- 50 of 50 AI passages caught
- 0 of 50 human passages flagged
- A scan in a median of 6.4 seconds
$5/mo
Try Undetectable AIThe most comprehensive AI detector test ever conducted
Out of 32, I tested and ranked the 10 best AI detectors to see which ones really live up to their claims. The whole list is on the AI detectors page. Part of the AI tools directory on RankDude.


Tested and written by Miles Foster
Based on hands-on reviews, each graded on the same 10 tests, ranked by grade.
I signed up to every AI detector on this page myself, gave each the same briefs, published the sites for real and ran the same 10 tests on what was live. Every grade, screenshot and measurement is on its review, and each test site stays monitored for 30 days after the test. Here are the 10 best, highest first.
The 10 AI detectors in grade order: three findings from each review, the grade out of 5 and the lowest plan price its pricing page lists. The 10 test grades are on each review.
$5/mo
Try Undetectable AI$5/mo
Try Winston AI$25/mo
Try Sapling$8.99/mo
Try ZeroGPT$8.25/mo
Try Quetext$12.95/mo
Try Originality.ai$12/mo
Try WriteHuman$4.17/mo
Try AI or Not$6.99/mo
Try Decopy AI Detector$12.99/mo
Try SolvelyI labelled every passage as AI or human before opening any detector, pasted each one in once, kept the first verdict and counted every wrong one. Here is what I found.
Use an AI detector when you want a quick first look at whether a text reads as machine-written. Catching AI text was easy for every detector I tested. The difference is false flags: some accused real writers of being AI, so read the highlighted sentences, not only the score. Never treat a score as proof against a person. To rewrite text so it reads as human, use an AI humanizer; to find copied text, use an AI plagiarism checker.
I tested, reviewed and ranked every AI detector below. Each block has my verdict, its best and worst test results, the video review and the full story.
My pick for a free check that was right on every clean passage. It caught all 50 AI passages, flagged none of 50 human ones, got all 24 verdicts right in four languages and answered in a median of 6.4 seconds. The checker needed no credits.
Video review · 2:33
Undetectable AI caught all 50 AI passages and flagged none of the 50 human ones, which came from 25 books written between 1776 and 1915. The highest human score was 11%. It also got all 24 passages in Spanish, French, German and Chinese right, and a scan took a median of 6.4 seconds.
The weak spot is joined text: it passed 5 of the 10 passages that put a human half next to an AI half, scoring them 36% to 50%. It shows no margin of error, takes one file at a time with no batch or web address check, and keeps no scan history. A Humanize button sits under every result, because the same site sells a rewriter.
I ran 151 scans, about 44,000 words, on an account with a balance of 0 credits, and nothing stopped me. If you want a fast, free first check of a whole text, this is the one; do not use the percentage as proof. All the AI passages came from one model family, and the human ones were old books.
My pick for highlights that land on the AI sentences. It caught all 5 AI passages, flagged none of 5 human ones, marked the AI half of both mixed passages and answered in about four seconds. The trial is small, 2,000 words in all.
Video review · 3:54
Winston AI caught all five AI passages, each at 0% human, and flagged none of five human passages, which scored 98% to 100% human. In both mixed passages it coloured the AI sentences red and the human ones green, and it got all four non-English passages right. Every scan took about four seconds.
The trial pays for only 2,000 words, so my samples are small and all the AI text came from one model. On mixed text the headline score still leaned human, at 67% and 64% human, and it states no margin of error. Download is locked on the trial, and the API sits on a separate developer account.
It reads pasted text, text files, PDFs, Word files, batches of up to 50 files and web addresses, and it keeps a history with a share link per scan. If you want a quick check with clear highlights, it is a good choice; pay once you need volume. It ties with the first place at 4.9 and sits second under the ranking rule.
My pick for the fastest scans. It scored in a median of 0.5 seconds, caught all 5 AI passages and checked a batch of six files in 2.0 seconds. But it flagged 7 of 30 human passages as AI.
