detection
Best AI Detector
Compare the best AI detectors — Pangram, Originality.ai, GPTZero, Copyleaks, Turnitin and Winston AI, with dated independent accuracy results, false-positive rates, pricing, a guide for teachers, and which to use for education, publishing or checking your own writing.
Quick answer: No AI detector is reliable enough to prove that a piece of text was written by AI, and the strongest independent research agrees — treat a detector score as a signal, never as evidence. For raw accuracy with the lowest false-positive rate, Pangram leads on both of the strongest independent studies: a University of Chicago Booth study (August 2025) found it was the only detector to keep false positives at or below 0.5% while still catching AI text reliably (Chicago Booth Review), and in a peer-reviewed June 2026 study of academic papers its average AI score on fully AI-written papers was 85.7%, against 14.3% or less for Turnitin, GPTZero and Copyleaks (Van Vlasselaer et al.). For publishers and SEO teams checking edited or paraphrased content, Originality.ai had the highest accuracy of the commercial detectors in the independent RAID benchmark, 85.0% against 71.0% for Winston AI and 66.5% for GPTZero, and 96.7% on paraphrased text (RAID). For a free check, Pangram’s free plan covers 2,000 words a day, and GPTZero runs a basic scan with no account. Two caveats matter most: a few passes through a “humaniser” tool defeat almost all detectors, and the 2023 evidence that detectors flag non-native English writing has been only partly offset by 2026 testing that found no such false positives.
This guide ranks the detectors that matter — Pangram, Originality.ai, Winston AI, GPTZero, Copyleaks, Turnitin and the best free options — with vendor claims and independent test results kept strictly separate. Two shifts define the moment: the peer-reviewed evidence has caught up with the marketing (and it is unflattering), and the field is moving away from a single ”% AI” verdict towards process transparency, policy rules and content watermarking. If you are choosing a detector to make decisions about students, writers or published content, read the false-positive section before you read the rankings. For individual tools, see our guides to Turnitin AI detection, GPTZero and the Grammarly AI detector; for the mechanism, how AI detectors work.
What is the most accurate AI detector?
Pangram, on the independent evidence available. It is the only detector that met the University of Chicago Booth study’s strict false-positive cap (0.5% or lower) while still catching AI text, and in the peer-reviewed June 2026 study of 160 academic papers its average AI score on fully AI-written papers was 85.7%, against 14.3% for GPTZero, 14.0% for Copyleaks and 0.0% for Turnitin; on hybrid papers it averaged 45.5% against an actual 46.3%. The authors report that all four tools correctly identified the human-written papers (Chicago Booth Review, IJEI). For paraphrased text specifically, Originality.ai has the strongest result in the RAID benchmark (96.7%). Every vendor, Pangram included, advertises 99% or more; those figures are measured on clean text and are a ceiling.
Which one should you pick?
| You are… | Pick | Why |
|---|---|---|
| A teacher or institution making decisions about students | Pangram, alongside process evidence | Lowest false positives and best detection in independent tests; see AI detectors for teachers |
| A student checking your own work | Pangram’s free plan or GPTZero’s free checker | Free self-checks; neither predicts what Turnitin will say |
| A publisher or SEO team vetting edited copy | Originality.ai | Best paraphrase robustness, bulk scanning and API |
| An enterprise needing many languages | Copyleaks | 30+ languages, plagiarism and AI in one platform |
| Already on Turnitin | Turnitin, as one input among several | It is what your institution runs; see its limits |
| Trying to prove you wrote something | Not a detector | Version history, drafts and a process tracker are stronger evidence |
The current state of AI detection
AI-writing detection has become a multi-hundred-million-dollar market built on a claim its own field cannot fully stand behind: that a statistical model can tell human writing from machine writing accurately enough to act on. It sometimes can. Often it cannot. Five shifts define where things stand.
1. Institutional trust has collapsed — and universities are pulling back. Curtin University in Australia disabled Turnitin’s AI writing detection from 1 January 2026 (Curtin); the University of Cape Town stopped using detection tools such as Turnitin’s AI Score from 1 October 2025, with the director of its teaching centre saying “AI detectors are simply not reliable” (UCT); and the University of Queensland made Turnitin’s AI Writing Indicator unavailable from Semester 2 2025 (UQ). The direction of travel among the institutions that know the tools best is towards trust-based assessment, not more detection.
