We independently research everything we recommend. When you buy through our links, we may earn a commission. Learn more ›
Writing · Buying Guides
The Best AI Detector (and Why You Shouldn't Trust Any of Them Blindly)
Pangram is the best AI detector for most editors, publishers and hiring managers: it had the lowest error rates in two independent academic studies. GPTZero is the better fit for classrooms. But no detector can prove that someone used AI, so treat any score as a reason to ask questions, not a verdict.
GPTZero did well on unedited AI text in the University of Chicago study and offers the most classroom-ready toolkit, including Canvas and Google Classroom integration and a Writing Replay feature.
Originality.ai bundles AI detection with plagiarism, fact and readability checks for $12.95 a month billed yearly, and it had a low false-positive rate in the University of Chicago study.
The Research Score rates each product on its own merits. Our picks also weigh price and who the product is for, so a pick can rank above a higher-scoring product. How we score
AI detectors promise to tell you whether a person or a chatbot wrote a piece of text. The honest answer is that the best of them are now quite good on longer passages, and the worst are close to guessing. Even the good ones still get things wrong, and a wrong flag can cost a student a grade or a freelancer a client.
Based on the independent research we reviewed, Pangram is the best AI detector for most people who need one. In a 2025 University of Chicago study, it was the only tool to meet a strict 0.5% false-positive cap, and a University of Maryland study found it was the only automatic detector that came close to a panel of expert human readers. GPTZero is our runner-up and the better choice for teachers, thanks to its classroom integrations and writing-process replay. Originality.ai is our budget pick for web publishers, and if your school already licenses Turnitin, use it, carefully.
Whichever you use, read our section on using results fairly first. We did not run hands-on tests for this guide.
We considered nine AI detectors: Pangram, GPTZero, Turnitin, Originality.ai, Copyleaks, Grammarly’s AI Detector, Winston AI, ZeroGPT and Sapling. Rather than rely on vendors’ own accuracy claims, we built this guide around independent research: a 2025 University of Chicago working paper, a 2025 University of Maryland study, a 2023 study on bias against non-native English writers, and guidance from Vanderbilt and MIT Sloan. We also read each company’s pricing, help and privacy pages, all checked on the date of publication. We scored our picks with our published Research Score. We did not run hands-on tests for this guide.
Who this is for
This guide is for people who have to judge other people’s writing: teachers and professors, editors and publishers who commission freelance work, and hiring managers reading cover letters and writing samples. If you’re looking for tools to help you write, see our guides to the best AI writing tools and the best AI tools for students.
The short version: no detector is proof
Any AI detector gives you a probability, not a fact. Here’s what the record shows:
OpenAI gave up on its own detector. OpenAI’s 2023 AI classifier correctly flagged only 26% of AI-written text and mislabeled human writing as AI 9% of the time. OpenAI withdrew it on July 20, 2023, “due to its low rate of accuracy.”
Detectors have been biased against non-native English writers. A 2023 study by Weixin Liang and colleagues found that seven popular detectors labeled 61% of TOEFL essays written by non-native English speakers as AI-generated, on average, while they were nearly perfect on essays by US eighth graders.
Universities have switched detection off. Vanderbilt disabled Turnitin’s AI detector in August 2023, citing false positives, bias and a lack of transparency. MIT Sloan’s teaching-technology team calls detection software “far from foolproof.”
The best tools have improved a lot, but not on everything. In the 2025 University of Chicago study, the top commercial tools were reasonably accurate on medium and long text but lost accuracy on passages under 50 words, and one detector largely failed on AI text that had been rewritten to evade detection.
How to use AI detector results fairly
If you’re going to use a detector, use it the way the vendors and researchers say to:
Treat a score as a question, not an answer. Turnitin says its percentage “should not be used as the sole basis for action.” GPTZero calls its result a “conversation starter.”
Set your policy before the assignment. Tell students, writers or applicants what AI use is allowed and that work may be checked. MIT Sloan recommends defining acceptable and unacceptable uses with examples.
Ignore short and low scores. The University of Chicago study found detectors less reliable on short passages. Turnitin hides scores below 20% for this reason and won’t score papers under 300 words of prose.
Look at process evidence. Drafts, Google Docs version history, GPTZero’s Writing Replay or Grammarly Authorship show how something was written. That’s stronger evidence than a percentage.
Talk to the writer. Ask them to walk you through their argument or sources. Vanderbilt recommends a direct conversation over relying on software.
Be extra careful with English learners, given the documented history of bias against non-native writers.
Don’t upload confidential text without checking the privacy policy. Some services may train on what you submit unless you opt out.