Video review · 3:43
Sapling returned a score in a median of 0.5 seconds across 42 scans, and it caught all five AI passages at 98.7% AI or more. In both mixed passages its sentence list found the AI half, with the AI sentences at 99.8% and 99.9% and the human ones at 0.0%. It got all four non-English passages right.
The problem is false flags: it called 7 of 30 human passages AI, with wrong scores from 62.6% to 95.4%. They included Frankenstein, Alice's Adventures in Wonderland, Jane Eyre and Walden, from books published between 1776 and 1899. On mixed text the headline score stayed low, at 2.9% and 2.4% AI.
It reads PDF, Word and text files, web pages and batches, and exports a CSV, but the free plan stops at 2,000 characters a check; I used the Pro trial for longer texts. Its Trustpilot score is 2.4 from 7 reviews, mostly about being flagged as AI. Use it as a quick first look, not as evidence against a writer.
My pick for a free plan with no stop. It caught all 20 AI passages, judged every non-English passage right, answered in a median of 1.2 seconds and took about a hundred scans with no limit. But it flagged 2 of 20 human passages.
Video review · 2:42
ZeroGPT labelled all 20 AI passages as AI, the lowest gauge at 88%, and judged all 20 Spanish, French, German and Chinese passages correctly. A verdict took a median of 1.2 seconds, and the free plan took about a hundred scans with no stop. It flagged 9 of 10 joined passages, five as AI and four with a hedge.
It also flagged two of 20 passages people wrote before 1930: a page of Alice in Wonderland at 75.7% and Pride and Prejudice at 39.2%. Scanning either again gave the same answer. It shows no reason for a flag, keeps no history, and on the free plan has no batch scan, no PDF report and no web address check.
Twice in about a hundred scans the button did nothing and showed no error, once for 90 seconds. Its terms say it does not train on your text and that a score is not definitive proof. Its Trustpilot score is 1.2 from 111 reviews. Use it for a quick check, not to accuse anyone.
My pick for naming mixed text as mixed. It caught all 5 AI passages at 100% AI, flagged none of 5 human ones, called both mixed passages Mixed and answered in about three seconds. My file upload failed.
Video review · 3:47
Quetext caught all five AI passages, each at 100% AI, and flagged none of five human passages, which scored 0.01% to 0.15% AI. It called both mixed passages Mixed, at 50% and 48% AI, and shaded the two AI sentences in each. It got all four non-English passages right, and a scan took about three seconds.
The free plan showed about 2,000 words, and I used 1,618 of them, so my samples are small and all the AI text came from one model. My file upload stopped with a message that AI detection needs at least 50 words, although the file held 56 and then 82, so that test carries a default grade of 3. Bulk checks and the PDF export are paid, and it states no margin of error.
Every scan is kept in a Reports list with its score and time. Quetext holds 4.6 on Trustpilot from 154 reviews, and I saw no charge on the free plan. If you want to see which sentences it marked, it is a good quick check; I could not test it at scale.
My pick for a clear call with a short explanation. It caught all 12 AI passages, flagged none of 7 human ones, flagged all 6 mixed passages and met 9 of my 10 clarity points. The free plan is 10,000 words, once.
Video review · 3:35
Originality.ai marked all 12 AI passages as AI, at 92 to 97 percent, and flagged none of 7 human passages, which came back 100 percent likely original. It flagged all 6 mixed passages, and a scan took a median of 5.5 seconds. It met 9 of my 10 clarity points: a plain call, sentence highlights, a short explanation, a share link, a history and the same result on three scans.
The free plan gives 100 credits once, which is 10,000 words with no refill, and I used 84 and stopped. On one of the three joined passages it marked the last human sentence in the same yellow as the AI one. I could not test other languages, humanisers or other AI models, and my human sample was small and old. Bulk scans need Pro, and the API is for Enterprise plans.
If you want a clear, quick answer on pasted text, it is a strong choice, but you get a few essays, not a month of work. Its Trustpilot score is 4.5 from 1,367 reviews, which I read by eye. It gives no margin of error.
My pick for a report in about a second, on short text. It caught all 3 AI passages and flagged none of 3 human ones. The free plan is 10 checks a month, and text goes in by pasting only.
