detectionguide
Subscribe

Share

Detectors · Guide

The Best AI Detectors of 2026, Ranked on the Evidence

Pangram 4 leads, GPTZero is close behind, and every vendor now claims near-perfect accuracy. Here’s what independent tests actually show, and where each tool still breaks.

11 min read
Author
Mara OkaforOctober 14, 20269:00 AM UTC
Section
DetectorsSection
Length
11 minutesReading time
70%HUMANAICONFIDENCE · UNCALIBRATEDDG/143
In the lab, the best detectors almost never flag human writing. Real writing is not a lab.

Picking an AI detector used to mean picking the least embarrassing one. That has changed. The best tools in 2026 rarely flag ordinary human writing in controlled tests, catch nearly all AI text written from simple prompts, and have started on a problem that looked hopeless two years ago: telling an essay a chatbot wrote apart from one a chatbot merely tidied up.

They also still fail, in ways that are now well documented. This guide ranks the main detectors on what independent researchers have found, not on what vendors print on their pricing pages, and is specific about where each one breaks.

How we ranked them

This ranking draws on published independent tests, chiefly a 2025 University of Chicago working paper, Epoch AI’s July 2026 benchmark, an August 2026 University of Notre Dame study and several adversarial tests, along with the vendors’ own technical documentation. It is not a DetectionGuide bench run. In line with how we test, we weight false positives most heavily: a detector that wrongly accuses one honest writer in a hundred is worse than one that lets a little more AI through.

One caveat before the list. The two leading models, Pangram 4 and GPTZero 4o, shipped in July and September. Every independent study cited here tested an earlier version, so the newest headline numbers for both are the vendors’ own.

At a glance

Rank Detector Best for Starting price
1 Pangram Best overall: editors, publishers, anyone who needs few false positives Free tier; $20/month
2 GPTZero Teachers and students; a cautious second opinion Free tier; paid from about $13/month
3 Turnitin Institutions that already license it Institutional licenses only
4 Originality.ai Web publishers and SEO teams From $14.95/month
5 Copyleaks Enterprise, LMS and multilingual use Free trial; paid and enterprise plans

Prices as listed by the vendors in mid-to-late 2026. Plans change often, so check before buying.

1. Pangram: best overall

Pangram Labs launched in 2023 as Checkfor.ai, founded by Max Spero and Bradley Emi, and took its current name in 2024. It has since become the detector researchers reach for when they need a strong baseline, and 2026 put it in front of a much wider audience. Substack built Pangram scans into its apps in July, the company raised $9 million in a round led by Menlo Ventures, and its Chrome extension now labels posts on X, LinkedIn, Reddit, Medium and Substack as you scroll. Pangram told The New York Times its monthly users grew from 2,700 in June 2025 to 120,000 in June 2026.

What changed in Pangram 4

Pangram 4, released on July 29, 2026, is the main reason it tops this list.

  • A much larger model. It’s six times the size of Pangram 3.3 and the company’s first mixture-of-experts model, which sends different kinds of input to different parts of the network.
  • Three labels, not two. Every token gets a score for human-written, AI-assisted and AI-generated, computed over overlapping 512-token windows rather than fixed chunks. That lets Pangram 4 mark where AI passages begin and end in a mixed document, and tell a human draft that AI polished from paragraphs AI wrote outright. Pangram says no other detector makes both distinctions in a single pass.
  • Fewer false positives. On about 2 million held-out human documents written before 2022, Pangram reports a false positive rate of 0.0041%, roughly one in 24,000. It says AI-polished human writing is now labeled fully AI-generated about 0.01% of the time, down from 0.18% with Pangram 3.3.
  • Fewer misses. Across roughly 520,000 samples from 26 models, the share of AI text it missed fell to 0.34%, from 1.99% with Pangram 3. No model family it tested topped 0.7%.
  • Humanizer resistance. Pangram says it detects the AI in humanized text 98.83% of the time across 13 commercial humanizers. That builds on Pangram 3.2, which the company said quadrupled its humanizer detection rate in February. It also gave an AI agent 24 hours of API access to hunt for a bypass. The agent found one, by imitating terse, dictated surgical pathology notes, which Pangram ruled out of scope.
  • Cleaner training data. Human training text is licensed or company-owned, and Pangram says it doesn’t train on what users submit.
  • Images. Pangram Image launched the same day as a research preview, and image scans are now included in every plan.

The trade-off is price. Pangram now bills per 100 words rather than rounding every scan up to 1,000, and API pricing rose to $0.05 per 100 words, which the company concedes is up to 10 times more for long documents. The Individual plan is $20 a month for up to 300,000 words. Pangram 3 was retired on September 30.

What independent testers found

Pangram’s case doesn’t rest on its own numbers. In a 2025 University of Chicago working paper by Brian Jabarian and Alex Imas, which compared Pangram, GPTZero and Originality.ai on 1,992 human passages across six genres, Pangram had essentially no false positives or false negatives on medium and long passages. It was the only tool to stay under a strict 0.5% false positive cap without giving up detection, it held up after text went through a humanizer, and it was the cheapest per correct detection.

