AI detectors for self-publishing authors: the cards that take a manuscript and say what happens to it
An AI detector fits a self-publishing author or a manuscript editor when the vendor states a document upload or a per-check limit of 5,000 words or more, and states what happens to the text afterwards, and 12 of the 72 text cards in the index meet both tests by the input limits, file types and retention fields read 8 to 10 November 2026: Sapling, GPTKit, Undetectable AI, AI or Not, Fast-DetectGPT, AI Detector Pro, Paperpal, Phrasly, Smodin, StealthWriter, Trinka and TruthScan (sapling.ai/ai-content-detector, gptkit.ai, undetectable.ai, docs.aiornot.com, fastdetect.net, aidetector.pro, paperpal.com, phrasly.ai, smodin.io, stealthwriter.ai, trinka.ai, truthscan.com). The reason the retention test comes second and not last is Amazon's own rule: Kindle Direct Publishing requires authors to inform it of AI-generated text, images or translations when publishing or republishing a book, and does not require disclosure of AI-assisted content, by the KDP Content Guidelines read live on 10 November 2026 (kdp.amazon.com, Content Guidelines), so an author who runs a 90,000-word manuscript through a detector before upload is checking their own disclosure, and the manuscript is the author's most valuable file. This page is one of the views under AI detectors by use case; the researchers view applies the limit rule without the retention test, the data retention ranking sorts every stated period, and the full index shows all 72 cards.
- Field 1Per-check limit as stated
- Field 2Upload formats as stated
- Field 3Retention as stated
- Field 4Price per 1,000 words where published
- Field 5Study records
- Field 6Source, read date
Three fields, read in order. The input limits field must state a document upload or a ceiling of 5,000 words or more per check (27,500 characters at the index's rate of 5.5 characters a word), because a novel is 60,000 to 120,000 words and a detector that takes 1,200 words at a time turns one check into eighty. The retention field must state a period or state that submissions are not stored, in the vendor's own words, because the text being checked is an unpublished book; a card whose policy says "as long as necessary" and nothing more does not qualify, however large its limit, so Winston AI (200,000 characters, no stated period), GPTZero (150,000 characters), Turnitin (30,000 words, kept inside the similarity report with no period), Detecting-AI.com (1,000,000 characters) and Decopy (200,000 characters) sit outside this view and on the researchers view. The third column quotes the upload formats as the vendor spells them, because a manuscript arrives as .docx or .pdf, and the price column quotes the published price per 1,000 words where one exists, because a 90,000-word check is a priced event; where a vendor sells credits or plans, the cell quotes the plan. Hive and SynthID state upload formats and a retention period but take media, not manuscripts (Hive's formats are image, video and audio; SynthID has no public text checker), and are left out by the modality. Field rules are in the methodology.
Which fields decide fit for a manuscript check?
| Detector | Per-check limit as stated | Upload formats as stated | Retention as stated | Price per 1,000 words where published | Study records | Source, read date |
|---|---|---|---|---|---|---|
 Sapling | Up to 100,000 characters per check on Pro and Enterprise (2,000 free); API up to 200,000 characters per request | .pdf, .docx, .pptx, .odt, .txt, .md, .html or .htm up to 30 MB; 5 files per batch free, 10 on Pro | "AI detection results are cached on the entire text for 3 days"; a zero_retention flag scores a text without a copy being kept | $0.0275, converted from $0.005 per 1,000 characters | 3 | sapling.ai/ai-content-detector, sapling.ai/privacy, 8 November 2026 |
 GPTKit | 153,600 characters per request on the Institution plan (30,720 Team, 10,240 Personal, 2048 as a guest) | html, txt, doc, docx, pdf, jpeg, png, gif, odt, rtf and more in the upload control | "Once the detection process is complete, the content is immediately removed from our servers" | Not published; plans from $5 a month | 2 | gptkit.ai, 9 November 2026 |
 Undetectable AI | Up to 10,000 words per check on the web tool, no sign-up; API under 30,000 words per request | "paste or upload a file", no format named | "Your text is never stored or used to train models" | $0.04, converted from API Starter at $29 a month for 72,500 word credits | 1 | undetectable.ai/ai-detector, help.undetectable.ai, 9 November 2026 |
 AI or Not | Text from 250 to 500,000 characters per API request | DOC, DOCX or TXT on the web tool | "uploaded media are processed and deleted immediately after inference" | $0.005, from $5 per million words | 1 | docs.aiornot.com, aiornot.com/pricing, 9 November 2026 |