How we picked
We looked for detectors that:
Have independent evidence of accuracy, especially a low false-positive rate, because wrongly accusing a human writer is the costliest mistake.
Hold up against paraphrased or “humanized” AI text, not just raw chatbot output.
Show their work, with sentence-level highlights and process evidence rather than a single number.
Fit the reader’s workflow: LMS integrations for teachers, bulk scans and plagiarism checks for publishers.
Handle your data responsibly, ideally without training on your submissions.
Offer a free tier or fair price.
Top pick · The most accurate AI detector in independent studies
P
Pangram
8.6/10
Pangram had the lowest error rates in two independent academic studies, offers 20 free checks a day, and says it never trains on the text you submit.
Pangram is our top pick because its accuracy holds up in research it didn’t run itself. In a working paper from the University of Chicago’s Becker Friedman Institute, economists Brian Jabarian and Alex Imas tested Pangram, GPTZero, Originality.ai and an open-source RoBERTa model on 1,992 passages across six genres, from news and blogs to résumés and novels. The institute’s summary says Pangram had “essentially zero” false positives and false negatives on medium-length and long passages, and was the only detector to meet a strict policy cap of a 0.5% false-positive rate without losing its ability to catch AI text. It also kept catching AI text after it had been rewritten by humanizer tools.
A second study points the same way. In a University of Maryland-led study, a majority vote of five people who often use ChatGPT for writing caught 99.3% of AI-written articles with no false positives. Of the automatic detectors tested (Pangram, GPTZero and three open-source tools), only Pangram’s humanizer-focused model matched that panel; its base model caught 98.0% with a 2.0% false-positive rate.
Pangram’s free plan includes 20 checks a day. The $20-a-month Individual plan covers 300,000 words a month, image detection, a similarity check on every scan and OCR for scanned documents. It works in more than 20 languages and plugs into Chrome, Firefox, Google Docs, Canvas, Google Classroom, Moodle and Brightspace. Its privacy policy says, “Your submissions to Pangram are not used to train our AI or machine learning systems.”
Flaws but not dealbreakers
Pangram’s marketing claim of a false-positive rate of “1 in 10,000” is its own figure, measured on aggregated public data sets, and the Maryland study’s 2.0% rate for the base model is a reminder that results depend on the text you feed it. Both studies also found that every detector, Pangram included, is less reliable on short passages. And the University of Chicago paper is a working paper that hadn’t been peer reviewed; its authors expect results to shift as AI models and humanizer tools change.
Pangram’s word allowance for the Team plan isn’t listed on its pricing page, and plagiarism detection and extra credits require a paid plan.
Who it’s for
Editors, publishers, hiring managers and schools that need the lowest risk of falsely flagging human writing.
Who should skip it
You need a full plagiarism database and gradebook workflow in one tool, which a Turnitin license already gives your school.
Pros
Lowest false-positive rate in the University of Chicago study, including on AI text run through humanizer tools
Only automatic detector to come close to expert human readers in a University of Maryland study
Free plan with 20 checks a day; Individual plan $20/month
Says submissions are never used to train its models; SOC 2 Type 2
Cons
Its own headline accuracy figures are company claims
Short passages remain less reliable, as with every detector
Plagiarism checking and extra credits require a subscription
Research Score
8.6/10
Features (30%)9
Value for money (25%)8
Privacy & security (20%)9
Compatibility (15%)9
Docs & support (10%)7
Free plan (20 checks a day); Individual $20/month; Team $40/month for 2 seats; Professional $65/month (at the time of research)
Affiliate link · price at the time of research, may have changed
Runner-up · The best AI detector for teachers
G
GPTZero
7.5/10
GPTZero did well on unedited AI text in the University of Chicago study and offers the most classroom-ready toolkit, including Canvas and Google Classroom integration and a Writing Replay feature.
GPTZero is the tool many teachers already know, and it earns its spot. In the University of Chicago study, GPTZero kept false positives below 1%, according to a summary in the Chicago Booth Review, and caught nearly all unedited AI text. Its pricing page lists a free plan with 10,000 words a month, a Premium plan with 300,000 words, and a Professional plan with 500,000 words, batch uploads of up to 250 files and LMS integration with Canvas and Google Classroom.
What sets GPTZero apart for classrooms is that it looks beyond the final text. Its educator page describes Writing Replay, which shows copy-and-paste activity and replays how a student wrote a document, plus a Chrome extension and reports inside Google Docs. Process evidence like that is fairer to rely on than a single score. GPTZero’s own FAQ calls the detector a “conversation starter” and recommends asking students to show their process and looking at patterns across several assignments.
