7 Best AI Writing Checkers: What They Catch and Miss

An AI detector reports the probability that a passage was machine-written, which leaves the verdict to you. Independent research supports that framing in both directions.

A 2023 study of 14 tools found none scoring above 80% accuracy, while a 2025 University of Chicago Booth study found that the strongest commercial detectors wrongly flagged fewer than 1% of human-written passages, provided the passages were long enough and nobody had run the text through a "humanizer" tool.

Detectors have improved, but the conditions under which they work reliably are narrower than their marketing suggests.

This guide compares seven widely used detectors: GPTZero, Pangram, Originality.ai, Turnitin, Copyleaks, Winston AI and Grammarly's AI detector.

It separates what independent researchers have measured from what vendors claim, shows where detectors tend to fail, and gives you a 15-minute method for testing any of them on your own writing before you rely on a score.

How an AI Detector Arrives at a Score

A language model writes by choosing a likely next word, so its output tends to be statistically smooth: common phrasing, steady sentence lengths and few surprises.

Most detectors look for that smoothness. Some measure perplexity, which is how predictable the words are to a model, and burstiness, which is how much sentence length and structure vary. Others train a classifier on large collections of human and AI text and report the probability that a passage belongs to the second group.

Our beginner's guide to how AI works, linked below, explains why next-word prediction leaves this kind of statistical trace. The design has a built-in weakness. The trace is a tendency, and no detector can read it as a mark.

Writing that happens to be predictable, such as a formulaic contract clause, a plain product description or an essay written in a second language with a limited vocabulary, can look machine-made, while AI text that has been edited, paraphrased or prompted to sound less uniform can look human.

Watermarking, where a model embeds a hidden statistical pattern as it generates, is a different approach. It only helps with text from a model that applies one, so it cannot check text from arbitrary sources.

Also Read: The AI Revolution: A Beginner's Guide to How AI Really Works

What Three Independent Studies Found

Vendor accuracy claims are hard to compare, because each company tests on its own data. Three studies by outside researchers give a firmer footing.

Weber-Wulff and colleagues, International Journal for Educational Integrity, 2023

• Tools tested: 14 tools, including Turnitin
• Main finding: All scored below 80% accuracy and only five above 70%; the tools leaned toward calling text human; accuracy fell on edited, paraphrased and machine-translated text
• Keep in mind: Tested 2023 versions on ChatGPT-era text; the tools have changed since

Liang and colleagues, Stanford, Patterns, 2023

• Tools tested: Seven detectors
• Main finding: On average, 61.3% of essays by non-native English writers were flagged as AI-written; US eighth-grade essays were handled far better
• Keep in mind: The detectors were not named, and the essays came from a single forum

Jabarian and Imas, University of Chicago Booth, 2025

• Tools tested: GPTZero, Originality.ai, Pangram and an open-source model
• Main finding: The three commercial tools kept false positives under 1% on medium and long text; the open-source model flagged 30% to 78% of human text; all lost accuracy on passages under 50 words
• Keep in mind: A working paper, not yet peer reviewed; only three commercial tools; text types were limited

The first study, published in 2023, tested 12 public tools plus Turnitin and PlagiarismCheck. Its authors concluded that the tools were neither accurate nor reliable, and noted that the errors ran mostly toward missing AI text, with false positives appearing in six of the 14 tools.

Turnitin was the only tool to classify all of the AI-generated documents correctly in one test, so even the best performer of that round sat in a weak category.

OpenAI offers a telling comparison: it released its own AI text classifier in early 2023 and retired it on July 20, 2023, citing a low rate of accuracy. At launch, OpenAI said the classifier correctly identified 26% of AI-written text, according to its announcement.

The Stanford study matters most for anyone judging other people's writing. The researchers ran seven detectors over 91 essays written for the TOEFL exam by non-native English speakers and 88 essays by US eighth graders.

The detectors flagged 61.3% of the TOEFL essays as AI-generated on average, all seven flagged 19.8% of them, and at least one detector flagged 97.8%, as the full paper in Patterns and its preprint report.

The authors link the pattern to perplexity: writers with a smaller range of vocabulary produce more predictable text, which reads to a detector as machine-like. They advise against using detectors in evaluative settings, particularly for non-native speakers.