Video review · 3:07
WriteHuman caught all three AI passages, each at 100% AI probability, and flagged none of three human passages, which scored 0% AI. A report appeared in about one second on all 8 scans. Every scan is kept in a history.
The free plan allows 10 detector checks a month, with a listed limit of 250 words a request, and text goes in by pasting only: no file upload, no batch, no web address. On my one mixed passage the headline read 0% AI and Likely Human-Written, although the two AI sentences were orange. That is a sample of one, and I could not test other languages, humanisers or six models.
WriteHuman is best known as a humaniser, and the detector sits beside it: reviewers on Trustpilot, where it holds 4.2 from 315 reviews, praise the humaniser far more. Use it for a quick check of a short text, and read the colours, not only the number.
My pick for a generous free plan on English text. It caught all 5 AI passages, flagged no English human passage and answered in half a second, with 1 million free words a month. But it flagged both human texts in other languages.
Video review · 3:14
AI or Not caught all five AI passages, at 87 to 100 percent AI, and flagged none of five English human passages, which scored 1 to 4 percent AI. A verdict took about half a second on every one of 19 scans, including a 2,008-word text. The free plan gives 1 million words of text checks a month, with no card.
In other languages it got two of four right: it caught the Spanish and German AI texts but called both human texts AI, Don Quijote (1615) at 99 percent and Les Miserables at 85 percent. It caught 1 of 2 mixed passages, scoring them 75% and 13% AI. Its upload box did nothing when I gave it a text file or a Word file, which may be a fault on my side.
It gives no explanation, no margin, no share link and no history. The same account also checks images, voice, music and video. Use it for quick English checks; do not use it on writing in other languages without checking the result.
My pick for a reason behind each flag. It caught my one AI passage, cleared my one human passage and gives a reason, a suggestion and a PDF report. The free plan allows about five short scans a month.
Video review · 4:10
Decopy AI Detector gave my AI passage 100% AI Probability and cleared a page of Pride and Prejudice from 1813 at 0%. It met 8 of 10 clarity points: a plain label, a percentage, a reason and a suggestion for each flagged sentence, a PDF report and a history kept for 30 days or 100 records.
The free plan gives 10 credits a month, and a scan cost 2, so I ran only four scans and stopped when 80 percent was used. On the mixed passage it highlighted the human paragraph red as well, and it missed an AI passage in Spanish at 34%. Each result rests on one passage, so none of it is a measurement.
A verdict took a median of about 16 seconds, from about 5 to 23, the slowest of the ten. I found no batch scan, web address check, extension or API for the detector. It holds 2.3 on Trustpilot from 8 reviews. Use it for a free first look, not to judge a person.
My pick for a free scan with no sign-up, but never for judging a person. It caught all 5 AI passages, but it flagged all 6 human passages as AI, and every scan landed between 65% and 87%.
Video review · 4:41
Solvely caught all five AI passages, scoring them 65% to 74%, but it flagged all six human passages as AI, at 65% to 87%. They included a page of Pride and Prejudice from 1813 at 84%. The human passages scored higher on average than the AI ones, so the score does not tell you who wrote the text.
It flagged both mixed passages, but their human halves too, and a sentence from a human book opened its list marked High AI. In other languages it got two of four right and flagged the Spanish and French human passages as AI. Every scan said 100% Confidence, even at 65%, and the methodology caps a detector that flags over 10 percent of human text at 2.
It is free to scan with no sign-up, takes PDF, DOCX and TXT files and a 2,500-word text, and gave the same answer each time, in about 8 seconds. Four of five engine scores are locked, and there is no history or export. It is mainly a study app; I would not use it to judge a person.