Flaws but not dealbreakers
The same University of Chicago study found that GPTZero largely lost its ability to detect AI text that had been rewritten by humanizer tools, with false-negative rates around 50% or higher in most genres and models. So a clean GPTZero result doesn’t mean much if a student is trying to evade detection.
GPTZero’s pricing page showed prices only in euros from our location, so we can’t quote US prices. And its privacy policy doesn’t clearly say whether text submitted through the website can be used for product development.
Who it’s for
Teachers and schools that want to look at how a document was written, not just a score.
Who should skip it
You often deal with AI text that's been paraphrased or run through a humanizer.
Pros
Free plan covers 10,000 words a month
Writing Replay shows copy-paste activity and typing patterns
Canvas and Google Classroom integrations on the Professional plan
Low false-positive rate in the University of Chicago study
Cons
Largely missed AI text that had been run through humanizer tools in that study
We couldn't confirm US prices on its pricing page
Privacy policy doesn't clearly rule out using submissions for product research
Research Score
7.5/10
Features (30%)8
Value for money (25%)7
Privacy & security (20%)7
Compatibility (15%)8
Docs & support (10%)7
Free plan (10,000 words a month); Premium and Professional paid plans, see current pricing
Affiliate link · price at the time of research, may have changed
Budget pick · For web publishers and content teams
O
Originality.ai
7.0/10
Originality.ai bundles AI detection with plagiarism, fact and readability checks for $12.95 a month billed yearly, and it had a low false-positive rate in the University of Chicago study.
Originality.ai is built for people who publish other people’s writing. Pro costs $12.95 a month billed yearly ($155.40 a year) or $14.95 month to month at the time of research, and covers 200,000 words a month. Monthly credits expire at month’s end. Beyond the AI checker, Pro includes a plagiarism checker, a fact and hallucination checker, readability and grammar checks, bulk scans, shareable reports, team management and full browser-extension access.
On accuracy, it sits in the middle. In the University of Chicago study, Originality.ai kept false positives below 1%, according to the Chicago Booth Review, which is the number that matters most if you’re deciding whether to pay a writer. But it missed more AI text than Pangram did.
Flaws but not dealbreakers
The Chicago Booth Review reports that Originality.ai’s false-negative rate ran from roughly 10% to 40%, depending on which AI model wrote the text. Its privacy policy also says, “Unless you opt out, we may use your scan history and scan results to help train and improve our models,” and it doesn’t list specific retention periods. If you’re scanning confidential manuscripts, opt out first.
Who it’s for
Editors and SEO teams who check freelance articles before publishing and want plagiarism and fact checks in one place.
Who should skip it
You need to catch every piece of AI text; it missed more AI writing than Pangram in independent testing.
Pros
Pro plan $12.95/month billed yearly, with 200,000 words a month
Combines AI, plagiarism, fact-checking and readability checks
Team management, shareable reports and bulk scans on Pro
Cons
Missed 10% to 40% of AI text, depending on the model, in the University of Chicago study
May train on your scan history unless you opt out
Free tier is very small
Research Score
7.0/10
Features (30%)7
Value for money (25%)8
Privacy & security (20%)6
Compatibility (15%)7
Docs & support (10%)7
Free (limited); Pro $12.95/month billed yearly ($155.40/year) or $14.95 monthly; Enterprise $136.58/month billed yearly (at the time of research)
Affiliate link · price at the time of research, may have changed
Also great · If your school already licenses it
T
Turnitin
Turnitin’s AI indicator is built into the Similarity Report that many instructors already use, and it hides low scores to reduce false accusations.
If your school already pays for Turnitin, its AI writing indicator appears inside the Similarity Report, visible only to instructors and administrators, including through LMSs such as Canvas, Moodle and Blackboard. It covers long-form English, Spanish, Japanese, Arabic, Dutch, German and Turkish, and in English it also flags text it thinks was reworked by paraphrasing or humanizer tools.
Turnitin is cautious about false positives. It says it aims to keep its false-positive rate under 1% for documents with more than 20% AI writing, and it shows scores from 1% to 19% only as an asterisk, with no percentage or highlights. It also states plainly that the percentage “should not be used as the sole basis for action or a definitive grading measure.”
Flaws but not dealbreakers
Turnitin’s accuracy numbers are its own, and it wasn’t among the tools in the independent studies we relied on. It also won’t analyze a paper with fewer than 300 words of prose, and it skips lists, headers, tables and code. Some universities have turned the feature off: Vanderbilt disabled it in 2023, estimating that even a 1% false-positive rate could have wrongly flagged around 750 of the roughly 75,000 papers it submitted the year before.
