The Best AI Detectors in 2026, According to Independent Tests

What published studies found about false positives, missed AI text and humanized writing, what each detector offers for free, and which one fits which job.

We make one of the detectors on this page. Figures for the others come from independent studies or from the vendors themselves, labeled as such. We haven't tested other detectors ourselves.

By the AI Checker Humanizer team. Published . Updated .

The best AI detectors at a glance

On false positives, Pangram had the strongest results here. GPTZero tied it at zero in the Epoch AI test; Originality had a lower rate than GPTZero in the NBER study. These results don't establish an overall accuracy ranking, and on humanized text results varied a lot from one detector to the next. Among the 32 colleges we checked, Turnitin was named most often, but only schools can buy it.

The best AI detectors at a glance
DetectorWhat tests foundFree use
PangramFalse positive rates of 0.1% or less in the NBER study and none in the Epoch AI test; stayed robust against a commercial humanizer in the NBER study, though the Karr study found it missed most text run through one humanizerA 7-day free trial of the Individual plan
GPTZeroFalse positive rates of about 0.7% in the NBER study and none of 495 texts in the Epoch AI test; missed about half of the humanized texts in the NBER studyA free plan with up to 10,000 words a month
TurnitinIn a Temple University test, scored 28 of 30 human papers as fully human and 23 of 30 AI papers as fully AI, but gave a partial score to only 13 of 30 mixed papersNone: sold to schools, not to individuals
OriginalityFalse positive rates of 0.1% to 0.3% in the NBER study and 3.8% in the Epoch AI testA limited free plan
Our detectorIn our own English test: flagged 1 of 1,096 human texts and none of 600 essays by non-native English writers, and caught 270 of 278 AI texts10 checks a day of up to 2,000 words without an account; 30 a day of up to 10,000 words with a free account
Other detectors' details from their own pages, checked November 14, 2026; Turnitin's from its own guides, checked November 11, 2026.

These are the independent studies behind those numbers. They used different texts and methods, so compare detectors within a study, not across studies.

Independent studies of AI detector accuracy
StudyTexts testedDetectorsWhat they found
Liang et al., Patterns (2023)91 TOEFL essays by non-native English writers, 88 US eighth-grade essays7 detectors, including GPTZero and OriginalityNon-native essays wrongly flagged 61% of the time on average; the US essays were judged almost entirely correctly
Weber-Wulff et al., International Journal for Educational Integrity (2023)Human, AI and edited AI texts14 tools, including TurnitinDescribed the tools as “neither accurate nor reliable”; edits and paraphrasing made results worse
Jabarian and Imas, NBER working paper (2025)1,992 human texts across six genres, plus matching AI texts from four modelsPangram, Originality, GPTZero, an open-source modelFalse positive rates of 0.1% or less for Pangram, 0.1% to 0.3% for Originality, about 0.7% for GPTZero; the open-source model was far worse
Epoch AI (2026)495 human texts of about 500 words; AI texts from three current modelsPangram, GPTZero, OriginalityNo false positives for Pangram or GPTZero, 3.8% for Originality; when told to imitate a human style, AI texts were missed 10% to 18% of the time
Karr et al., preprint (2026)642 abstracts of published papers, 2013–15 and 2023–25Pangram, GPTZeroNo pre-2016 abstracts flagged; abstracts lightly refined with AI flagged 38% to 80% of the time; text run through a humanizer was missed over 96% of the time
Sources checked November 11 and 12, 2026.

How we picked the studies

  • Published methods and numbers. Each study says what texts it used and reports how often human writing was wrongly flagged, not just how much AI text was caught.
  • Mostly recent. Three of the five are from 2025 and 2026. We kept two from 2023: the Liang study, which tested essays by non-native English writers, and the Weber-Wulff study, which included edited AI text.
  • Vendor figures kept apart. When a detector's maker reports its own accuracy, we label it vendor-reported.

We list our detector's results separately and don't rank it against the others.

Detector-by-detector summary

Pangram

In the NBER study, Pangram's false positive rate was 0.1% or less, the lowest range of the detectors tested, and it stayed robust against a commercial humanizer. In the Epoch AI test it had no false positives, as did GPTZero. The Karr study, though, found it missed most text run through one humanizer. Pangram itself reports a false positive rate of about 0.01% on its own tests (vendor-reported). Its Individual plan costs $20 a month for up to 300,000 words, with a 7-day free trial, and Professional costs $65 a month. It also works inside Google Docs and Gmail.

