The short answer
For a personal draft check, compare GPTZero’s passage highlighting with Grammarly’s percentage-based free check. For classroom workflows, examine the integrations documented by GPTZero, Copyleaks and Turnitin. For a record of the writing process, look at writing-replay or Authorship features.
A flag should prompt a review of the text and its context, not an automatic rewrite of genuine human writing. Retain drafts and sources. Trying known-authorship examples can reveal practical problems, but a few examples cannot establish a detector’s accuracy.
Which detector fits which need
These documented features were checked on October 14, 2026. Compare the task each tool fits with the limits in the final column.
Scroll horizontally to see the full table. Keyboard: focus the table and use the left and right arrow keys.
| Your need | Documented features that fit | What to check before relying on it |
|---|---|---|
| Check my own draft before I submit it | GPTZero and ZeroGPT document passage highlighting; Grammarly offers a free percentage check. Copyleaks adds detail with an account. Proofademic offers one free scan for new users. | No vendor page we read says its result will match the tool your instructor or client uses. Keep drafts and notes, and read your own text for the passages that were flagged. |
| Classroom or learning-system use | Copyleaks lists Canvas, D2L, Moodle, Blackboard, Schoology, Edsby and Sakai. GPTZero lists Canvas and Google Classroom. Originality.ai lists a Moodle plugin. GPTZero and Copyleaks also describe security certifications (SOC 2) on their own pages. | Your institution’s policy on detector scores as evidence, student data terms, and the cost for your size. We did not verify pricing or billing for any product. |
| Scanning content at volume for a publishing or editorial team | Originality.ai lists bulk scanning, team roles, an API, plagiarism and fact checking. Copyleaks lists an API and a WordPress integration. ZeroGPT lists batch file upload and an API. GPTZero lists an API. | How the vendor handles a disputed flag, how scan credits are counted, and whether the false positive rate holds for your genre. The sources reviewed here do not settle all of these questions. |
| Evidence of how a document was written, not just a score | GPTZero’s Google Docs writing replay, Grammarly Authorship and the Originality.ai Chrome extension’s writing replay record the writing session and show how text arrived in the document. | These only help if they were running while the document was written. They show activity, and a recorded session is not proof of who had the ideas. |
| Writing in a language other than English | Copyleaks says it supports more than 30 languages and publishes its own per-language figures. GPTZero says it fully supports English, German, Portuguese, French and Spanish. ZeroGPT says it supports all languages. | Those are vendor statements. ZeroGPT shows no per-language data on the page we read. Test with known human text in your language before using a result. |
| Text that mixes human writing and AI edits | Most of these checkers highlight passages, while Grammarly’s passage-level detail is in its paid detector agent. GPTZero reports 96.5% accuracy on mixed documents. The pages reviewed do not describe that evaluation. Proofademic describes sentence-level scoring and a study set that included 2,600 mixed-authorship documents. | Both figures are vendor-reported. The GPTZero pages reviewed do not describe that test; Proofademic used synthetically spliced documents. RAID and Weber-Wulff report that edited or paraphrased text lowers detector performance in their tests. |
| Programmatic access | API access is listed by GPTZero, Copyleaks, Originality.ai and ZeroGPT. | Terms, rate limits and prices. We did not read API pricing or documentation. |
| Worry about false flags on non-native English writing | GPTZero says it is tuned to reduce false positives on English-as-a-second-language writing. Copyleaks says it has low false positive rates on non-native English text and links a post about it. Proofademic reports a false positive rate on non-native English writing of 0.5% in its study headline and 0.6% in the study table (620 documents, 0.85 confidence threshold); the page is inconsistent between the two. | Liang et al. found a 61.22% average false positive rate across seven detectors on TOEFL essays in 2023. The studies summarized below do not verify these current vendor claims. |
Key takeaways
- Choose by the job first: a free pre-submission check, classroom or learning-system use, volume scanning, proof of how a text was written, a language other than English, or programmatic access.
- The current product accuracy claims summarized here are vendor-reported. First-place claims can refer to different datasets, dates and metrics, so they do not establish a comparable ranking.