The 2025 Chicago Booth working paper is the most favorable to detectors and the most careful about conditions. GPTZero, Originality.ai and Pangram all kept false positives below 1% on medium and long passages, and Pangram's were essentially zero.

The paper also tested "humanizers," tools that rewrite AI text to evade detection. Pangram stayed robust against a current humanizer, while GPTZero's false negative rate rose to around 50% or more across most genres, according to the University of Chicago's research summary.

The authors also point out a trade-off built into every detector: raising the threshold at which text counts as AI lowers false positives and raises false negatives, so the same tool can look strong or weak depending on its setting.

Read together, the studies say accuracy depends on four things: text length, the specific tool, the threshold and whether the text was modified. Any single accuracy figure that ignores those conditions is incomplete.

Seven Detectors, and How Much Evidence Stands Behind Each

The most useful way to compare these tools is by the quality of the evidence behind them. Features and limits change often, so check each provider's pricing page for current plans.

GPTZero

• Typical user: Educators and anyone wanting a free first check
• Independent evidence: Chicago Booth 2025: low false positives, including on short passages
• What to watch: Lost much of its ability to catch AI text after humanizer processing; the company has published its own re-run of the Chicago benchmark, which is a vendor result

Pangram

• Typical user: Publishers and editors who most want to avoid false accusations
• Independent evidence: Chicago Booth 2025: essentially zero false positives on medium and long text; robust to a current humanizer
• What to watch: Newer tool with one major independent study behind it, which is a working paper

Originality.ai

• Typical user: Publishers and content teams; also bundles plagiarism checking
• Independent evidence: Chicago Booth 2025: strengths in separating AI from human text at default settings
• What to watch: Higher false positive rate than GPTZero in the same study; results shift with the threshold

Turnitin

• Typical user: Schools and universities, through institutional access
• Independent evidence: Weber-Wulff 2023: strongest of the 14 tools on one test, within a weak category
• What to watch: Its own published figures are a document false positive rate under 1% for documents with 20% or more AI writing and about 4% at the sentence level

Copyleaks

• Typical user: Multilingual and enterprise users
• Independent evidence: None in the three studies above
• What to watch: Advertises over 99% accuracy and 30+ languages; those are vendor claims

Winston AI

• Typical user: Agencies and publishers who want documented reports
• Independent evidence: None in the three studies above
• What to watch: Advertises very high accuracy; no comparable independent evaluation was found

Grammarly AI detector

• Typical user: Quick free checks inside an existing Grammarly workflow
• Independent evidence: None in the three studies above
• What to watch: Grammarly itself says no AI detector can conclusively determine whether AI was used

The tools with independent evidence. GPTZero, Pangram and Originality.ai are the three the Chicago team tested, and the authors describe GPTZero and Originality.ai as a second tier behind Pangram, with partial strengths.

Their guidance on choosing between them is practical: if you mostly want to avoid false accusations, GPTZero's lower false positive rate favors it, and if you mostly want to catch AI text, Originality.ai's stronger separation favors it.

Pangram led on both measures in that study, but one study is thin evidence for a recommendation, and the text types it used were limited.

Turnitin deserves separate handling because most readers meet it through a school. Turnitin says that its sentence-level false positive rate is around 4%, that false positives are more common when less than 20% of a document is flagged (so it shows an asterisk in place of a score), and that most false-positive sentences sit next to real AI text, according to its own explanation.

Its instructor guidance says the result should not be the only basis for action against a student. Vanderbilt University announced in August 2023 that it would disable the feature, citing concerns about reliability.

Copyleaks, Winston AI and Grammarly are widely used, and there is a gap in the independent record for them. That doesn't mean they perform badly.

It means the three studies above did not test them, and their accuracy figures come from the companies. Grammarly says plainly on its own detector page that no tool can be certain.

Also Read: AI Tools That Are Actually Changing How Professionals Work

Where Detectors Fail: Short Text, Humanizers and Non-Native Writers

The same weak points appear across the studies.