Who it’s for
Instructors at schools with Turnitin Feedback Studio who need AI checks inside their existing grading workflow.
Who should skip it
You're an individual buyer, or your school has switched the AI indicator off.
Pros
Built into the Similarity Report and LMS workflows
Doesn't display scores below 20%, to limit false positives
Detects text in seven languages, plus paraphraser and bypasser detection in English
Cons
Needs at least 300 words of prose; ignores lists, tables and code
No public pricing we could verify
No independent accuracy data in the studies we reviewed
Research Score
Not scored – not enough reliable information yet.
Through institutional licenses; no public price we could verify (at the time of research)
Affiliate link · price at the time of research, may have changed
Also great · For a free, quick second opinion
GA
Grammarly AI Detector
Grammarly’s AI detector is free to use, and Grammarly Authorship records how a document was written, which is better evidence than any score.
Grammarly’s AI detector is free on its website. More useful, for students and freelancers in particular, is Grammarly Authorship, which the company says sorts text by source, such as typed by you, pasted from an online source or generated by AI, and keeps a record of how a document was created. If you’re ever accused of using AI, a writing history like that is far more persuasive than any detector score.
Grammarly is also frank about the limits: its detector page says “there is no AI detector that can conclusively or definitively determine whether AI was used.”
Flaws but not dealbreakers
Grammarly claims 99% detection accuracy and a top ranking on the RAID benchmark, but those are the company’s claims, and Grammarly wasn’t part of the independent studies we reviewed. It also acknowledges that lightly edited AI text may evade detection and that human writing can be flagged.
Who it’s for
Writers who want to check their own work, or show how they wrote it, before submitting.
Who should skip it
You need to make decisions about other people's work; we found no independent accuracy data for it.
Pros
Free AI detector on the web
Authorship tracks typed, pasted and AI-generated text
Clear warnings that no detector can prove AI use
Cons
Accuracy claims come from the company
Lightly edited AI text may slip through, Grammarly says
Research Score
Not scored – not enough reliable information yet.
Free AI detector; AI Detector agent requires Grammarly Pro (at the time of research)
Affiliate link · price at the time of research, may have changed
Other good options
Copyleaks combines AI and plagiarism detection in a single report. Its pricing page lists a Personal plan at $13.99 a month billed annually ($167.88 a year), covering up to 300,000 words, with AI detection in more than 30 languages. It wasn’t in the independent studies we relied on, so we couldn’t compare its accuracy with our picks.
The competition
Winston AI: We found no independent accuracy data for it, and its pricing page mixed regular and promotional prices in a way we couldn’t reconcile.
ZeroGPT and Sapling both offer free checks with no sign-up. We didn’t find them evaluated in the independent studies we reviewed, so we can’t recommend relying on them for anything with consequences.
Open-source detectors such as the RoBERTa-based model tested by the University of Chicago researchers performed far worse than the commercial tools, with false-positive rates of roughly 30% to 78% depending on the scenario. The researchers called it unsuitable for high-stakes use.
We don’t cover “humanizer” tools that rewrite AI text to slip past detectors. Turnitin says it now flags that kind of rewriting in English submissions.
What to look forward to
Detection is an arms race. The University of Chicago authors recommend regular, published audits because results will shift as new AI models and humanizer tools appear, and Turnitin says it updated its English detection model in October 2026. We’ll update this guide as new independent evaluations are published. For students, see also the best AI writing tools for students.
FAQ
Which AI detector is the most accurate?
In the independent studies we reviewed, Pangram had the lowest error rates: it was the only tool to meet a 0.5% false-positive cap in a University of Chicago study, and the only automatic detector to approach expert human readers in a University of Maryland study.
Can an AI detector prove that someone used ChatGPT?
No. Every vendor in this guide says its score shouldn’t be the only evidence, and even the best tools make mistakes, especially on short text. Use a score as a reason to talk to the writer and look at drafts or version history.
Are AI detectors biased against non-native English speakers?
They can be. A 2023 study by Liang and colleagues found that seven detectors labeled 61% of TOEFL essays by non-native English writers as AI-generated, on average. Newer detectors perform better in recent studies, but the risk is a reason to be careful with English learners.
Is Turnitin's AI detector accurate?
Turnitin says it aims for a false-positive rate under 1% for documents with more than 20% AI writing, and it hides scores below 20%. We found no independent figures for it in the studies we reviewed, and some universities, including Vanderbilt, have turned it off.