- The independent studies below used different texts, detectors and dates, so their numbers cannot be combined into a league table, and study aggregates are not results for one product.
- Short text, non-native English writing and edited or paraphrased text are the conditions where the research and the vendors’ own pages report the most trouble.
- Use a detector score as a prompt for questions, never as proof about who wrote a text.
Detector profiles: capabilities and limits
Each row combines what the vendor says the product does with a limit the vendor or the evidence supports. Where a row has an AIHT guide, it covers that product in more depth.
Scroll horizontally to see the full table. Keyboard: focus the table and use the left and right arrow keys.
| Detector | What the vendor documents | Stated or evidence-based limit | Evidence status |
|---|---|---|---|
| GPTZero | Free web scan (10,000 characters before an account is needed), sentence highlighting, Chrome extension, Google Docs writing replay, Canvas and Google Classroom integrations, API, plagiarism and hallucination checkers. Five languages listed as fully supported. | Its own FAQ says accuracy improves with longer inputs, is strongest on English prose, and that results should not be used to punish or as a final verdict. | Accuracy figures are vendor-reported, including a RAID-based claim and a university partner. |
| Originality.ai | AI checker with plagiarism, bulk scan, readability, fact checking and team roles. Chrome and Firefox extensions, Moodle plugin and an API. Free check of up to 2,000 words, 3 scans a day. | Pricing and credit rules were not read. Fact-checking and quality features are separate from AI detection and do not support its accuracy. | Says it is the most accurate in third-party studies and links its own round-up, which we did not review. |
| Copyleaks | Free scan up to 25,000 characters, 30+ languages, AI Logic explanations of flagged phrases, extensions, Google Docs add-on, LMS integrations, API and enterprise controls such as adjustable sensitivity. | Its detector page notes that grammar tools such as Grammarly can be flagged as potential AI content. Some detail needs an account or a paid plan. | Accuracy and false positive figures, including per-language numbers, are vendor-reported. |
| ZeroGPT | Free web check with highlighted sentences (15,000-character input box), file upload, batch upload, PDF reports and an API. The company also sells an AI humanizer and paraphraser. | The homepage gives accuracy and false positive figures without the data or method behind them. | Vendor-reported only. |
| Grammarly | Free percentage score, plagiarism check, citations, an Authorship activity report, and an AI Rewriter for flagged phrases. Grammarly also sells a separate AI Humanizer tool. A paid detector agent gives phrase-level detail. | Its FAQ says no detector is 100% accurate and that lightly edited AI text may evade detection. It also says its model was trained on texts created before 2021 while updating continuously. | Says it ranks first on RAID. That is Grammarly’s statement, and we did not verify the ranking. |
| Proofademic | A one-time free scan of up to 1,000 words with sentence-level highlighting and per-segment confidence. AIHT’s detector directory lists it among the scorecard detectors. | Its research page says robustness against humanizer tools is an active area of work, not a solved problem. The 0.2% document false positive rate comes from one internal vendor evaluation: 4,200 human-written academic texts, a 0.85 confidence threshold, English only. | Internal vendor study, May 2026. The HTML report was reviewed; AIHT has not reproduced the results. |
| Turnitin | Document report for long-form prose, with no score or highlights for results in the 1% to 19% range. | Requires sufficient qualifying prose; the FAQ describes a 300-word minimum in evaluated sentences, depending on language. A report can miss AI text. | Vendor validation with a stated reporting threshold; see the vendor-claims section below for the conditions. |
AIHT has not run its own detector accuracy test, so none of the figures on this page are AIHT measurements. The stored detector figures on AIHT comparison pages are editorial baselines, not a new detector study, and they are not used here. See the detector directory and the testing methodology for their status.
Accuracy, false positives and false negatives
A detector makes two kinds of mistakes, and they have different costs. Accuracy blends them together, which is why it can mislead.
