How Accurate Is Sapling's AI Detector?

Sapling's AI detector is accurate enough for a quick first look and not accurate enough to settle a question about who wrote a text. Sapling's own detector page says that "accuracy varies with the dataset, text length, and decision threshold," and that no detector "should be used as a standalone check." Independent evidence is thin. The best-documented test of Sapling found that it flagged 18 of 30 human-written physics essays as AI. Sapling upgraded its model in September 2026, so older results describe an earlier version.

If Sapling flagged your own writing, the score describes patterns in the text, and it isn't evidence of how you wrote it.

How accurate is Sapling's AI detector?

Sapling's own pages give no single accuracy percentage. The detector page says accuracy depends on the dataset, the text length and the threshold. It advises you to test the detector on your own examples and count both false positives and missed AI text. It also says the detector "becomes much more accurate after 50 or so words."

Sapling does describe its validation work. A technical whitepaper, "Sapling AI Content Detection: Technical Overview and Validation Report", is an 11-page document with an evaluation on a 1,000-sample held-out set, including false-positive and false-negative analysis. You request it through a form on the detector page, and the results aren't printed there.

Sapling's changelog gives the most recent accuracy statement. On 17 September 2026 it upgraded the default model, with higher precision on recent AI models and higher detection of humanized AI text. It reports that false positives fell by 60% on internal held-out evaluations.

Two published tests used Sapling's detector, and both predate that upgrade. The table shows what each one measured.

Both studies tested the web detector as it stood at the time. Sapling's current model may behave differently, and nobody has published an independent test of the September 2026 version.

What is Sapling and who makes it?

Sapling is a language-tools company, and the AI detector is one product in a larger toolkit. Sapling Intelligence, Inc. makes it, and its about page says the team's work comes out of the Berkeley and Stanford AI labs and that the company is backed by Y Combinator.

Sapling's homepage describes it as a "language model toolkit for developers and businesses." Its customers include support and CX teams, and its other tools are built for those workflows. The detector is one of the pieces that anyone can use on the website.

Sapling's other products

The same site lists several tools next to the detector:

• An API and SDK with endpoints and editor components for grammar checking, autocomplete and rephrasing.
• An MCP server that lets clients such as Claude call Sapling's language tools.
• Grammar and quality checking with scoring and error reports.
• Autocomplete for messaging platforms, and a shared library of text-expansion snippets.
• Rephrase and chat-assist tools, plus an AI punctuation check and a sentence rewriter on the website.
• A browser extension and integrations with Zendesk, ServiceNow, Salesforce, Twilio Flex, and editors such as TinyMCE and Quill.

The breadth is useful if you already work in a support or sales tool that Sapling supports.

What does Sapling's free AI detector allow?

The free detector checks the first 2,000 characters of whatever you paste, which is roughly 300 to 350 words of English. Anything after that is not scored. Paid plans raise the limit to 100,000 characters per check.

These are Sapling's detector limits and prices in October 2026.

New accounts can start a one-month Pro trial without a credit card. The API gives new keys 50,000 characters a day and 250,000 characters a month free, and it supports up to 200,000 characters per request.

The 2,000-character cap matters more than it looks. A thesis chapter has tens of thousands of characters, so the free score describes only the opening paragraphs. Sapling's evaluation guide says this itself: the free check scores the first 2,000 characters, so the result tells you nothing about the rest of a longer document.

The detector also takes PDF and Word files. It extracts the text first, and Sapling's guide suggests checking the extracted text, since extraction can change the reading order.

How does Sapling score text?

Sapling gives one overall percentage and a separate score for each sentence. The overall result is an AI-generation score, and the flagged portions are highlighted in the text. A higher score means the text is more likely to be AI-generated.

Sapling describes the method as a Transformer-based model that estimates, for each token, the probability that it is AI-generated. Sentence scores combine the probabilities of the tokens inside them. The API also returns the token-level probabilities, and it can return an HTML heatmap.

