The best AI detectors are now quite accurate on long, unedited text, but they still make mistakes, especially on short, edited or non native English writing. For most people we recommend Pangram at $20 a month, which had the lowest false positive rate in a 2025 University of Chicago study. Treat any score as a reason to ask questions, never as proof on its own.
Top detectors perform well in independent tests on long passages, but accuracy drops on short or edited text, and no score is reliable enough on its own to punish a student, employee or freelancer.- Our pickPangramStarts at $20/month for Individual
- Budget pickGPTZeroStarts at $10/month, $8.33/month billed yearly
- Best free optionQuillBotFree (6 scans/day, up to 1,200 words a scan); Premium removes limits
- Upgrade pickCopyleaksStarts at $16.99/month for Personal, $13.99/month billed yearly; Pro $74.99/month billed yearly
- Also greatOriginality.aiStarts at $14.95/month, $12.95/month billed yearly
Pangram
Best for teachers, editors and businesses that want the fewest false accusations.
Pangram did best in the most rigorous independent comparison we found. Researchers Brian Jabarian and Alex Imas tested GPTZero, Originality.ai, Pangram and the open source RoBERTa detector in a 2025 working paper summarized by Chicago Booth Review Chicago Booth study ↗. Pangram had the lowest false positive rate of the group, essentially zero at most thresholds, and missed only about 2 to 4 percent of AI text Chicago Booth study ↗. The researchers said it was the only detector that held what they called policy grade performance across all four AI models they tested Chicago Booth study ↗.
The Individual plan costs $20 a month for up to 300,000 words, with a 7 day free trial, and adds image detection, similarity checks and OCR for scanned documents Pangram pricing ↗. It works in Google Docs and Gmail and through a browser extension Pangram pricing ↗, and Pangram also lists Google Classroom and Canvas integrations Pangram ↗. A Team plan costs $40 a month for two seats Pangram pricing ↗.
Be careful with Pangram's own marketing. Its homepage claims 99.98 percent accuracy and a false positive rate of 1 in 10,000, but it does not explain how the headline figure was measured Pangram ↗. The independent study is the better guide, and even it found that every detector struggled on passages under 50 words Chicago Booth study ↗. Hands on testing is in progress; we will update this guide with results.
Why we like it
- Lowest false positive rate in independent research
- Caught most AI text across four models
- Google Docs, Gmail, Classroom and Canvas support
- 7 day free trial
Flaws but not dealbreakers
- Marketing accuracy claims are not fully explained
- Free plan limits are unclear
- Weak on very short passages, like all detectors
GPTZero
Best for teachers and individuals who want a strong detector for less, with a free plan to start.
GPTZero is the best known AI detector, and in the same 2025 Chicago study it had a false positive rate below 1 percent and missed only about 0 to 2 percent of AI text Chicago Booth study ↗. That is a big improvement on 2023, when an academic study found that half of GPTZero's AI flags were false accusations 14 detector study ↗. Detectors change quickly, which is why we weight recent research.
GPTZero's pricing page loads prices in the browser, so we confirmed them through a pricing tracker dated August 27, 2026: Essential is $10 a month, or $8.33 billed yearly, for 150,000 words, and Premium is $16 a month, or $12.99 yearly, for 300,000 words GPTZero prices ↗. The free plan covers 10,000 words a month and does not expire GPTZero prices ↗. GPTZero also offers plagiarism checks, a Chrome extension and Authorship and Writing Replay, which shows how a document was written GPTZero plans ↗. Confirm the price at checkout, since it comes from a third party.
Why we like it
- Low false positive rate in recent research
- Free plan with 10,000 words a month
- Writing Replay shows how a document was written
- Annual billing saves about 45 percent
Flaws but not dealbreakers
- Prices not readable on its own page without a browser
- Older versions had many false positives
- Free plan caps each scan at 10,000 characters
QuillBot
Best for a quick free check of a single document, with a clear explanation of each flag.
QuillBot's AI detector is free for up to six scans a day, at up to 1,200 words a scan, and needs at least 80 words of text QuillBot detector ↗. It gives a 0 to 100 percent likelihood score, says whether text looks fully AI written, fully human or human with AI refinement, and highlights the passages it is unsure about QuillBot detector ↗. QuillBot cites a 99 percent detection rate on RAID, a third party benchmark QuillBot detector ↗.