Scroll horizontally to see the full table. Keyboard: focus the table and use the left and right arrow keys.
| Term | Plain meaning | Why it matters |
|---|---|---|
| Accuracy | Share of all tested texts labeled correctly. | Depends on how many AI and human texts were in the test. |
| False positive rate | Share of human-written texts wrongly flagged as AI. | This error lands on the writer. |
| False negative rate | Share of AI-written texts the detector misses. | Changing a decision cutoff can trade false positives against false negatives; the result depends on the detector and test data. |
| Threshold (cutoff) | The score above which a text is called AI. | Moving it trades one error for the other. |
Illustrative example, not from any study: imagine 200 essays, of which 10 used AI. A tool that labels every essay as human is 95% accurate and catches none of the 10. The number sounds impressive and tells you nothing about the case you care about.
Why test-set balance and thresholds change the story
Balance. If a test contains mostly human text, accuracy mostly measures how well the tool recognizes human text. If it contains mostly AI text, the reverse. A claim is easier to interpret when the numbers of human and AI texts are stated, along with who wrote the human texts. Liang et al. found very different behavior on essays by non-native English writers than a general-purpose test would suggest (see the table below).
Thresholds. A cutoff changes which texts are flagged. Turnitin defines false positives as fully human-written text incorrectly identified as AI-generated. Its FAQ describes a target below 1% for reports showing more than 20% AI writing, and acknowledges that AI text can be missed. It suppresses scores below 20%. These conditions matter when comparing its headline rate with another detector.
Scale. Illustrative arithmetic: a 1% false positive rate sounds small, yet across 1,000 fully human essays it would still mean about 10 wrongly flagged writers.
How to read vendor claims
Vendor pages are useful for what a tool says it does. They are not independent evidence of how well it works. These are the accuracy statements we found, each attributed to the vendor.
- GPTZero reports 99% accuracy on its homepage. Its October 2025 RAID evaluation post reports a 95.7% detection rate at a fixed 1% false-positive operating point, without adversarial attacks. That is a vendor-run comparison, with a specified cutoff.
- Originality.ai says it is the most accurate detector in third-party studies and links a round-up of them. We did not review those studies, so we treat the claim as the vendor’s. Checked October 14, 2026.
- Copyleaks advertises over 99% accuracy and 0.03% false positives on its product page. Its V11 methodology, tested September 10, 2026, reports 0.30% false positives and 0.64% false negatives at default sensitivity on English text. These different vendor figures should not be treated as interchangeable.
- ZeroGPT reports 98.4% detection accuracy and a false positive rate below 1% on its homepage, checked October 14, 2026. The page does not give the evaluation data or method. See our ZeroGPT guide for that product in depth.
- Grammarly says its detector reaches 99% detection accuracy and ranks number one on the RAID benchmark. Both are Grammarly’s own statements. Our Grammarly guide covers what is and is not independently known.
- Proofademic publishes one study of its own sentence-level model, evaluated on 4,200 human-written texts and 2,600 mixed-authorship documents, reporting a 0.2% document-level false positive rate. The vendor describes limitations, but this guide does not establish independent reproduction of the result.
- Turnitin reports a target below 1% false positives under its stated reporting conditions. It describes tests on more than 700,000 additional academic papers written before widely used AI chat tools. Those are vendor validation records.
What three well-known studies measured
These three studies examine different aspects of detector reliability. Each row states its date, scope, measured outcome and main limit. Read the rows side by side, not as a ranking.
Scroll horizontally to see the full table. Keyboard: focus the table and use the left and right arrow keys.
| Study and date | Test scope | Measured outcome | Main limit |
|---|---|---|---|
| Liang et al., 2023 | Seven detectors, accessed March 2023. 91 human-written TOEFL essays and 88 US eighth-grade essays. | Average false positive rate of 61.22% across the seven detectors on the TOEFL essays. 89 of 91 essays were flagged by at least one detector, and 18 of 91 by all seven. | One writer group and small samples. The rate is an average across detectors, not a rate for any one of them. Tools have changed since. |
| Dugan et al., RAID, ACL 2024 | 12 detectors (8 open-source, 4 closed-source). | The authors report that current detectors are easily fooled by adversarial attacks and by variations in how text is generated. | Benchmark conditions, not your documents. Per-detector figures are in the paper’s tables, and we do not restate them. |
| Weber-Wulff et al., published December 2023 | 14 tools (12 public systems plus Turnitin and PlagiarismCheck), 54 test cases each, 756 tests in total. | The authors conclude the tools were neither accurate nor reliable overall, and that machine paraphrasing reduced performance. | Small case set, 2023 versions. Study-wide findings are not rates for any one tool. |
What the studies do and do not tell you
They do show that detector errors are documented, that false positives can fall unevenly on some writers, and that edits and generation settings can change results.