| Fast-DetectGPT | No maximum stated; the hosted demo's terms say length limits apply with no figure; the code runs locally | .txt, .pdf or .docx up to 10 MB on the hosted demo | "Detection text is only temporarily stored during processing and deleted immediately after completion" (the hosted demo's terms); a local run sends nothing | Not published; free open-source code | 1 | fastdetect.net, github.com/baoguangsheng/fast-detect-gpt, 10 November 2026 |
 AI Detector Pro | 40,000 characters per paid scan, "about 6000-7000 words" | No upload format list published; the API takes 4,000-character units | "automatically deleted at the beginning of the 4th month after it was entered" (90 days) | $0.124, converted from $0.09 per 4,000-character unit, Unlimited plan required | 0 | aidetector.pro/pages/faq, aidetector.pro/pages/privacy, 10 November 2026 |
 Paperpal | 1,200 words per free scan; Prime 25,000 words of detection a month and Pro unlimited, with no per-scan ceiling stated | .docx up to 25 MB | "We don't store your drafts or use them to train our AI models" | Not published; Prime $25 a month | 0 | paperpal.com/tools/ai-detector, 10 November 2026 |
 Phrasly | "no character limits on detection" on the web tool; Business API 15,000 characters per request | "paste text or upload file", no format named | "deleted immediately after analysis. We don't store your writing" | $0.02 on the Business API plan ($100 a month including $100 in credits) | 0 | phrasly.ai/ai-detector, phrasly.ai/terms/privacy, 8 November 2026 |
 Smodin | 50,000 characters per text or document, about 9,100 words | .txt, .doc, .docx or .pdf | "Your text is never stored or shared" | Not published; Starter $12.99 a month | 0 | smodin.io/ai-content-detector, 9 November 2026 |
 StealthWriter | 5,000 words per input on every paid plan (1,000 free, 500 on the demo) | Paste only; no upload | "Content submitted for AI processing (humanization or detection) is not stored at all" | Not published; Starter $20 a month | 0 | stealthwriter.ai/pricing, stealthwriter.ai/privacy, 10 November 2026 |
 Trinka | No per-check ceiling stated; 1 credit per 5,000 words; the marketplace takes 25,000 words per file | "Upload your MS Word or PDF file" | "Premium user's processed data is automatically deleted from the server every 90 days"; real-time deletion on the Confidential Data plan | Not published; $10 for 4 credits, 1 credit per 5,000 words | 0 | trinka.ai/features/ai-content-detector, trinka.ai/trust-center, 10 November 2026 |
 TruthScan | Under 30,000 words per API request; the web box takes pasted text | Word, PDF and Google Docs document processing for enterprise users; jpg, png, jpeg, tif, webp for images | "Retained for 30 days by default, then deleted, unless retention is required for service improvement or otherwise agreed" | Not published per word; Starter $24 a month for 1,000 results | 0 | truthscan.com, text API documentation, truthscan.com/privacy, 9 November 2026 |
Values as each vendor states them, ordered by study records then name, read on the dates shown; a limit in characters is converted at 5.5 characters a word only where the cell says "about". The study records column counts every record on the card, text or otherwise, and only Sapling's count carries a measured false positive rate on human text (0 of 42 papers, Walters 2023).
Why each card is in this view
- AIDI-D-0010Sapling
a 100,000-character paid check, eight document formats and the index's only stated zero-retention switch, a flag an author sets per request (sapling.ai/privacy, read 8 November 2026); 0 of 42 human papers flagged in Walters 2023 with 31% returned uncertain in the same test, and 33% of 42 GPT-4 papers correct (Walters 2023); Liang 2023 tested it on TOEFL essays and prints per-tool rates in a chart only (Liang 2023).
- AIDI-D-0045GPTKit
153,600 characters a request on the Institution plan, ten upload formats and deletion on completion (gptkit.ai, read 9 November 2026); 0 of 114 human submissions flagged in Orenstrakh 2023 and 4.5% of manipulated AI text caught in Perkins 2024, so a manuscript that passed through a humanizer will mostly pass GPTKit too (Orenstrakh 2023; Perkins 2024).
- AIDI-D-0024Undetectable AI
10,000 words a check with no sign-up and a no-storage line on the detector page (undetectable.ai/ai-detector, read 9 November 2026); the one record, Kar 2024, scored a single article, so no rate on human text exists for the card.
- AIDI-D-0025AI or Not
500,000 characters per API request, DOC, DOCX or TXT on the web tool, deletion after inference and the lowest published text price in the index, $5 per million words (docs.aiornot.com, aiornot.com/pricing, read 9 November 2026); its one record is on images, not text (Ha and Passananti 2024).