The API documentation adds some detail that helps when you read the web result:

• The overall score runs from 0 to 1, where 0 is the highest confidence in human writing and 1 is the highest confidence in AI.
• The overall score is computed over the whole text at once. It is not the average of the sentence scores.
• A field called ai_fraction gives the share of text in sentences scored 0.5 or above, so it shows how much of a mixed document looks AI-written.
• Sapling recommends at least 300 characters per request.

Sapling's guide also says the score is an estimate of how likely the checked text is to be AI-generated. It is not the share of words written by AI. The sentence highlights are scored separately, so they may disagree with the overall score, and Sapling asks you to review both.

Languages

Sapling checks English, Spanish, Portuguese, French, German, Italian, Dutch and Japanese. Spanish, Portuguese, French, German and Japanese were added on 9 October 2026, and Italian and Dutch were added in preliminary form on 13 October 2026. Text in any other language goes through the English model, and Sapling says those results are less reliable.

For researchers who write in a language outside that list, this means a Sapling score on a Thai or Arabic text isn't based on a model built for it.

What did the independent studies find?

The independent evidence on Sapling comes from two papers, and neither was designed to test Sapling alone.

Liang et al. (2023): seven detectors and second-language essays

Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu and James Zou at Stanford tested seven detectors. The study, "GPT detectors are biased against non-native English writers", was published in the journal Patterns. Sapling was one of them, along with Originality.AI, Quil.org, OpenAI's detector, Crossplag, GPTZero and ZeroGPT. The authors accessed all seven on 15 March 2023.

The detectors were close to perfect on the US eighth-grade essays. On the 91 TOEFL essays they flagged an average of 61.22% as AI-generated. All seven agreed that 18 of the 91 essays were AI, and 89 of the 91 were flagged by at least one detector.

The TOEFL essays that every detector flagged had significantly lower perplexity than the others. The authors read this as a sign that detectors penalize writers with limited linguistic range. The paper gives the average across the seven detectors as its headline number, so it doesn't isolate Sapling's own rate in the text. We cover the full study in why AI detectors flag non-native writers.

Yeadon et al. (2024): physics essays at Durham

Will Yeadon and colleagues at Durham University tested five detectors on short physics essays, half written by students before ChatGPT existed and half written by GPT-4. They published it as an arXiv preprint in March 2024. The tools were ZeroGPT, QuillBot, Hive Moderation, Sapling and Radar.

Their confusion matrix for Sapling lists 30 true positives, 18 false positives, 12 true negatives and no false negatives. That is 70% accuracy and 62% precision. Sapling caught every AI-written essay and also flagged 18 of the 30 human essays. The paper notes that Sapling had the highest number of false positives of the five tools.

Read the result as a statement about that test. The essays were short, formal and about one subject, which is the kind of text Sapling's own FAQ says is more likely to be flagged. The study also tested a version from before the 2026 upgrade.

What the two studies share

Both studies show Sapling catching AI text well and also flagging a lot of human text. In the Durham test it missed no AI essays and flagged 18 human ones. A detector with that profile suits a first look at a long text, where a high score tells you where to read closely. It suits a verdict much less.

Why Sapling flags some human writing

Sapling's FAQ answers this directly. If you paste your own writing and it says AI, false positives can happen, and shorter, more general and more essay-like text is more likely to be flagged. Sapling's evaluation guide adds that formulaic human writing can be flagged, and that edited AI writing can be missed.

Academic prose has all of those features. Methods sections, abstracts and literature summaries are formal, general and built on stock phrases. Writers working in a second language often use a narrower range of vocabulary, which is the pattern Liang et al. linked to lower perplexity and more false positives. Our guide to perplexity in AI detection explains the measure, and burstiness is the sentence-rhythm signal that detectors are often described as using.

If you want to look at those signals in your own text, the site has a free perplexity checker and a free burstiness checker.

How to read a Sapling score on your own writing

Treat a Sapling score as a pointer to passages worth rereading, and don't treat it as a finding about authorship. The steps below follow Sapling's own guidance.

1. Check how much text was scored. The free check covers the first 2,000 characters only.
2. Open the sentence-level view and see which sentences scored high. A low overall score can hide a few high-scoring sentences.
3. Check longer texts in sections. Sapling's FAQ suggests splitting a long text, so paste each section on its own and compare the scores.
4. Read the flagged sentences as a reader would. Stock phrases, long lists of hedges and very even sentence lengths are the usual causes.
5. Keep your drafts, notes and sources. Dated drafts are better evidence of authorship than any detector score.