2. Independent science caught up, and it is unflattering. The rigorous, non-commercial studies now exist. The most recent is a peer-reviewed study by Van Vlasselaer, Van Droogenbroeck and Spruyt in the International Journal for Educational Integrity (June 2026), which found Turnitin, GPTZero and Copyleaks gave fully AI-written academic papers average AI scores of 14.3% or less, while the authors report every tool correctly identified the human papers (IJEI). A University of Chicago Booth paper (“Artificial Writing and Automated Detection”, August 2025) tested detectors against a strict policy cap and found only one — Pangram — could hold false positives at or below 0.5% while still detecting AI text reliably (Chicago Booth Review). The RAID benchmark — 6 million-plus generations across 11 models, 8 domains and 11 adversarial attacks — found that current detectors are “easily fooled” by paraphrasing, sampling changes and unseen models (RAID).
3. The humaniser arms race tilted towards evasion. “Humaniser” tools that rewrite AI text to beat detectors are now cheap and effective. Independent testing found that after roughly three passes through a quality humaniser, no detector consistently identified the content as AI, with GPTZero’s detection rate falling to around 18% on humanised text (GradPilot). Bypass tools such as Undetectable.ai report 85–88% success against major detectors (Phrasly). Detection can win a round, but it is structurally on the back foot.
4. Watermarking emerged as the “honest” alternative — and in August 2026 it jumped labs. Google’s SynthID Detector portal — opened to early testers at I/O 2025 (20 May 2025) and expanded since — scans images, audio, video and text for an embedded watermark (Google). And on 11 August 2026, Anthropic became the first frontier lab to watermark text at production scale: Claude output now carries an invisible statistical mark worldwide, with C2PA metadata on generated files, driven by EU AI Act Article 50 (enforceable 2 August 2026) — see our AI watermarking guide for how it works. The limitation remains fundamental: watermark detection only works on content from models that add the watermark, and Anthropic itself stresses a mark proves content was processed by Claude, not authored by it. Text from ChatGPT still carries no watermark — OpenAI has never shipped a public text watermark (SynthID docs) — and open-weight models never will.
5. Detection went segment-level and policy-based. The better tools no longer return a single “80% AI” number. Pangram 3.0 (December 2025) added segment-level analysis and a four-tier classification that separates “lightly AI-assisted” from “fully AI-generated”, and Pangram 4 (29 July 2026) targeted mixed human-AI text and humaniser output (Pangram). GPTZero’s 4o model (24 September 2026) added AI Patterns, which names the stylistic features behind a score (GPTZero). Turnitin’s Clarity add-on shifted towards policy profiles — “No AI”, “disclosure required”, “outline only” — plus draft-timeline version history that shows how a document was written rather than only guessing whether AI touched it (Turnitin Clarity).
How AI detectors actually work
AI detectors do not “see” a hidden fingerprint in the way a plagiarism checker matches copied strings. Most predict how likely a language model is to have produced the text, using two core signals:
- Perplexity measures how predictable each word is given the words around it. Language models are trained to choose high-probability words, so machine text tends to have low, even perplexity. Human writing is spikier and less predictable.
- Burstiness measures variation in sentence length and structure. Humans mix long and short sentences; models often produce steadier, more uniform prose.
Modern detectors layer a trained classifier on top of these signals — a model taught on large sets of human and AI text to output a probability. This design explains both their strengths and their central flaw. They are good at spotting unedited, default-settings output from a known model. They are weak wherever text sits outside their training distribution: heavily edited or paraphrased writing, output from a model they have not seen, and — critically — human writing that happens to be low-perplexity, such as formulaic technical prose or the simpler vocabulary of a non-native English speaker. A detector cannot tell the difference between “predictable because a machine wrote it” and “predictable because the human writer used common words”. That single ambiguity is the source of nearly every false accusation.