Epoch AI’s July 2026 test of Pangram 3.3.2 found the same pattern on newer models: no false positives on 495 human passages from 99 authors, all 297 basic-prompt AI passages caught, and the lowest miss rate of the three detectors when AI imitated a specific author’s style (10.1%, against GPTZero’s 10.8%).

Where it falls short

  • It flags more academic writing than GPTZero. In the Notre Dame study, Pangram 3.2 flagged none of 306 abstracts published from 2013 to 2015. But it flagged 14.9% of abstracts from 2023 to 2025, against GPTZero’s 8.9%. The authors stress these aren’t confirmed false positives, since some authors may have used AI without saying so. The rate swung sharply by field, from 2.1% in chemistry to 26.6% in theology.
  • It treats light editing as AI. When Gemini was asked only to refine an abstract, Pangram flagged 64% of the results even when the original predated ChatGPT by years. GPTZero flagged 38%. Whether that counts as a false positive depends on your rules, and plenty of journals and courses allow that kind of edit. Pangram 4’s AI-assisted label is meant for exactly this case, but nobody has tested it independently yet.
  • It can be beaten by someone who targets it. In a June 2026 paper, Kevin Ren, Manish Raghavan and Nikhil Garg used an automated loop to evolve rewriting prompts against Pangram and cut its detection rate from 94.7% to 24.1%. The journalist Alexios Mantzarlis built 588 samples around weaknesses he’d found, such as rhyme and repetition, and Pangram called 86% of them human. He still came away “relatively impressed” by how hard it was to fool.
  • New models open a window. The same paper found that after GPT-5.4’s release on March 5, Pangram recognized just 34% of its abstracts as AI, until Pangram 3.3 shipped in May. Pangram’s own analysis had put that figure at 99.5%, and the researchers couldn’t explain the gap.
  • Humanizers still beat older versions. In the Notre Dame study, after one pass through Undetectable AI, Pangram 3.2 missed more than 96% of AI-rewritten abstracts. GPTZero did no better.

The Substack rollout also showed the social cost of a good detector. Some writers compared it to a witch hunt, and Substack itself says Pangram can’t measure how much human effort went into a post or tell whether AI was only used for research.

Verdict: The best detector available, on the strength of its independent record on false positives and the most serious attempt yet to handle mixed writing. Keep a human in the loop, and be especially careful with academic prose outside the sciences.

2. GPTZero: best for classrooms

GPTZero was the first detector most teachers heard of, and its new model narrows the gap with Pangram. GPTZero 4o became the default for all users on September 20, 2026. The company says it produced one false positive in 10,402 essays from the PERSUADE student-writing corpus, and that it was rebuilt to resist paraphrasing.

The new features are aimed at people who have to explain a result. AI Patterns names the stylistic habits behind a score. Masked Detection ignores headings. And GPTZero says adding or removing em dashes or semicolons no longer changes a result, a direct answer to em-dash panic.

On benchmarks, the two leaders now trade blows. Both have re-run Epoch AI’s public test on their newest models, and each reports no false positives and under 3% of author imitations missed. GPTZero also claims a higher score than Pangram 4 on the DetectRL benchmark. None of these results are independent.

What the independent evidence does show is a difference in temperament. In Epoch’s July test, GPTZero matched Pangram with zero false positives and missed only slightly more. In the Notre Dame study it flagged fewer recent abstracts (8.9% to Pangram’s 14.9%) and far fewer lightly refined ones. But it also missed far more AI: on abstracts AI wrote from scratch, GPTZero missed 14.9% and Pangram 2.7%. The 2025 Chicago study found GPTZero lost much of its detection power once text went through a humanizer.

Verdict: A close second, and arguably the better fit for schools. It’s more conservative, and when the cost of a wrong call is a student’s record, missing some AI is the safer failure. The free tier covers occasional spot checks.

3. Turnitin: the one you don’t choose

Most students don’t pick Turnitin; their institution does. It isn’t sold to individuals, and none of the independent studies in this guide tested it, so judging it means relying largely on Turnitin’s own claims.

What’s worth knowing:

  • Since August 2025, it also looks for text run through “AI bypassers,” Turnitin’s term for humanizers. The check covers English only and runs automatically for institutions with AI detection switched on.
  • It won’t score fewer than 300 words of prose.
  • Since July 2024, scores below 20% appear as an asterisk rather than a number, because Turnitin found more false positives in that range.
  • Turnitin says its false positive rate is under 1% for longer submissions. That’s its own figure.
  • Clarity, its newer writing-process tool, records how a piece was written, which is stronger evidence than any percentage.

Verdict: If your school uses it, learn to read it. If you’re the one reading it, Turnitin’s own guidance is right: the score is an indicator, not a finding. And if you’ve been flagged, here’s what to do.

4. Originality.ai: for publishers

Originality.ai is built for web publishers, SEO agencies and editors screening freelance copy, and bundles AI detection with plagiarism checking. It offers several models, from Lite, which tolerates light AI editing, to Turbo, which is tuned to catch humanized text, plus an Academic model. In June 2026 it added an “AI Allowance” setting that lets users choose how much AI to tolerate, from 0% to 40%.