GPTZero

GPTZero's false positive rate on ordinary human text was about 0.7% in the NBER study, and it flagged none of 495 texts in the Epoch AI test. In the NBER study it missed about half of the humanized texts. GPTZero says its error rate is under 2% when it reports high confidence, and reports a 1.1% false positive rate on TOEFL essays (both vendor-reported). Its free plan covers up to 10,000 words a month, and paid plans raise that limit. GPTZero describes itself as an official AI detector partner of the American Federation of Teachers.

Turnitin

Turnitin reports a document-level false positive rate under 1% for documents with more than 20% labeled AI (vendor-reported). In a Temple University test, it scored 28 of 30 human papers as fully human and 23 of 30 AI papers as fully AI, but gave a partial score to only 13 of 30 papers that mixed the two. Turnitin sells to schools, not to individuals, and each school decides whether AI detection is switched on. Our article on Turnitin AI detection has the details.

Originality

Originality's false positive rate was 0.1% to 0.3% in the NBER study, and 3.8% in the Epoch AI test. It has a limited free plan; Pro costs $14.95 a month, or $12.95 a month billed yearly, for 200,000 words a month. Plagiarism checking is included, and it pitches itself at website owners, content marketers and publishers.

Winston AI

None of the 2025 and 2026 studies in our table tested Winston AI. It offers a 14-day trial with 2,000 credits; Essential costs $18 a month, or $10 a month billed yearly, for 100,000 credits a month. Its plans also cover scanning pictures and handwriting, and AI image detection.

ZeroGPT

None of the 2025 and 2026 studies in our table tested ZeroGPT. It claims 98.4% detection accuracy and a false positive rate under 1% (vendor-reported). It has a free plan, and the checker on its home page takes up to 15,000 characters at a time. It also bundles paraphrasing, summarizing and other writing tools.

AI or Not

None of the studies in our table tested AI or Not. Its free plan includes 1 million words of text detection a month and 20 image checks; Pro costs $5 a month. An API key is included even on the free plan, and paid plans also check images, audio and video.

Our detector

Our own detection engine gives an overall score and colors each sentence in English and Portuguese, so you can see where the score comes from. On English texts it hadn't seen before:

Our detector on English texts it hadn't seen before
TextsHow manyResult95% bound
Human-written texts1,0961 flagged as likely AI (0.09%)At most 0.52%
Essays by non-native English writers6000 flagged as likely AIAt most 0.64%
AI-generated texts from 6 model families278270 caught as likely AI (97.1%)At least 94.4%
English, November 2026, detection engine version w0p5-38c6d591.

It scores English, Spanish, Portuguese and French, needs at least 150 words (200 in Spanish, Portuguese and French) for a verdict, and is free: 10 checks a day of up to 2,000 words without an account; 30 a day of up to 10,000 words with a free account. More on the AI detector page.

For QuillBot's and Grammarly's detectors, see our article on whether AI detectors are accurate.

Most accurate vs easiest to use

For accuracy, compare false positives and missed AI text within the same study, and check its date and detector version. For everyday use, check the word limit, account requirement and sentence highlighting.

Best free AI detectors

Free AI detectors compared
DetectorFree usePaid plans
PangramA 7-day free trial of the Individual planIndividual: $20 a month for up to 300,000 words; Professional: $65 a month
GPTZeroA free plan with up to 10,000 words a monthPaid plans raise the monthly limit
OriginalityA limited free planPro: $14.95 a month, or $12.95 a month billed yearly, for 200,000 words a month
Winston AIA 14-day trial with 2,000 creditsEssential: $18 a month, or $10 a month billed yearly, for 100,000 credits a month
ZeroGPTA free plan; the checker on its home page takes up to 15,000 charactersPersonal, business and API plans
AI or Not1 million words of text detection a month, plus 20 image checksPro: $5 a month
TurnitinNoneSold to schools, not to individuals
Our detector10 checks a day of up to 2,000 words without an account; 30 a day of up to 10,000 words with a free accountChecking stays free; paid plans are for the humanizer
Other detectors' details from their own pages, checked November 14, 2026; Turnitin's from its own guides, checked November 11, 2026. Prices in US dollars.

For an occasional essay, check the input limit first. Pangram and Winston AI offer time-limited trials; Turnitin has no free public version. If you check every week, compare how many words each free plan covers a month and whether you need an account. You can try ours below.

0 / 2,000 words

Best AI detector for teachers

For teachers, four things are worth checking: false positives, highlighted passages, price, and how the detector fits the way work is handed in.