Top AI detectors compared
Two numbers exist for every detector, and they disagree, so we show both. Vendor accuracy is measured on clean, unedited AI text under controlled conditions and is effectively a ceiling. Independent accuracy is measured on edited, paraphrased, mixed and adversarial text — closer to what you will actually paste in.
What the vendors claim
| Detector | Claimed accuracy | Claimed false-positive rate |
|---|---|---|
| Pangram | 99.8%+ | Under 1% |
| Winston AI | 99.98% | Low |
| Copyleaks | 99%+ | Under 1% |
| Originality.ai | ~99% | Under 1% |
| GPTZero | 99%+ | Low |
| Turnitin | 97% (2023 launch figure) | Under 1% (at 20%+ AI) |
What independent testing finds
| Detector | Independent finding | Date of test | Source |
|---|---|---|---|
| Pangram | 65% on fully AI papers (97.5% counting borderline scores); 92.5% on hybrid and humanised; all human papers correct | Published June 2026 | IJEI |
| Pangram | Only detector to meet a false-positive cap of ≤0.5% while still detecting AI reliably | August 2025 | UChicago Booth |
| Turnitin | 0% on fully AI papers; 60.0% hybrid; 50.0% humanised; all human papers correct | Published June 2026 | IJEI |
| GPTZero | 0% on fully AI papers; 0% hybrid; 2.5% humanised; all human papers correct | Published June 2026 | IJEI |
| Copyleaks | 0% on fully AI papers; 30.0% hybrid; 22.5% humanised; all human papers correct | Published June 2026 | IJEI |
| Originality.ai | 85.0% at a 5% false-positive rate, against 71.0% for Winston AI and 66.5% for GPTZero; 96.7% on paraphrased text, under 10% on character-level attacks | Published 2024 | RAID |
| Copyleaks | One of three detectors, with Turnitin and Originality.ai, found to have high accuracy on 126 essays | Published 2023 | Open Information Science |
| Winston AI | First of 55 on RAID: 99.89% at a 1% false-positive rate, attacks included; run by Winston, scored by RAID | Submitted September 2026 | RAID |
| Winston AI | 67% accuracy, fifth of 14 tools; no human texts flagged | Tested March 2023, published December 2023 | IJEI |
The June 2026 study’s figures are the share of papers each tool classified correctly. Its AI texts were generated before publication, and vendors have shipped model updates since — Turnitin several times, Pangram 4 in July and GPTZero 4o in September 2026 — so treat the dates as part of the result.
Why the two tables disagree — and which to trust. Every vendor advertises a 98–99%+ number, and every vendor’s number is real under its own test conditions: clean AI text, one known model, no editing. Independent tests break all three assumptions and accuracy falls, sometimes by 20–30 points. Treat vendor claims as a ceiling and independent figures as the floor; your real result sits between them and depends on how the text was written and edited. One more caveat on sources: most “best AI detector” rankings online are published by detector or humaniser companies with a commercial stake in the result — Scribbr, for example, sells its own detector and ranks it first. Where possible we anchor on peer-reviewed and academic work (the 2026 IJEI study, UChicago Booth, Stanford, the RAID benchmark) rather than vendor-run comparisons.
The best AI detectors, reviewed
We have ordered these by overall standing for the most common job — deciding whether text is AI-written without wrongly accusing a human. That weighting puts false-positive control first.
1. Pangram — most accurate, lowest false positives
Price: Free (2,000 words a day); Individual $20/month (300,000 words); Professional $65/month (1.5 million words plus API credit); API $0.05 per 100 words (Pangram) Access: Web dashboard, Chrome extension, Google Docs, API, LMS integrations (Canvas, Moodle, Google Classroom, Brightspace) Best model coverage: Multilingual (20+ languages), segment-level, four-tier classification
Pangram is the detector the independent evidence supports most clearly. In the University of Chicago Booth study, it was the only tool to satisfy a stringent false-positive cap (≤0.5%) without losing the ability to detect AI text — recording near-zero false positives on long and medium passages and never above 1% on short ones (Chicago Booth Review). In the peer-reviewed June 2026 study its average AI score on fully AI-written academic papers was 85.7%, far above the other three tools, and on hybrid papers it averaged 45.5% against an actual 46.3% (IJEI). Its 3.0 release added segment-level analysis and a four-tier scale that distinguishes lightly AI-assisted work from fully AI-generated work, and Pangram 4 (29 July 2026) targets mixed text and humaniser output.
