    under 30,000 words per API request, 30-day default retention and Zero Data Retention on Business and above (truthscan.com/privacy, read 9 November 2026); no study record.

  18. AIDI-D-0009Scribbr

    premium checks priced by document band up to 50,000+ words and a no-storage statement (scribbr.com/ai-detector, read 8 November 2026); 2% of 42 human papers in Walters 2023.

What the evidence says for proposal review

The funders' rules come before any score. NIH's notice of 23 June 2023 tells peer reviewers that "AI tools have no guarantee of where data are being sent, saved, viewed, or used in the future" and revises the reviewer confidentiality agreement to prohibit uploading or sharing content from an application to online generative AI tools (grants.nih.gov NOT-OD-23-149, read 9 November 2026). NSF's notice of 14 December 2023 says that "any information uploaded into generative AI tools not behind NSF's firewall is considered to be entering the public domain" and that proposers "are encouraged to indicate in the project description the extent to which, if any, generative AI technology was used" (nsf.gov merit review policy, read 9 November 2026). An AI detector is a third-party service that takes the proposal text as input, so a reviewer bound by either notice should treat the retention column above as the first field, and the funder's programme officer, not a score, as the authority on whether a check is permitted at all.

One record in the index scored scientific text. Gao 2023 asked ChatGPT to write 50 abstracts in the style of five medical journals from real titles and ran them and the 50 originals through the GPT-2 Output Detector: an AUROC of 0.94, a median "fake" score of 99.98% on generated abstracts and 0.02% on originals, and 86% sensitivity with 94% specificity at the paper's cutoff (DOI 10.1038/s41746-023-00819-6, read 9 November 2026; Gao 2023 record). Blinded human reviewers in the same study caught 68% of the generated abstracts and called 14% of the real ones generated. The GPT-2 Output Detector is a 2019 model run locally with no vendor policy, which is why it is absent from the table; on 1,992 human passages in Jabarian and Imas 2025 the same model flagged 0.9503 as AI at threshold 0.5, while Pangram flagged 0.0000, Originality.ai 0.0011 and GPTZero 0.0071 (NBER w34223, Table 3, read 9 November 2026; Jabarian and Imas 2025 record). That study's false negative rates at a 0.0001 false positive cap reached 1.0000 for GPTZero on all four generators, so the threshold a reviewer would need to avoid a wrong flag on a careful applicant is one at which some tools catch nothing.

The applicant profile a reviewer reads is the one with the worst false positive record. Liang 2023 reported an average false positive rate of 61.22% across seven detectors on 91 essays by non-native English writers (arXiv 2304.02819, read 8 November 2026; Liang 2023 record), and Van Vlasselaer 2026 ran Pangram alone on 1,163 real theses and saw it flag 45.5% (DOI 10.1007/s40979-026-00226-w, Table 6, read 9 November 2026). The non-English evidence ranking counts every record on translated or non-native text. No detector score should be the only evidence in a funding, misconduct or publishing decision, and a reviewer who is barred from uploading a proposal has no score to weigh.

Sources

  • https://grants.nih.gov/grants/guide/notice-files/NOT-OD-23-149.html (read 9 November 2026)
  • https://www.nsf.gov/policies/ai/merit-review (read 9 November 2026)
  • https://gowinston.ai/privacy-policy/ (read 8 November 2026)
  • https://www.pangram.com/data-privacy (read 8 November 2026)
  • https://gptzero.me/privacy-policy.html (read 8 November 2026)
  • https://copyleaks.com/privacy-policy (read 8 November 2026)
  • https://phrasly.ai/terms/privacy (read 8 November 2026)
  • https://sapling.ai/privacy (read 8 November 2026)
  • https://justdone.com/legal/privacy-policy (read 8 November 2026)
  • https://originality.ai/privacy-policy (read 8 November 2026)
  • https://www.zerogpt.com/privacy-policy (read 8 November 2026)
  • https://proofademic.ai/privacy-policy/?utm_source=aidetectorindex.com&utm_medium=owned&utm_campaign=for_grant_reviewers (read 8 November 2026)
  • https://guides.turnitin.com/hc/en-us/articles/28477544839821-Turnitin-s-AI-writing-detection-capabilities-FAQs (read 8 November 2026)
  • https://aidetectplus.com/privacy-policy (read 9 November 2026)
  • https://undetectable.ai/privacy (read 9 November 2026)
  • https://smodin.io/ai-content-detector (read 9 November 2026)
  • https://support.compilatio.net/hc/en-us/articles/4416953631505 (read 9 November 2026)
  • https://www.quetext.com/privacy-policy (read 9 November 2026)
  • https://truthscan.com/privacy (read 9 November 2026)
  • https://www.scribbr.com/ai-detector/ (read 8 November 2026)
  • http://web.archive.org/web/20251209115654/https://writer.com/ai-content-detector/ (read 9 November 2026)
  • https://doi.org/10.1038/s41746-023-00819-6 (read 9 November 2026)
  • https://www.nber.org/papers/w34223 (read 9 November 2026)
  • https://arxiv.org/pdf/2304.02819v3 (read 8 November 2026)
  • https://doi.org/10.1007/s40979-026-00226-w (read 9 November 2026)

Revision log: 2026-11-09: view created at index version 2026.11.92026-11-10: self-publishing view linked

FAQ: AI detectors for grant reviewers

May an NIH or NSF reviewer run a proposal through an AI detector?

The notices do not name detectors. NIH's NOT-OD-23-149 (23 June 2023) prohibits uploading or sharing content from an application to online generative AI tools, and NSF's notice (14 December 2023) prohibits uploading proposal content to non-approved generative AI tools and treats anything uploaded outside its firewall as public (grants.nih.gov, nsf.gov, read 9 November 2026). The programme officer decides.

Which AI detectors take a full proposal and state a retention policy?

Eighteen of 72 cards on 8 and 9 November 2026, from Winston AI's 200,000 characters per scan to Scribbr's 50,000-word band; Turnitin and Compilatio need an institutional account. Four state a zero-retention option (Pangram, Sapling, ZeroGPT, TruthScan) and five state deletion on analysis or no storage (Phrasly, Undetectable AI, Smodin, Scribbr, and Copyleaks for its API).

Has any study tested AI detectors on research writing?

One in the index: Gao 2023, where the GPT-2 Output Detector reached an AUROC of 0.94 on 50 ChatGPT-written and 50 real medical abstracts (DOI 10.1038/s41746-023-00819-6, read 9 November 2026). No indexed study scored grant proposals or full manuscripts.

How low can a false positive rate go on human text?

To 0.0000 of 1,992 passages for Pangram, 0.0011 for Originality.ai and 0.0071 for GPTZero at threshold 0.5 in Jabarian and Imas 2025, against 0.9503 for the 2019 RoBERTa detector (NBER w34223, read 9 November 2026). At a 0.0001 false positive cap, GPTZero's false negative rate was 1.0000 on all four generators in the same study.

Which vendors train on submitted text?

Copyleaks (unless a direct customer opts out), Originality.ai (scan history, unless the user opts out), Sapling (content "used only to improve the Service") and Quetext (policy section 6.1, with an email opt-out) state that submissions may train or improve their models, all as read on 8 and 9 November 2026. Pangram, ZeroGPT, Proofademic, Undetectable AI and Winston AI state that they do not.

Does a flagged proposal prove anything?

No. The vendors say a score is a signal, not a verdict; Turnitin's FAQ says its indicator "should not be used as the sole basis for action". Liang 2023's 61.22% average false positive rate on non-native English writers (arXiv 2304.02819, read 8 November 2026) describes the applicant pool a funder reviews.

Related reading