We like that QuillBot is upfront about limits. Its page says scores are signals, not verdicts, warns that formulaic human writing such as legal or academic text can be scored wrongly, and advises against making academic or employment decisions on a score alone QuillBot detector ↗. It was not part of the Chicago study, so we have less independent evidence for it than for Pangram or GPTZero. For more than a quick check, GPTZero's free plan is the better free choice GPTZero prices ↗.
Why we like it
- Free, six scans a day
- Explains why passages were flagged
- Honest about what scores mean
Flaws but not dealbreakers
- 1,200 word limit per free scan
- Not included in the major independent study
- Downloadable reports need Premium
Copyleaks
Best for schools, agencies and teams that need AI and plagiarism checks across many languages and users.
Copyleaks combines AI detection and plagiarism checking in one report. AI detection covers more than 30 languages and plagiarism detection more than 100, and the Personal plan includes AI image detection, a browser extension and a Google Docs add on Copyleaks pricing ↗. That breadth makes it the better choice if you review work written in several languages or need to check for copying as well as AI.
The Pro plan costs $99.99 a month, or $74.99 a month billed yearly, and adds 25 user seats, website scanning by sitemap, cross language translation detection and an analytics dashboard Copyleaks pricing ↗. The Personal plan is $16.99 a month, or $13.99 billed yearly Copyleaks pricing ↗. Education plans are priced by student numbers and include LMS integrations and grade syncing Copyleaks pricing ↗. Copyleaks was not in the Chicago study, so we have less independent accuracy data for it.
Why we like it
- AI and plagiarism in one report
- AI detection in 30+ languages
- 25 seats on Pro
- Website scanning by sitemap
Flaws but not dealbreakers
- Not in the main independent accuracy study
- No free plan listed
- Monthly Personal plan covers only 25,000 words
Originality.ai
Best for publishers and agencies checking freelance web content, as long as you accept some missed AI text.
Originality.ai is built for content publishers. Pro costs $14.95 a month, or $12.95 billed yearly, for 2,000 credits, where one credit checks 100 words for AI Originality pricing ↗. It can scan full sites and URLs, tag scans and manage a team, and it has a Chrome extension that works in Google Docs Originality pricing ↗. The free plan allows three AI scans a day of up to 2,000 words Originality pricing ↗.
In the Chicago study Originality.ai rarely accused human writers, with a false positive rate below 1 percent, but it missed about 10 to 40 percent of AI text depending on the model Chicago Booth study ↗. That trade off suits a publisher that wants to avoid wrongly rejecting a writer, but it means AI text will slip through.
Why we like it
- Site and URL scanning
- Low false positive rate
- Cheap Pro plan
Flaws but not dealbreakers
- Missed 10 to 40 percent of AI text in research
- Credits expire monthly
Compare them
| Tool | Starting price | Free plan | Free trial | Independent false positive data | Plagiarism check | Best for |
|---|---|---|---|---|---|---|
| PangramOur pick | $20/mo | Yes, limits unclear | 7 days | Essentially 0% (Chicago 2025) | Similarity detection | Fewest false accusations |
| GPTZeroBudget pick | $10/mo ($8.33 yearly) | Yes, 10,000 words/mo | Not stated | Below 1% (Chicago 2025) | Yes | Budget and teachers |
| QuillBotBest free option | Free; Premium available | Yes, 6 scans/day | Not stated | Not in study | Not stated | Quick free checks |
| CopyleaksUpgrade pick | $16.99/mo ($13.99 yearly) | Not listed | Not listed | Not in study | Yes, 100+ languages | Multilingual teams |
| Originality.aiAlso great | $14.95/mo ($12.95 yearly) | Yes, 3 scans/day | Not stated | Below 1%, misses 10 to 40% of AI | Yes | Web publishers |
| Turnitin | Institutions only | No | Not stated | Claims under 1% | Yes | Schools already using it |
Prices and features from each vendor’s site, checked Oct 9, 2026. Prices change often.
Who this is for
This guide is for teachers and schools, editors and publishers who buy freelance writing, agencies, hiring managers reading cover letters, and business owners who want to know whether content they paid for was written by a person.