Why it wins: The lowest false-positive rate in independent testing, which is the single most important property if a wrong result has real consequences for a person. The free tier and Chrome extension make it easy to trial.
Limitations: The free plan is capped at 2,000 words a day; heavy or institutional use needs a paid plan. It is newer and less known than Turnitin inside universities, so buy-in can lag the evidence.
Best for: Educators, admissions teams and anyone making high-stakes decisions where a false accusation is unacceptable.
2. Originality.ai — best for publishers, SEO and edited text
Price: Pro $14.95/month (2,000 credits); Enterprise $179/month (15,000 credits + API); pay-as-you-go $30 for 3,000 credits Access: Web app, Chrome extension, API, team dashboards, bulk scanning Notable: Combines AI detection, plagiarism, fact-checking and readability in one workflow
Originality.ai is built for content teams rather than classrooms, and it had the highest accuracy of the four commercial detectors in the RAID benchmark, 85.0% at a 5% false-positive rate against 71.0% for Winston AI and 66.5% for GPTZero (RAID). Under paraphrase attacks it scored 96.7%, the best of any detector tested, but character-level tricks such as homoglyphs and zero-width spaces cut it to under 10%. Originality.ai markets accuracy of around 99%.
Why it wins: The highest accuracy of the commercial detectors in RAID and the best result under paraphrasing, plus bulk scanning and an API that suit publishing pipelines and agencies vetting freelance work.
Limitations: No genuine free tier — only ~50 credits at signup. Character-level tricks such as homoglyphs defeated it in RAID. Remember that Google does not penalise content for being AI-made — it judges quality — so treat Originality.ai as an editorial-standards tool, not an SEO safety net.
Best for: Publishers, marketing teams, SEO agencies and editors enforcing a human-writing standard at scale.
3. Winston AI — premium reporting for publishers and educators
Price: Essential ~$18/month (80,000 words); Advanced ~$29/month (200,000 words); Business per-seat above Access: Web app, shareable PDF reports, plagiarism, OCR, AI-image and deepfake detection Notable: Professional reporting and document handling
Winston AI markets a 99.98% accuracy figure. In September 2026 it ranked first of 55 detectors on the RAID leaderboard, at 99.89% with a 1% false-positive rate; Winston ran that test itself and RAID scored it, on AI text from older models. In a 2023 peer-reviewed test of 14 tools it placed fifth at 67% accuracy, with no human texts wrongly flagged (IJEI).
Why it matters: Polished PDF reporting, OCR, plagiarism and image/deepfake detection make it a tidy all-in-one for publishers and educators who want documentation.
Limitations: Its only peer-reviewed result dates from 2023, and its RAID result covers older models and includes no Claude output.
Best for: Publishers and educators who value reporting and document handling and will corroborate every flag.
4. GPTZero — best free tier, education favourite
Price: Free basic checker (no account); Premium (300,000 words/month) and Professional (500,000 words/month) — GPTZero’s pricing page shows prices in the visitor’s local currency, so check the live page; Team, Enterprise and API tiers for institutions Access: Web app, Chrome extension, Canvas / Google Classroom / Moodle integrations, API Notable: “Paraphraser Shield” tuned against humaniser tools; Writing Replay; GPTZero 4o model since 24 September 2026
GPTZero, founded by Edward Tian, is a detector built for teaching. Its basic checker needs no account. GPTZero claims 99% accuracy, and its 4o model claims one false positive in 10,402 student essays (GPTZero); independently, in the June 2026 peer-reviewed study the authors report it correctly identified the human papers, while its average AI score on fully AI-written papers was 14.3% (IJEI). Superhuman, the company formerly named Grammarly, announced a deal to acquire GPTZero on 23 June 2026, and said on 30 September 2026 that GPTZero “is now part of Superhuman” (Superhuman). For how its score works and what independent testing found, see our GPTZero guide.
Why it ranks here: A free no-account checker, strong LMS integrations and low false positives in the 2026 study — offset by the weakest detection of AI-written papers in that same study.