They do not show which commercial detector is best now, how any tool performs on your own documents, or how a detector updated after the study would score. None of the three is a test of AIHT, and AIHT’s ratings of humanizers are not detector accuracy measures. The three studies summarized here do not test all seven products in their current versions. The current product figures above therefore remain vendor-attributed in this guide.
What can change a result
- Length. Turnitin describes a minimum of 300 words in evaluated sentences, depending on language. GPTZero’s FAQ says accuracy improves with longer inputs.
- Writing type. Turnitin says it analyzes prose sentences and does not reliably detect lists, code or short unconventional writing.
- Editing and paraphrasing. RAID and Weber-Wulff both report that transformed text reduces detector performance in their tests.
- Writing tools. Copyleaks’s detector page notes that grammar checkers built on generative models can be flagged as potential AI content.
- Version changes. Turnitin says a submitted document’s AI percentage changes only if the document is resubmitted for processing, and that it updated its English AI writing detection model architecture in October 2026.
A checklist for judging detector evidence
- Who ran the test? Independent researchers or the vendor?
- What texts were used? Number of human and AI texts, topics, length, and who wrote the human ones (including non-native writers).
- How was the AI text made and edited? Untouched output behaves differently from edited or paraphrased text.
- Which errors are reported? Look for the false positive and false negative rates separately, not accuracy alone.
- What cutoff was used? A different threshold gives a different trade-off.
- Which version and date? Detectors update, so ask when the tested version ran.
- Is it an aggregate? An average across tools is not a result for one tool.
- Does it resemble your use? Match the test to your length, subject and writers before you rely on it.
- What happens after a flag? Decide in advance how you will review a flagged text without treating the score as proof.
If a detector flags your own writing
A flag is a reason to look closer, not a verdict. Keep your drafts and notes, and read our guide on why an AI detector may flag human writing. For how detectors score text in general, see how AI detectors work, and for why one text can score differently on different runs, why AI detector scores change.
Sources
- Liang et al., GPT detectors are biased against non-native English writers (2023)
- Dugan et al., RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors (ACL 2024)
- Weber-Wulff et al., Testing of detection tools for AI-generated text (2023)
- Turnitin, AI writing detection capabilities FAQs (vendor; page showed “Updated October 14, 2026”)
- GPTZero homepage and FAQ (vendor; checked October 14, 2026)
- Originality.ai AI checker page (vendor; checked October 14, 2026)
- Copyleaks AI detector page (vendor; checked October 14, 2026)
- ZeroGPT homepage (vendor; checked October 14, 2026)
- Grammarly AI detector page (vendor; checked October 14, 2026)
- Proofademic AI detection research page (vendor; checked October 14, 2026)
Frequently asked questions
What is the most accurate AI detector in 2026?
The three studies reviewed in this guide do not identify a current most-accurate detector. Their texts, detector versions and evaluation dates differ, so their findings do not establish a present-day ranking, and the product figures we found are vendor-reported.
Which AI detector should I use to check my own writing?
Several vendors offer a free check with highlighted passages, and the table above lists their stated free limits. Use the result as a prompt to review your draft, not as a prediction of what another tool will show.
Is a 98% or 99% accuracy claim believable?
It may be true on the vendor’s own test set, but without the data, the human-to-AI mix and the false positive rate you cannot tell what it means for your text.
Should I check a text with two detectors?
A second result is not independent proof. In Liang et al., 18 of 91 human-written essays were flagged by all seven detectors tested, so agreement between tools can still be wrong.
Do detectors catch edited or humanized text?
In the RAID and Weber-Wulff tests, transformed text reduced detector performance. Turnitin says it also looks for text modified by paraphrasing or bypasser tools in English submissions and does not name which tools. The studies summarized here do not verify that current product claim.