Do AI Detectors Work? GPTZero vs Turnitin vs Pangram (2026)

“Is this written by AI?” is now a question teachers, editors, recruiters and clients ask every day, and a whole industry of AI detectors promises an answer. But the numbers on their homepages (“99% accurate”) rarely match what independent researchers find, and a false accusation can cost a student a grade or a writer a client. In this guide I look at what the best available evidence says about the main AI detectors in 2026, including GPTZero, Turnitin, Originality.ai, Pangram and Copyleaks: how often they catch AI text, how often they wrongly flag human writing, who they are biased against, why dozens of universities have switched them off, and what you should do if a detector flags your own work.

TL;DR

Short answer: the best commercial detectors are good at spotting raw, unedited AI text in long passages, and much weaker on short, edited or paraphrased text.
Most accurate in independent research: Pangram, then Originality.ai and GPTZero. In the University of Chicago Booth working paper (August 2025), Pangram was the only tool to keep false positives under 0.5% without losing detection power.
False positives are real: about 1% or less on long passages for the top tools, but 2 to 3% on short passages, and free open-source detectors flagged 30 to 69% of human text.
Bias: a 2023 Stanford study found older detectors flagged 61.3% of essays by non-native English writers as AI.
Universities are backing off: at least 37 US colleges have disabled or declined AI detection, and Curtin University in Australia switched off Turnitin’s AI detector from January 1, 2026.
Never treat a detector score as proof. Use it as a reason to look closer, not a verdict.

How AI Detectors Work

AI text detectors are themselves machine-learning classifiers. They are trained on large sets of human-written and AI-written text and learn statistical differences between them. Early tools relied on measures like perplexity (how predictable each next word is) and burstiness (how much sentence length and structure vary). AI text tends to be smooth and predictable; human writing is lumpier.

Modern detectors such as Pangram, GPTZero and Originality.ai use deep classifiers trained on output from many models, but the basic idea is the same: they estimate a probability. They cannot “see” whether you used ChatGPT. That is why a detector can never be 100% certain, and why simple, formal or formulaic human writing can look “AI-like”.

What Independent Research Says

Vendor claims are marketing. These are the most useful independent sources:

University of Chicago Booth working paper (Jabarian and Imas, BFI WP 2025-116, August 2025): the most rigorous head-to-head comparison so far. Pangram was the only detector that met a strict 0.5% false-positive cap while still catching AI text well; Originality.ai came a close second, and GPTZero also performed well on longer passages. An open-source RoBERTa baseline flagged 30 to 69% of human text as AI, showing how bad cheap detectors can be. GPTZero disputes parts of the ranking.
Liang et al., Stanford, published in Patterns (2023): seven popular detectors of the time flagged 61.3% of TOEFL essays written by non-native English speakers as AI, while classifying native-speaker essays almost perfectly. 97.8% of the TOEFL essays were flagged by at least one detector.
Small 2026 tests by review sites keep finding the same pattern: raw AI output is caught most of the time, edited or “humanized” text much less often, and every tool produces some false positives. One August 2026 test of five detectors concluded that none reached 80% overall accuracy across all its sample types, although its sample was small.

False Positive Rates by Tool

A false positive means human writing labelled as AI. This is the number that matters most when someone’s grade or job is at stake.

Detector	False positives (independent)	Vendor claim
Pangram	About 0% on medium and long passages; never above about 1% on short ones (Booth)	About 0.01% overall
Originality.ai	1% or less on medium to long; about 2 to 3% on short (Booth)	Low
GPTZero	1% or less on medium to long; about 2.4% on short (Booth)	About 1%
Turnitin	Third-party reports of about 4% at sentence level	Under 1% at document level
Copyleaks	Not independently verified	About 0.2%
Open-source RoBERTa detector	30 to 69% (Booth)	–

Even a 1% rate adds up. Vanderbilt University noted when it disabled Turnitin’s AI detector in 2023 that with about 75,000 papers a year, a 1% false-positive rate would mean around 750 students wrongly flagged.

The Main AI Detectors Compared

Pangram

The best performer in independent research so far, with very low false positives on essays and long text. It also reports near-zero false positives on non-native English essays in its own re-tests. Best for: institutions and editors who want the lowest risk of false accusations.

Originality.ai

Popular with publishers and SEO agencies. Strong results in the Booth paper and in small 2026 tests, including better-than-average detection of paraphrased text. Also includes plagiarism and fact checking. Best for: content teams checking freelance work.

GPTZero

The best-known detector among teachers and students, with a free tier, sentence highlighting and writing-process replay in its Docs integration. Good on long passages, weaker on short ones. Best for: teachers who want a free first check plus context.

Turnitin AI detection

Built into the plagiarism tool most universities already license, so it is the detector students are most likely to meet. Turnitin added “AI bypasser” detection for humanized text in 2025. It is only available to institutions, it shows a percentage, and it does not report scores under 20% as a number to reduce false alarms. Best for: institutions already on Turnitin, used carefully.

Copyleaks

An enterprise detector with many languages and LMS integrations, and decent results against humanized text in some small tests. Its very low false-positive claim is not independently verified.

Free tools (ZeroGPT and similar)

Free detectors are the least reliable. In one 2026 test, ZeroGPT gave a different result for the same text 30% of the time. Do not use them to make decisions about people.

Bias Against Non-Native English Writers

The Stanford study is the strongest warning about detectors. Writers who use simpler vocabulary and more predictable grammar, which is common for people writing in a second language, look “AI-like” to perplexity-based tools. When the same TOEFL essays were rewritten with richer vocabulary, the false-positive rate dropped from 61.3% to 11.6%. Newer detectors claim to have fixed this, and Pangram reports near-zero false positives on ESL essays in its own tests, but those are vendor figures. Neurodivergent students and people who write in a very formal, structured style report similar problems.