Limitations: GPTZero’s own FAQ says it is not trained on heavily modified AI text and that its results “should not be used to punish students”. Independent results measure models before 4o.
Best for: Teachers, students self-checking, and anyone who wants a free, LMS-integrated detector for a first-pass signal.
5. Copyleaks — best for enterprise and multilingual
Price: Free scan up to 25,000 characters without an account, with monthly credits on a free account (Copyleaks); Essential ~$10/month (annual); Advanced ~$16/month (annual); enterprise and API pricing above Access: Web app, browser extension, LMS and API integrations, plagiarism + AI in one scan Notable: 30+ languages; SOC 2 and enterprise compliance
Copyleaks pairs AI detection with a mature plagiarism engine and is the natural pick where compliance, languages and volume matter. In a 2023 peer-reviewed study of 126 essays, it was one of three detectors, with Turnitin and Originality.ai, found to have high accuracy (Open Information Science). Copyleaks advertises accuracy above 99%. In the peer-reviewed June 2026 study the authors report it correctly identified every human paper, and its average AI score on fully AI-written papers was 14.0% (IJEI).
Why it wins: Strong multilingual coverage, low false positives, enterprise compliance, and combined plagiarism + AI detection in a single tool.
Limitations: In the June 2026 study it classified none of the fully AI-written papers as AI, and 30.0% of hybrid and 22.5% of humanised papers correctly (IJEI). The interface is built for organisations more than individuals.
Best for: Enterprises, publishers and institutions needing multilingual detection, plagiarism checking and compliance in one platform.
6. Turnitin — the education incumbent (institution-only)
Price: Institution licence only, sold as an add-on to Turnitin Feedback Studio / Originality; no consumer plan Access: Embedded in university and school LMS workflows; not available to individuals Notable: Clarity add-on adds policy profiles and draft-timeline version history
Turnitin is the detector most students will actually be measured by, because it is embedded across thousands of institutions. Turnitin says its false-positive rate is below 1% for documents with more than 20% AI writing (Turnitin), and at its 2023 launch said it identified 97% of ChatGPT and GPT-3 writing (Turnitin). Independent testing tells a more cautious story: in the peer-reviewed June 2026 study it classified every human paper correctly but reported 0% on all 40 fully AI-written papers, with 60.0% on hybrid and 50.0% on humanised papers (IJEI). Turnitin itself states its detection “may not always be accurate” and “should not be used as the sole basis for adverse actions against a student” (Turnitin). Our Turnitin AI detection guide covers the score, the thresholds and the case law.
Why it matters: Ubiquity and LMS integration, and a 2026 shift with Clarity towards writing-process transparency — version history and policy rules — rather than a single detection score.
Limitations: Institution-only, so you cannot pre-check your own work against it. Its false-positive profile on non-native writing has driven several universities to switch it off (false-positive cases).
Best for: Institutions already inside the Turnitin ecosystem — used as one input among several, never as proof.
7. Free detectors — Scribbr, QuillBot and ZeroGPT
For a quick, zero-cost check, three free tools are widely used. Scribbr and QuillBot share the same detection engine, and Scribbr says its free detector identified 78% of texts correctly in a comparison Scribbr ran itself (Scribbr). ZeroGPT needs no sign-up; in the RAID benchmark it scored 65.5% and could not meet the benchmark’s 5% false-positive target (RAID). None of the three was tested in the peer-reviewed studies above.
Quetext, which we cover as a plagiarism checker, also sells an AI detector and offers free scans of up to 1,000 words. Quetext says it is close to 99% accurate on AI content; we found no independent test of it.