  • False positives. Compare rates within one study, such as the NBER and Epoch AI rows in the table at the top. Be more careful with essays by English learners, where the 2023 detectors did far worse (more on that below).
  • Passages to review. A score alone doesn't show which sentences were flagged. Look for sentence or phrase highlighting; GPTZero and our detector both show it.
  • Price. GPTZero's free plan covers up to 10,000 words a month; ours is free for 30 checks a day with an account.
  • School systems. Turnitin runs inside the school's learning platform where the school has switched it on. Among the colleges we checked, Utah State University uses Copyleaks, on by default in Canvas. Originality offers a Moodle plugin. Ours has no LMS integration or class accounts; you paste or upload one piece at a time.

Whatever you use, a score is a reason to look closer, not proof. Many universities say in their policies that a detector score can't be the only evidence, and Turnitin says the same about its own tool. Our AI detector for teachers page covers false positives and how to talk to a student about a flag, and what AI detector colleges use lists 32 schools' policies.

Where detector results need extra care

  • Writing by non-native English speakers. In Liang's 2023 study, seven detectors wrongly flagged 61% of TOEFL essays on average. Newer tools say they have addressed this; GPTZero, for example, reports a 1.1% false positive rate on TOEFL essays (vendor-reported).
  • Short texts. A short piece gives a detector less to go on. Turnitin won't score fewer than 300 words of prose, and ours needs at least 150 words (200 in Spanish, Portuguese and French).
  • Humanized and paraphrased text. Research on the DIPPER paraphraser in 2023 showed rewriting AI text with another model could take one detector's catch rate from 70% to under 5%. Results on current tools vary by detector, as the studies above show. Check humanized text for changed numbers, missing negations and misplaced citations; our best AI humanizer guide has examples.
  • Writing that mixes human and AI. Turnitin gave a partial score to only 13 of 30 mixed papers in the Temple test, and in the Karr study abstracts lightly refined with AI were flagged 38% to 80% of the time.

Our article on whether AI detectors are accurate goes through each of these in more detail.

Choosing the best AI detector for your job

Students checking their own work

If your school uses Turnitin, you can't run it yourself, because Turnitin sells only to schools, and other tools will score your work differently. Use a free detector with sentence highlighting to find passages worth another look, and keep your drafts. Our AI essay checker covers the patterns teachers notice in essays.

Editors and SEO teams

Before checking a batch of articles, compare the price per word and the cost of extra checks. Originality includes a plagiarism checker, and Pangram's Individual plan covers up to 300,000 words a month. Look at which sentences a detector flags before you send a piece back.

Publishers and academic journals

A wrong flag on a submitted paper is costly, so look closely at false positive rates in the studies above. Treat any flag as a reason to ask the author, not a verdict: in the Karr study, abstracts lightly refined with AI were flagged 38% to 80% of the time.

Frequently asked questions

What is the most accurate AI detector?

No study we cite ranks detectors on overall accuracy. On false positives, Pangram had the lowest rate in the NBER study and tied with GPTZero at zero in the Epoch AI test. To compare overall accuracy, check missed AI text too, and the date of the test.

What is the best free AI detector?

It depends on how much you check. GPTZero's free plan covers up to 10,000 words a month and AI or Not's covers 1 million words of text a month. Ours runs 10 checks a day of up to 2,000 words without an account. Free limits change often, so check the date in the table above.

Which AI detector do universities use?

Turnitin appears most often in the policies we checked: 24 of 32 colleges named it. Many have its AI detection switched off, so these mentions don't show how widely it's used. Our article on what AI detector colleges use lists each school's policy.

Is GPTZero or Turnitin more accurate?

We can't point you to a recent independent test of both on the same texts. GPTZero's figures on this page come from the NBER study and the Epoch AI test, neither of which included Turnitin, and from GPTZero itself. Turnitin's come from a Temple University test of Turnitin alone and from Turnitin itself.

Can any AI detector catch humanized text?

Sometimes. In the NBER study, Pangram stayed robust against a commercial humanizer while GPTZero missed about half of the humanized texts. In the Karr study, text run through a humanizer was missed over 96% of the time. Turnitin added detection of humanized English text in August 2025.

Why don't you test every detector yourselves?

We make one of the detectors, so a test we ran on the others wouldn't be independent. Published studies with open methods are a fairer guide, and we list our own detector's numbers separately.

When was this page last checked?

See the “Updated” date at the top.

Pangram, GPTZero, Turnitin, Originality, Winston AI, ZeroGPT, AI or Not, Copyleaks, QuillBot and Grammarly are trademarks of their owners. AI Checker Humanizer isn't affiliated with any of them.