The ten above in one table: the measured average, my verdict, the final grade, and every test grade behind it.
| Rank | Tool | Price from | Test average | My verdict | Final grade | Tier |
|---|---|---|---|---|---|---|
| 1 | Undetectable AI | $5/mo | 4.38 of 10 tests | +0.6 | 4.9/5 | Excellent |
| 2 | Winston AI | $5/mo | 4.38 of 10 tests | +0.6 | 4.9/5 | Excellent |
| 3 | Sapling | $25/mo | 4.08 of 10 tests | +0.6 | 4.6/5 | Excellent |
| 4 | ZeroGPT | $8.99/mo | 4.08 of 10 tests | +0.6 | 4.6/5 | Excellent |
| 5 | Quetext | $8.25/mo | 4.48 of 10 tests | 0.0 | 4.4/5 | Good |
| 6 | Originality.ai | $12.95/mo | 4.17 of 10 tests | +0.2 | 4.3/5 | Good |
| 7 | WriteHuman | $12/mo | 3.67 of 10 tests | +0.4 | 4.0/5 | Good |
| 8 | AI or Not | $4.17/mo | 3.58 of 10 tests | +0.1 | 3.6/5 | Good |
| 9 | Decopy AI Detector | $6.99/mo | 3.77 of 10 tests | -0.2 | 3.5/5 | Good |
| 10 | Solvely | $12.99/mo | 2.98 of 10 tests | -0.9 | 2.0/5 | Below average |
Before I opened any detector I built one set of passages and labelled every one as AI or human. The human passages come from books written long before public AI, so a flag on one can only be a mistake. I pasted each passage in once, kept the first verdict and counted it against the label. Then I tried mixed passages, other languages, every way of loading a file, and timed each scan. On free plans the sample shrinks to what the allowance pays for, and the review says how many passages it rests on. Up to ten tests, graded 1 to 5; the grade is their average plus my verdict, which can move it by up to one point. A test I could not run properly gets no grade, and the review says so. The measurements sit in the test results linked from each review.
| Test | What I do | Example: Undetectable AI (#1) | Grade |
|---|---|---|---|
| Accuracy on AI-Written Text | Paste in AI-written passages once each; count how many it labels AI. | 50 of 50 AI passages caught | 5/5 |
| False Positives on Human Text | Paste in human passages from old books; count every one it flags as AI. | 0 of 50 human passages flagged; the highest score was 11% | 5/5 |
| Resistance to Humanisers & Paraphrasers | Run AI passages through humanisers; check the detector still catches them. | Not graded on this review; the humaniser test was left out | |
| Mixed & Lightly-Edited Text | Part-AI, part-human passages: does it spot the AI, and does it highlight the right part? | 15 of 20 mixed passages flagged; 5 of the 10 joined passages passed as human | 4/5 |
| AI Model Coverage | Catch rate for each AI model that wrote the passages, to find the blind spot. | Not graded on this review; all the AI passages came from one model family | |
| Non-English Text | AI and human passages in other languages; count wrong verdicts, human ones flagged first. | All 24 verdicts right in Spanish, French, German and Chinese | 5/5 |
| Document & File Support | Ten ways text gets in and out: paste, files, batch, long text, web address, export, API. | 6 of 10 points worked; no batch, web address check or free API | 3/5 |
| Score Clarity & Evidence | Ten clarity points: label, percentage, highlights, reasons, repeat scans, history. | 7 of 10 clarity points; no margin of error and no scan history | 3/5 |
| Scan Speed | Time every scan from pressing check to the verdict on screen. | A median of 6.4 seconds a scan; the slowest took 36.1 seconds | 4/5 |
| Pricing & Free-Tier Limits | Use the free plan until it stops; read every plan and what it unlocks. | The checker is free: 151 scans, about 44,000 words, on 0 credits | 5/5 |
Every measured grade comes from a tool I built and ran on every builder the same way. Each has its own page with how it works and the figure it gave every review.