Why Universities Are Switching Detectors Off

A list maintained by Programs.com (updated November 3, 2026) counts 37 US colleges and universities that have disabled, declined or discontinued AI writing detection, including Vanderbilt, UCLA, UT Austin, Northwestern, Yale, NYU, Johns Hopkins, UC Berkeley and, most recently, Washington State University (February 2026). Outside the US, Curtin University in Australia turned off Turnitin’s AI detector from January 1, 2026 while keeping its plagiarism checker. The common reasons:

False positives that can harm innocent students.
Bias against non-native and some neurodivergent writers.
Scores that cannot be explained or appealed in a meaningful way.
A shift toward assessment design: oral checks, in-class writing, drafts and process evidence.

When Detectors Fail

Short text. Under a few hundred words, error rates rise sharply.
Edited AI text. A human who rewrites AI drafts substantially often produces text no detector can reliably flag.
Mixed writing. Essays where some paragraphs are AI-assisted get confusing scores.
Formulaic human writing. Lab reports, legal text, translations and templates often look AI-like.
New models. Each new generation of models (GPT-6, Claude 5.5, Gemini 3.x and open models like DeepSeek V4) shifts the statistics, and detectors have to retrain to keep up.

Falsely Accused? What to Do

Stay calm and ask for the evidence, including the tool, the score and which passages were flagged.
Show your process. Google Docs or Word version history, outline notes, drafts, browser history of your research and handwritten notes are strong evidence.
Offer to explain your work in a short conversation. Being able to discuss your argument and sources is the best proof of authorship.
Point to the research. Share the Stanford study and your institution’s own guidance; many universities state that detector scores alone are not enough for a misconduct finding.
Use the formal appeal process if needed, and ask a student advocate or union for help.

To protect yourself in future, write in a tool that keeps version history and keep your notes. If you use AI for brainstorming or grammar where it is allowed, say so openly.

Advice for Teachers and Editors

Treat a high score as a reason for a conversation, never as proof.
Use the tools with the lowest independent false-positive rates, and avoid free detectors for decisions.
Check at least two signals: drafts, version history, knowledge of the topic in conversation, sudden style changes.
Set clear AI-use policies per assignment. Many problems come from unclear rules, not cheating.
Design tasks where process matters: personal reflection, in-class drafts, oral defence.

How Writers and Students Can Protect Themselves

Write in Google Docs or Word online so version history records every edit with timestamps. GPTZero and some other tools can also replay writing sessions in Docs.
Keep your research trail: notes, outlines, bookmarks, PDFs you annotated and rough drafts.
Declare allowed AI use in a short note, for example “I used Grammarly for spelling and Gemini to brainstorm three topic ideas.”
Do not run your work through “humanizer” tools. They often make writing worse, and Turnitin now flags humanized text specifically.
Keep your own voice: personal examples, opinions and specific details from your class or job are hard for both AI and detectors to fake.

How to Test a Detector Before Trusting It

If you are a teacher or editor choosing a tool, run your own small check rather than relying on marketing claims:

Collect 10 pieces of writing you know were written by people before 2022, including some by non-native English writers.
Collect 10 pieces you generate with current models (ChatGPT, Claude, Gemini), plus a few that you lightly edit by hand.
Run all 20 through each detector and note the scores.
Count how many human pieces were flagged (false positives) and how many AI pieces were missed. A tool that flags even 1 in 10 human texts should not be used for decisions about people.

Frequently Asked Questions

What is the most accurate AI detector in 2026?

In the most rigorous independent comparison (the Booth working paper), Pangram was the most accurate with the fewest false positives, followed closely by Originality.ai, with GPTZero also strong on longer text.

Can Turnitin detect ChatGPT?

Turnitin usually detects raw, unedited ChatGPT text in long essays. It is less reliable on short or heavily edited text, and third-party reports suggest sentence-level false positives around 4%.

Is GPTZero accurate?

GPTZero is accurate on longer passages, with false positives of about 1% or less in independent testing, rising to about 2.4% on short passages.

Can AI detectors be wrong?

Yes. Every detector produces false positives, and some groups, such as non-native English writers, are flagged more often. A detector score should never be used as the only evidence.

Do universities still use AI detectors?

Many do, mostly through Turnitin, but at least 37 US colleges and some international universities like Curtin have disabled AI detection because of reliability and fairness concerns.

Can AI detect text from Claude or Gemini, not just ChatGPT?

Commercial detectors are trained on output from many models, and raw output from ChatGPT, Claude and Gemini is flagged fairly reliably in tests. Accuracy varies with each new model release.

Do AI humanizer tools beat detectors?

Sometimes, but not reliably, and some detectors including Turnitin now look for humanized text. They also tend to reduce quality and can still count as misconduct under your institution’s rules.

Final Thoughts

AI detectors in 2026 are better than the early tools of 2023, and the best ones (Pangram, Originality.ai, GPTZero) rarely flag long human essays. But “rarely” is not “never”, short and edited text remain hard, and bias is a real risk. Use detectors as one signal among several, keep your drafts and version history, and focus on the process of writing rather than a single percentage. If you are a student looking for legitimate ways to use AI to learn faster, see my guide to the best AI tools for students, and if you want a private model that runs on your own PC, start with the best Ollama models.

Do AI Detectors Work in 2026? GPTZero vs Turnitin vs Originality.ai vs Pangram, False Positives and What Students Should Know