Sapling's evaluation guide gives the same warning: "Don't base a consequential decision on the score alone." An instructor or editor may have used a score to question your work. If so, our appeal letter template for ESL researchers gives you wording for a reply.

Sapling compared with other detectors

Sapling is one of several free detectors, and each has a different approach and different published evidence. Our other detector posts cover them one at a time:

• Pangram's AI detector, reviewed for accuracy and false positives.
• GPTZero, including the independent tests of it.
• ZeroGPT, which had the best result in the Durham study above.
• Winston AI and how accurate its checker is.

Sapling is free to try, shows a score for every sentence, accepts PDF and Word files, and has an API with a free monthly quota for developers. It also supports eight languages. Its free limit of 2,000 characters is the main constraint for coursework, since a full essay or chapter needs the paid plan or several separate checks.

For an overview of how all of them performed across independent tests, see how accurate AI detectors are in 2026.

What Sapling's September 2026 update means for humanized text

Sapling says the 17 September 2026 model is better at detecting humanized AI text. That claim is worth knowing about, and it can't be checked from the outside, because no independent test of the new model has been published.

For a researcher, the practical point is that detector results change with every model update. A draft that scored low last spring can score differently today. That's why no humanizer, ours included, can promise a score.

If you use AI to help draft a paper, keep a record of your own work and follow your journal's or university's disclosure rules.

Frequently asked questions

How accurate is Sapling's AI detector?

Sapling's pages say accuracy varies with the dataset, the text length and the decision threshold, and they point to a whitepaper for results. The one independent test that reports Sapling's own numbers is a 2024 Durham University preprint on physics essays. It found all 30 AI essays caught and 18 of 30 human essays flagged, for 70% accuracy.

Is Sapling's AI detector free?

Yes. The free version checks the first 2,000 characters per check, and the Pro plan raises that to 100,000 characters for US$25 a month or US$12 a month billed annually. Sapling also offers a one-month Pro trial without a credit card.

Who makes Sapling?

Sapling Intelligence, Inc. makes it. The company says its team comes out of the Berkeley and Stanford AI labs and that it is backed by Y Combinator. Sapling sells grammar, autocomplete, rephrasing and AI-detection tools to developers and businesses.

Can Sapling's detector flag human writing as AI?

Yes, and Sapling says so on its detector page. It says short, general and essay-like text is more likely to be flagged, and that formulaic human writing can be flagged. In the Durham physics-essay test, 18 of 30 human essays were flagged.

Does Sapling's AI detector work on languages other than English?

It works on Spanish, Portuguese, French, German, Italian, Dutch and Japanese as well as English. Italian and Dutch support is preliminary. Text in other languages is scored by the English model, and Sapling says those results are less reliable.

Is Sapling's AI detector biased against non-native English writers?

Sapling is one of seven detectors that Liang et al. (2023) tested, and the group as a whole flagged an average of 61.22% of non-native English essays as AI. The paper doesn't report Sapling's own rate in its text. Sapling's tested version dates from March 2023, so a current score may differ.

How does Sapling score a text?

Sapling shows an overall AI-generation percentage and a score for each sentence. A Transformer-based model estimates the probability that each token is AI-generated, and sentence scores combine those token probabilities. The overall score is calculated over the whole text, so it isn't the average of the sentence scores.

Should a university use Sapling's score as evidence?

No. Sapling's own guide says not to base a consequential decision on the score alone, and it says no current AI detector should be used as a standalone check. Universities have reached the same conclusion about detector scores in general.

Can I lower a Sapling AI score on my own draft?

You can revise the passages the sentence view highlights, which usually means varying stock phrases and sentence length, and adding your own examples and reasoning. The ProofreaderPro.ai humanizer is designed to lower AI scores on AI-assisted drafts, with citations preserved. Detectors change with each update, so no tool can promise a particular score.

AI Proofreader

Fix grammar, word choice and academic tone in your own writing, with every change shown so the final wording stays yours.