You can see all the checkers I built on the checkers page.
| Tool | What it does | Reviews graded |
|---|---|---|
| Detector passage set | Gives every AI detector the same labelled passages: AI-written text, older human writing, mixed text and text in other languages. | 10 |
| Detector scan runner | Pastes each passage into the detector, presses Scan and times the wait until the verdict is on screen. | 10 |
| Detector file probe | Tries every way text gets into a detector: pasting, a text file, a PDF, a Word file, a batch of files and a web address. | 10 |
| Detector corpus | Gives every AI detector the same set of passages, each labelled as human, AI or mixed before any detector is opened. | 10 |
| Detector scanner | Pastes each passage into a detector in the held test browser, presses the check button and times the verdict. | 10 |
| File intake check | Uploads one passage as a text, PDF and Word file and checks that the detector reads each the same way as pasted text. | 10 |
| False flag counter | Counts how many passages written by people the detector accuses of being AI, using the detector's own label. | 10 |
| Detector scan timer | Pastes each passage into the detector, presses scan, and times it from the press to the verdict on screen. | 10 |
These monitors keep measuring the test sites after the review is live. You can see all of them, with their charts, in the monitors lobby.
| Monitor | How often | Builders |
|---|---|---|
| Uptime monitor | Every 5 minutes | 25 |
| Page speed tests | Every day | 23 |
| Security and privacy checks | Every day | 25 |
| Price tracker | Every day | 130 |
The place on this page is the headline grade of the review, highest first. When two tools share a grade, the higher measured score goes first, then the name. The grade words mean: 4.5 and up excellent, 3.5 to 4.4 good, 2.5 to 3.4 average, under 2.5 below average or poor.
The full procedure is on the AI detectors methodology page, and every tool I test with is on the checkers page.
Test a detector on something you know is human before you trust it on anyone, because some flagged human writing as AI. Check it in a language you write in, since one got both human texts in other languages wrong. Read the highlighted sentences, not only the percentage. Look for a margin of error or a reason behind the score, which most did not give. Check the free allowance before you plan work around it.
Ten detectors, the same kinds of labelled passages. One came out on top.

Undetectable AI won because it was right on every clean passage and the checker is free. It caught all 50 AI passages, flagged none of the 50 human ones from old books, got all 24 verdicts right in four languages, and gave a verdict in a median of 6.4 seconds. It is the only review here run on the full set of 144 passages: 151 scans, about 44,000 words, on an account with 0 credits and no limit I could find. Another detector scores the same 4.9; the ranking rule puts this one first.
It is not perfect. It passed 5 of the 10 passages that join a human half to an AI half, it shows no margin of error, it takes one file at a time, and it keeps no scan history. The human passages were old books, so I cannot say how it treats a modern essay, and the AI passages came from one model family. Use it as a fast first check and treat the percentage as a hint, not as proof. I ran every test on its free account.

The other nine are in the complete rankings, each with a one-line verdict on its card.
Undetectable AI ranks first at 4.9 out of 5 in my tests. Winston AI (4.9) and Sapling (4.6) follow.
Sapling, at 4.6 out of 5, is the highest-ranked AI detector here whose own pricing page offers a free plan (free plan at $0 with basic writing assistance). ZeroGPT (4.6), Quetext (4.4), WriteHuman (4.0), AI or Not (3.6), Decopy AI Detector (3.5), Solvely (2.0) also have a free plan. A free trial that ends, or a plan that needs a card up front, does not count as free.
In my tests Solvely flagged all 6 human passages as AI, with scores from 65% to 87%. Sapling flagged 7 of 30 old-book passages and ZeroGPT 2 of 20. AI or Not was right on every English human passage but called a Spanish and a French human text AI. Undetectable AI, Winston AI, Quetext, Originality.ai, WriteHuman and Decopy AI Detector flagged none of the human passages they scanned, on samples from 1 to 50 passages.
All ten were tested on a free account or trial. Undetectable AI took 151 scans, about 44,000 words, with no limit I found, and ZeroGPT took about a hundred without a stop. AI or Not gives 1 million words of text checks a month. Originality.ai gives 10,000 words once, Winston AI and Quetext about 2,000 words, WriteHuman 10 checks a month and Decopy AI Detector about five short scans a month. Solvely scans without an account, and Sapling checks up to 2,000 characters at a time on the free plan.
Each tool's grade is the average of the same 10 tests, run by hand on sites I published, plus my verdict, which can move it by up to one point and is explained on every review. The list is in grade order, highest first; a tie goes to the higher measured score. The full procedure is on the AI detectors methodology page.
Written by Miles Foster, Owner and tester, RankDude.