If you only want to check your own writing before you submit it, the free tools here are enough. If you plan to act on a result, such as failing a student or refusing to pay a writer, read the accuracy section first. The tools are better than they were, but they are still not evidence on their own Chicago Booth study ↗ 14 detector study ↗.
Why you should trust us
We did not rely on vendor accuracy claims, which are often unexplained Pangram ↗. Instead we read the published research: the 2025 University of Chicago comparison of commercial detectors Chicago Booth study ↗, the 2023 Stanford study of bias against non native English writers Stanford bias study ↗, a 2026 follow up that tested that bias again 2026 bias follow up ↗, and a 2023 academic test of 14 detectors 14 detector study ↗. We also read each vendor's pricing page on October 9, 2026, and every fact links to its source.
We have not tested these detectors hands on for this guide yet. Hands on testing is in progress, using known human and AI samples, and we will update this page with results. No vendor paid for placement.
How we picked
The most important number for an AI detector is the false positive rate: how often it wrongly says a human wrote with AI. A false negative means some AI text slips through. A false positive can mean an honest student is punished or a freelancer loses a client. So we ranked tools first on independent evidence of low false positives, then on how much AI text they catch, then on price and features.
Only three commercial tools have strong recent independent data, from the Chicago study: Pangram, GPTZero and Originality.ai Chicago Booth study ↗. For the others we relied on vendor documentation and were more cautious.
- Independent false positive rate, weighted most heavily
- Share of AI text caught across different AI models
- Honest explanation of results and limits
- Words per month for the price
- Plagiarism checking, integrations and team features
- A free plan or trial so you can try it
How accurate are AI detectors, really?
Early detectors were poor. OpenAI launched its own AI classifier in January 2023 and shut it down on July 20, 2023 because of its low accuracy: it caught only 26 percent of AI text and wrongly flagged human text 9 percent of the time OpenAI classifier ↗. A 2023 academic study of 14 tools, including Turnitin, found none reached 80 percent accuracy, and machine paraphrased AI text was caught only 26 percent of the time 14 detector study ↗.
The best tools have improved sharply since then. In the 2025 Chicago study, Pangram's false positive rate was essentially zero at most thresholds, and GPTZero and Originality.ai were both below 1 percent Chicago Booth study ↗. But the open source RoBERTa detector was close to random guessing in many cases, and every commercial tool did worse on passages under 50 words Chicago Booth study ↗. Results also varied by which AI model wrote the text and which threshold was used Chicago Booth study ↗. In other words, which detector you use matters a great deal.
Even a 1 percent false positive rate adds up. Vanderbilt University disabled Turnitin's AI detector in August 2023, noting that at Turnitin's stated 1 percent false positive rate, about 750 of the 75,000 papers it submitted in 2022 could have been wrongly flagged Vanderbilt ↗. Turnitin itself says its tool does not determine misconduct and that instructors should give students the benefit of the doubt when the evidence is unclear Turnitin ↗.
Editing defeats detectors more than anything. In the 2023 study, manually edited AI text was misclassified as human about half the time 14 detector study ↗. Tools that rewrite AI text to dodge detectors, often called humanizers, are part of what the Chicago researchers called a likely technical arms race Chicago Booth study ↗.
False positives and non native English writers
The best known study on detector bias came from Stanford in 2023. Testing seven detectors, Liang and colleagues found an average false positive rate of 61 percent on TOEFL essays by non native English speakers, while essays by US eighth graders were classified almost perfectly Stanford bias study ↗. About one in five of the TOEFL essays was flagged as AI by all seven detectors Stanford bias study ↗. The likely cause is that simpler, more predictable wording looks like AI to some detectors.
A 2026 follow up suggests things have improved but not vanished. Researchers at Charles University found no systematic bias against non native writers in Czech, and in an English side test a commercial detector's false positive rate on the same TOEFL essays fell to 23 percent, down from the 61 percent average in 2023, though a gap between native and non native essays remained 2026 bias follow up ↗. The 2023 academic study also found that human text translated by machine was far more likely to be flagged 14 detector study ↗.
If you teach or employ non native English speakers, set a higher bar before acting on a flag. Pangram and GPTZero performed best in the latest independent tests Chicago Booth study ↗, but no study has shown that any tool is free of this problem.