The independent record suggests Turbo leans aggressive. In Epoch’s test, Originality.ai Turbo 3.0.2 was the only one of three detectors to flag human writing, wrongly labeling 19 of 495 human passages (3.8%), and it missed the most author imitations (17.9%). The Chicago study also found it produced more false positives than GPTZero.

Verdict: Reasonable for a content team that controls the follow-up and wants a strict screen. A poor choice for judging students.

5. Copyleaks: for enterprises

Copyleaks sells mostly to institutions and businesses, through an API, learning-management integrations and broad language support. It claims more than 99% accuracy and a 0.03% false positive rate. Independent results are sparse and mixed, and it wasn’t included in the major 2026 studies above.

Verdict: Worth a look if you need many languages or already have an enterprise contract. Test it on your own writing first.

Also ran, and what to avoid

Winston AI claims 99.87% accuracy for its latest model on its own evaluation, but there’s little independent data to check that against. Treat any “best AI detector” list from a detector company accordingly. GPTZero’s own, last updated in June, ranks GPTZero first and Pangram second.

Three things to avoid entirely:

  • Chatbots used as detectors. In the Notre Dame study, GPT-5 Nano, used as a detector, flagged 46.7% of human abstracts written between 2013 and 2015, years before ChatGPT existed. Even OpenAI shut down its own classifier.
  • Off-the-shelf open-source classifiers. The open-source RoBERTa baseline in the Chicago study misclassified 30% to 78% of human text, depending on the scenario.
  • Free checkers that don’t publish error rates or model versions. If a tool won’t tell you how often it’s wrong, assume often.

What no detector does well yet

The leaders have closed most of the easy gaps. The hard ones remain:

  • Imitation. When Epoch AI had models mimic samples of real authors’ work, all three detectors missed 10% to 18% of the results, and a quarter or more of the scientific passages.
  • Targeted attacks. Anyone willing to iterate against a specific detector can usually find a way past it.
  • Brand-new models. Detectors learn from the models that already exist. A new release can open a gap for weeks or months.
  • Short text. Error rates rise as passages shrink, which is why Turnitin won’t score fewer than 300 words.
  • The line between help and authorship. Whether a refined abstract counts as AI writing is a policy question. A detector can report that AI touched a text; it can’t tell you whether that was allowed.

How to use any of them

  • Scan at least 300 words, and whole documents rather than excerpts.
  • Read the label, not just the number. “AI-assisted” and “AI-generated” mean different things.
  • When the stakes are high, run a second detector from a different vendor. Disagreement is information.
  • Never treat a score as proof on its own. Drafts and version history are better evidence, for the accuser and the accused.
  • Check the date. Every tool here changed at least once in 2026, and a review from last spring describes a different product.

Pangram is the best detector you can use in October 2026, and GPTZero isn’t far behind. Neither is good enough to be the only evidence against anyone.

Common questions

What is the most accurate AI detector in 2026?

On the independent evidence available in October 2026, Pangram. It had the lowest error rates in a 2025 University of Chicago study and in Epoch AI’s July 2026 test, though GPTZero was close behind in the latter. The newest models from both companies, Pangram 4 and GPTZero 4o, have not yet been independently tested.

What changed in Pangram 4?

Released on July 29, 2026, Pangram 4 is six times larger than its predecessor and labels each part of a document as human-written, AI-assisted or AI-generated, so it can separate light AI editing from passages AI wrote. Pangram reports a false positive rate of 0.0041%, about one in 24,000 documents, and says it detects the AI in humanized text 98.83% of the time. Those figures come from Pangram’s own benchmarks.

Is Pangram better than GPTZero?

Slightly, on most independent tests so far. In Epoch AI’s July 2026 test, neither flagged any of 495 human passages, and Pangram missed 10.1% of author-imitating AI passages to GPTZero’s 10.8%. In a Notre Dame study of academic abstracts, GPTZero flagged recent and lightly edited human writing less often than Pangram but missed more AI text. Which is better depends on whether you are more worried about false accusations or about missed AI.

Can AI detectors catch humanized text?

Sometimes. Pangram and Turnitin both now train against humanizers, and a 2025 University of Chicago study found Pangram held up well against one. But a 2026 Notre Dame study found that after a single humanizer pass, Pangram 3.2 and GPTZero both missed more than 96% of AI-rewritten abstracts, and researchers have repeatedly built prompts that evade a specific detector.

Can I use ChatGPT as an AI detector?

No. In a 2026 Notre Dame study, a general-purpose model (GPT-5 Nano) used as a detector flagged 46.7% of human-written abstracts from 2013 to 2015, years before ChatGPT existed. Chatbots are not calibrated detectors.

  • #pangram
  • #gptzero
  • #turnitin
  • #originality-ai
  • #copyleaks
  • #comparison
  • #guide

Written by

Mara Okafor

Mara has spent a decade covering the strange places where software meets judgment. She reads model cards for fun and has never once trusted a confidence score.

More from Mara

Keep reading

All stories
detectionguide
Subscribe

The arms race between the machines that write and the machines that judge.

↑↓ to move↵ to open/ to search anywhere