- AIDI-D-0049Fast-DetectGPT
the one card an author can run on their own machine from MIT-licensed code, where no text leaves the computer; the hosted demo takes .txt, .pdf or .docx and states deletion on completion (github.com/baoguangsheng/fast-detect-gpt, fastdetect.net, read 10 November 2026); RAID 2024 printed 23.2% of human documents flagged at the naive 0.5 threshold (RAID 2024).
- AIDI-D-0048AI Detector Pro
40,000 characters a paid scan, a chapter at a time, with a dated 90-day deletion clause (aidetector.pro/pages/privacy, read 10 November 2026); no study record, and the Terms state two months where the policy states the fourth month.
- AIDI-D-0050Paperpal
a .docx upload up to 25 MB and a no-storage line on the detector page, from a vendor built for academic manuscripts (paperpal.com/tools/ai-detector, read 10 November 2026); the free scan stops at 1,200 words and paid plans meter words a month rather than per scan; no study record.
- AIDI-D-0012Phrasly
no character limit on the web detector and deletion immediately after analysis, with the lowest published API price among the text-only cards, $0.02 per 1,000 words (phrasly.ai/ai-detector, business.phrasly.ai, read 8 November 2026); the same company sells a humanizer, and no study tested the detector.
- AIDI-D-0029Smodin
50,000 characters a check with .txt, .doc, .docx or .pdf upload and "never stored or shared" on the detector page (smodin.io/ai-content-detector, read 9 November 2026); no study record, and the vendor's 99.8% claim carries no base.
- AIDI-D-0056StealthWriter
5,000 words per input on paid plans, paste only, and a privacy policy that says submitted text "is not stored at all" (stealthwriter.ai/privacy, read 10 November 2026); the detector is the second tab of a humanizer and no study tested it.
- AIDI-D-0055Trinka
a Word or PDF upload with a PDF report back, 1 credit per 5,000 words, a 90-day deletion default and a $500-a-year plan with real-time deletion (trinka.ai/trust-center, read 10 November 2026); the privacy policy reserves the right to save documents, and no study record exists.
- AIDI-D-0034TruthScan
under 30,000 words per API request, document processing for enterprise users and a 30-day default retention with a Zero Data Retention enterprise tier (truthscan.com/privacy, read 9 November 2026); no study record, and the 99%+ claim carries no base.
What the evidence says
What the KDP rule asks, and what a detector can answer
The KDP Content Guidelines, read live on 10 November 2026, require an author "to inform us of AI-generated content (text, images, or translations) when you publish a new book or make edits to and republish an existing book through KDP", define AI-generated as content "created by an AI-based tool" even "if you applied substantial edits afterwards", define AI-assisted as content the author created and then used AI tools "to edit, refine, error-check, or otherwise improve", and state that disclosure of AI-assisted content is not required (kdp.amazon.com, Content Guidelines). The distinction is about how the text was made, which an author already knows, and not about how it reads, which is all a detector measures; the same page says Amazon enforces its guidelines "using a combination of machine learning, automation, and dedicated teams of human reviewers" and names no detector. A detector score therefore answers a different question from the one KDP asks, and the index records no page on which Amazon says what score, tool or threshold it uses.
What the evidence adds is a sense of how wrong a score can be in each direction. On the human side, Liang 2023 ran 91 TOEFL essays by non-native English writers through seven detectors and found an average false positive rate of 61.22%, against near-perfect accuracy on 88 essays by US eighth-graders (DOI 10.1016/j.patter.2023.100779, read 8 November 2026; Liang 2023), and Ibrahim 2023 found 16 of 120 (13.3%) ESL student essays on the GPT-2 Output Detector and 18 of 120 (15%) on Crossplag scoring below the 95% originality cut the author used to mark a text as human-written (DOI 10.1186/s40468-023-00260-2, read 9 November 2026; Ibrahim 2023); an author writing in a second language is the reader those two records speak to. On the other side sits the human baseline: Fleckenstein 2024 found that 89 pre-service teachers identified 45.1% of ChatGPT essays and 200 experienced teachers 37.8%, so an editor reading a manuscript unaided catches fewer than half of fully generated texts (DOI 10.1016/j.caeai.2024.100209, read 9 November 2026; Fleckenstein 2024). No detector score should be the only evidence in a misconduct, hiring or publishing decision, and a disclosure to KDP is the author's statement about process, not a detector's statement about style.