• Very short text. All three commercial tools in the Chicago study lost accuracy on passages under 50 words. A product description, a caption or a short answer gives a detector little to work with.
• Edited or paraphrased AI text. The 2023 study found accuracy dropped once AI output was edited by hand, paraphrased or machine-translated, which is how much real-world AI text ends up.
• Humanizer tools. Rewriting tools designed to evade detection can defeat some detectors, and the Chicago results show wide differences between tools.
• Non-native English writing. The Stanford finding is the clearest case of a systematic false positive. It applies to any setting where someone's writing is being judged, including hiring, school and client work. Our guide to the ethics and risks of AI in the workplace, linked below, covers the broader problem of automated tools that disadvantage groups of people.
• Formulaic human writing. Legal clauses, technical manuals and templated emails are predictable by design, which is the pattern detectors read as machine-made.

Also Read: The Ethics and Risks of AI in the Workplace: What Every Business Needs to Know

A 15-Minute Test You Can Run on Any Detector

Published studies cannot tell you how a detector handles your kind of writing. A small test can. It works as a sanity check, since 15 samples cannot produce a reliable accuracy rate, yet it can reveal whether a tool is wrong in ways that matter to you.

1. Gather five human samples you know are human. Use your own older writing from before 2022, a formal document typical of your field and, with permission, a colleague's writing if they write in a second language.
2. Generate five AI samples. Use your usual AI tool and leave the output unedited.
3. Make five hybrids. Edit two AI drafts lightly, edit one heavily and run two through a paraphrasing tool.
4. Run all 15 at the default setting and record the result for each in a table like the one below.
5. Count the errors. Human samples flagged as AI are false positives. AI samples passed as human are false negatives. Note what happens to the hybrids.

Record each sample on one line, for example:

• Sample 1 (human, 2021 article): Score ___, Verdict ___, Correct? ___
• Sample 6 (raw AI draft): Score ___, Verdict ___, Correct? ___
• Sample 11 (AI draft, lightly edited): Score ___, Verdict ___, Correct? ___

If even one of your five human samples comes back flagged, do not use the tool to judge other people's work. Keep samples above about 100 words, since short text is unreliable, and repeat the test when the vendor announces a model update.

Before you upload anything, check the provider's data terms. A confidential client draft pasted into a free detector may not stay private.

Who Should Use a Detector, and What to Check Instead

Bloggers and publishers. Google does not use a detector score. Its guidance says that appropriate use of AI or automation is not against its guidelines, and that using automation primarily to manipulate rankings is, as its guidance on AI-generated content states.

Its generative AI page adds that AI is useful for researching a topic and structuring original content, while generating many pages without adding value may violate its policy on scaled content abuse. The practical risk lies in how thin a page is, whoever wrote it.

A better review than a detector is a short checklist: are the facts verified against primary sources, does the page say something the top results do not, and would a reader finish it with their question answered? Our guides on E-E-A-T and on SEO fundamentals, linked below, cover what quality signals look like.

Teachers: Turnitin's own advice is to treat a score as a prompt for conversation. Assignments that produce process evidence, such as drafts, outlines and in-class writing, give a firmer basis than a percentage.

Clients and agencies commissioning writers: Put the expectation in the contract: whether AI use is allowed, whether it must be disclosed and what happens if a detector flags a delivery. Ask for outlines and version history, and talk to the writer before drawing a conclusion.

Hiring managers: Writing samples and cover letters from non-native English speakers carry the highest risk of false flags, so a detector score should not screen anyone out.

Also Read: E-E-A-T Explained: How to Build Topical Authority

If a Detector Flags Your Own Writing

Being flagged by mistake is a documented outcome, and a calm, factual response works best.

1. Ask what was flagged and by which tool. The tool, version, score and threshold all matter, and which passages triggered the flag.
2. Gather process evidence. Version history in Google Docs or Word, earlier drafts, notes, outlines and source material show how the text came together.
3. Ask for a human review and point to the published research on false positives, including the Stanford finding if you write in a second language.
4. Keep copies of the flagged document and the report.

Conclusion

AI detectors measure statistical patterns, and a pattern is evidence of nothing more than that. The best of them perform well under favorable conditions and poorly outside them, and for most of the popular tools, the independent record is thin or missing.

The responsible use is as one signal among several: run your own 15-minute test, never judge a person by a score alone, and keep your attention on the thing Google and readers both reward, which is a page that is accurate, original and useful.

Also Read: SEO Fundamentals: Complete Guide to Ranking Higher on Google