How to use an AI detector fairly
Every vendor we checked, including Grammarly, QuillBot and Turnitin, says its scores should not be used alone to decide misconduct Turnitin ↗ Grammarly detector ↗ QuillBot detector ↗. The most useful features are the ones that give you more evidence, such as GPTZero's Writing Replay, which shows how a document was typed GPTZero plans ↗.
Write down your policy before you need it, and tell students or freelancers that you use a detector and that false positives happen. Turnitin recommends this approach Turnitin ↗.
- Never act on a score alone; talk to the writer first
- Ask for drafts, notes or version history
- Be extra cautious with short text under about 300 words
- Be extra cautious with non native English writers
- Run a second detector before raising a concern
- Check your privacy policy before uploading student or client work
What to look forward to
The Chicago researchers proposed that schools and companies set a strict cap on false positives, such as 0.5 percent, and compare detectors against that same standard, with regular published audits Chicago Booth study ↗. We would like to see vendors publish results that way instead of headline accuracy numbers.
OpenAI said when it closed its classifier that it was researching provenance techniques for text OpenAI classifier ↗. Watermarking or content credentials built into AI tools could one day make detection less of a guess. Until then, use detectors as one signal among several. We will update this guide when our own testing is complete.
The competition
TurnitinInstitutions only
Turnitin is sold to schools, not individuals, and claims a false positive rate under 1 percent Turnitin ↗. It ranked highest in a 2023 academic test 14 detector study ↗, but Vanderbilt disabled its AI detector over transparency and false positive concerns Vanderbilt ↗. Use it if your school already has it, with the same caution as any detector.
GrammarlyFree; Pro $12/month billed yearly or $30 monthly
Grammarly offers a free AI detector with no sign up, and it is honest that no detector is 100 percent accurate and that human text can be flagged Grammarly detector ↗. It does not publish accuracy figures and was not in the independent studies we used, so we did not pick it.
ZeroGPTNot stated
ZeroGPT has a free plan and paid plans, but its pricing page lists no prices or word limits and gives only general claims about accuracy ZeroGPT pricing ↗. Without independent evidence, we cannot recommend it for decisions about people.
Questions people ask
Are AI detectors accurate?
The best ones are now fairly accurate on long, unedited text. In a 2025 University of Chicago study, Pangram's false positive rate was essentially zero and GPTZero's was below 1 percent Chicago Booth study ↗. But accuracy drops on text under 50 words and on edited AI text, and some tools are close to random Chicago Booth study ↗ 14 detector study ↗.
Can an AI detector be wrong about human writing?
Yes. A 2023 Stanford study found seven detectors wrongly flagged 61 percent of essays by non native English speakers on average Stanford bias study ↗. Even at a 1 percent false positive rate, Vanderbilt estimated about 750 of its 75,000 yearly papers could be wrongly flagged Vanderbilt ↗.
What is the most accurate AI detector?
In the most rigorous recent independent comparison, Pangram had the lowest false positive rate and was the only tool that held policy grade performance across all four AI models tested Chicago Booth study ↗. GPTZero was close behind Chicago Booth study ↗.
What is the best free AI detector?
GPTZero's free plan covers 10,000 words a month and does not expire GPTZero prices ↗. QuillBot's free detector allows six scans a day of up to 1,200 words and explains each flag QuillBot detector ↗. Grammarly's free detector needs no sign up Grammarly detector ↗.
Can teachers prove a student used ChatGPT?
Not with a detector alone. Turnitin says its tool does not determine misconduct Turnitin ↗, and a 2023 academic study concluded detectors offer no verifiable evidence 14 detector study ↗. Drafts, version history and a conversation with the student are better evidence.
Does editing AI text fool detectors?
Often. In a 2023 study, manually edited AI text was classified as human about half the time, and machine paraphrased AI text was caught only 26 percent of the time 14 detector study ↗.
Why did OpenAI shut down its AI detector?
OpenAI withdrew its classifier on July 20, 2023 because of its low accuracy. It caught 26 percent of AI text and wrongly flagged 9 percent of human text OpenAI classifier ↗.
Where this comes from
Checked Oct 9, 2026. Vendor pages describe their own products; independent sources add outside views.
01