Sources
- https://kdp.amazon.com/en_US/help/topic/G200672390 (read 10 November 2026)
- https://sapling.ai/ai-content-detector (read 8 November 2026)
- https://sapling.ai/privacy (read 8 November 2026)
- https://sapling.ai/pricing (read 8 November 2026)
- https://gptkit.ai/ (read 9 November 2026)
- https://undetectable.ai/ai-detector (read 9 November 2026)
- https://help.undetectable.ai/en/article/detector-api-1cf74il/ (read 9 November 2026)
- https://undetectable.ai/pricing (read 9 November 2026)
- https://docs.aiornot.com/ (read 9 November 2026)
- https://www.aiornot.com/pricing (read 9 November 2026)
- https://fastdetect.net/ (read 10 November 2026)
- https://github.com/baoguangsheng/fast-detect-gpt (read 10 November 2026)
- https://aidetector.pro/pages/faq (read 10 November 2026)
- https://aidetector.pro/pages/privacy (read 10 November 2026)
- https://paperpal.com/tools/ai-detector (read 10 November 2026)
- https://phrasly.ai/ai-detector (read 8 November 2026)
- https://phrasly.ai/terms/privacy (read 8 November 2026)
- https://business.phrasly.ai/docs/ai-detection-api (read 8 November 2026)
- https://smodin.io/ai-content-detector (read 9 November 2026)
- https://stealthwriter.ai/pricing (read 10 November 2026)
- https://stealthwriter.ai/privacy (read 10 November 2026)
- https://www.trinka.ai/features/ai-content-detector (read 10 November 2026)
- https://www.trinka.ai/trust-center (read 10 November 2026)
- https://truthscan.com/truthscan-ai-text-detection-api-documentation (read 9 November 2026)
- https://truthscan.com/privacy (read 9 November 2026)
- https://doi.org/10.1016/j.patter.2023.100779 (read 8 November 2026)
- https://doi.org/10.1186/s40468-023-00260-2 (read 9 November 2026)
- https://doi.org/10.1016/j.caeai.2024.100209 (read 9 November 2026)
Revision log: 2026-11-10: view created at index version 2026.11.11; the KDP Content Guidelines fetched live the same day
FAQ: AI detectors for self-publishing authors
Which AI detectors take a whole manuscript in one check?
Of the twelve cards in this view, AI or Not states the largest text ceiling, 500,000 characters per API request, about 91,000 words; GPTKit 153,600 characters on its Institution plan and Sapling 100,000 characters per paid check follow, and Phrasly states no character limit on its web detector (docs.aiornot.com, gptkit.ai, sapling.ai, phrasly.ai, read 8 and 9 November 2026). The input limits ranking sorts every stated ceiling, including the cards this view leaves out for want of a retention statement.
Does Amazon KDP run an AI detector on uploads?
The KDP Content Guidelines, read 10 November 2026, name no detector; they require authors to inform KDP of AI-generated text, images or translations, exempt AI-assisted content from disclosure, and say enforcement uses "machine learning, automation, and dedicated teams of human reviewers" (kdp.amazon.com). The index records no Amazon page stating a tool, score or threshold.
Which of these detectors state that the manuscript is not stored?
Eight state deletion on completion or no storage: GPTKit, Undetectable AI, AI or Not, Fast-DetectGPT (the hosted demo), Paperpal, Phrasly, Smodin and StealthWriter, each quoted in the table from the page read 8 to 10 November 2026; Sapling caches results for 3 days with a per-request zero-retention flag, TruthScan keeps content 30 days by default, and AI Detector Pro and Trinka state 90 days.
Which has the lowest published price for a 90,000-word manuscript?
AI or Not, at $5 per million words, prices a 90,000-word check at $0.45 through its API; Phrasly's $0.02 per 1,000 words gives $1.80 on the Business API plan, Sapling's $0.005 per 1,000 characters about $2.48, and Undetectable AI's $0.04 per 1,000 words $3.60 (aiornot.com/pricing, business.phrasly.ai, sapling.ai/pricing, undetectable.ai/pricing, read 8 and 9 November 2026); the API pricing ranking converts each at 5.5 characters a word.
What false positive rate should an author writing in a second language expect?
No single figure: Liang 2023 found an average of 61.22% on 91 TOEFL essays across seven detectors in 2023, Ibrahim 2023 found 13.3% and 15% of 120 ESL essays below a 95% originality cut on two detectors, and Turnitin's own research reports 0.014 for ELL writers (Liang 2023; Ibrahim 2023; Turnitin claims record). Only Sapling in this view carries a measured rate on human text, 0 of 42 papers in Walters 2023.
Can an editor spot AI text without a detector?
Not reliably, by the one human-rater record in the index: 89 pre-service teachers identified 45.1% of ChatGPT essays and 200 experienced teachers 37.8% in Fleckenstein 2024 (DOI 10.1016/j.caeai.2024.100209, read 9 November 2026).