Feature comparison: the full matrix
| Feature | Pangram | Originality.ai | GPTZero | Copyleaks | Turnitin | Winston AI |
|---|---|---|---|---|---|---|
| Free tier | Yes (2,000 words/day) | No (signup credits) | Yes (basic checker) | Yes (25,000 characters a scan) | No (institution) | Trial |
| Lowest paid | $20/mo | $14.95/mo | See live page | $10/mo | Licence | $18/mo |
| Independent false-positive rate | 0% (IJEI 2026) | Data not available | 0% (IJEI 2026) | 0% (IJEI 2026) | 0% (IJEI 2026) | 0% (IJEI 2023) |
| Accuracy on paraphrased or humanised text | 92.5% (IJEI 2026) | 96.7% (RAID 2024) | 2.5% (IJEI 2026) | 22.5% (IJEI 2026) | 50.0% (IJEI 2026) | 52.6% (RAID 2024) |
| Plagiarism check | No | Yes | Add-on | Yes | Yes | Yes |
| Multilingual | 20+ langs | Yes | Partial | 30+ langs | Yes | Yes |
| LMS integration | Yes | No | Yes | Yes | Native | No |
| API | Yes | Yes | Yes | Yes | Enterprise | Yes |
| Individual access | Yes | Yes | Yes | Yes | No | Yes |
| Best-known for | Low false positives | Publishing/SEO | Free + education | Enterprise/multilingual | LMS ubiquity | Reporting |
Which AI detector should you use?
Best overall accuracy
Winner: Pangram (free tier; $20/month)
The only detector to meet a strict false-positive cap in independent academic testing while still catching AI text, and the highest-scoring of four tools on fully AI-written papers in the 2026 peer-reviewed study. Alternative: Originality.ai for paraphrased publishing copy.
Best for educators and schools
Winner: Pangram
The best independent results on both false positives and detection, with Canvas, Moodle, Google Classroom and Brightspace integrations. GPTZero is the alternative for its free checker and LMS integrations. See AI detectors for teachers below.
Best free AI detector
Winner: Pangram’s free plan (2,000 words a day)
The tool with the strongest independent results also has a usable free plan. GPTZero’s basic checker needs no account; Copyleaks’ free 25,000-character scan and Scribbr and QuillBot (78% in Scribbr’s own test) are backups. ZeroGPT is convenient but weaker. For every free checker ranked, with its limits, see best free AI detector.
Best for publishers, SEO and content teams
Winner: Originality.ai ($14.95/month)
Best robustness to paraphrasing plus bulk scanning, plagiarism and an API. But remember the strategic point: Google does not penalise AI content by origin — it rewards quality and E-E-A-T — so use a detector to enforce your own editorial standard, not to chase a non-existent ranking penalty.
Best for enterprise and multilingual
Winner: Copyleaks ($13.99/month billed annually and up)
Low false positives, 30-plus languages, plagiarism plus AI in one compliant platform. Alternative: Originality.ai Enterprise ($179/month) for content-heavy operations.
Best for checking your own writing before you submit
Winner: Pangram or GPTZero free tiers
Run your own draft to see whether human writing is being wrongly flagged — a common problem for non-native English writers and for clean, formulaic prose. If your genuine work is flagged, keep your drafts, notes and version history as evidence. It is not “cheating” to check your own writing; it is protecting yourself against an unreliable tool. Grammarly also offers a free checker, but by its own documentation it is tuned to miss AI text rather than risk false positives — see our Grammarly AI detector guide.
Best for avoiding a false accusation
Winner: Pangram — corroborated, never alone
No detector should trigger a penalty on its own. Pangram’s near-zero false-positive rate makes it the safest single input, but the correct process is always detector signal plus human judgement plus process evidence.
AI detectors for teachers
Teachers have fewer choices than the market suggests, because the detector that matters is often the one the institution already licenses.
What you can run. If your institution licenses Turnitin, its AI writing indicator appears inside the Similarity Report; students cannot see it. Outside that, Pangram and GPTZero both sell classroom and institutional plans with LMS integrations (Canvas, Moodle, Google Classroom and others).
What the evidence says. In the June 2026 peer-reviewed study of long academic papers, the authors report all four tools correctly identified human-written papers, but on fully AI-written papers Pangram’s average AI score was 85.7% against 14.3% or less for Turnitin, GPTZero and Copyleaks. The study’s authors conclude that detection tools “can provide useful initial flags” but “should not be used as sole evidence in high-stakes decision-making” (VUB).
What the vendors say. Turnitin says its detection “should not be used as the sole basis for adverse actions against a student”, and GPTZero’s FAQ says its results “should not be used to punish students”. Grammarly’s detector is a writer-facing self-check with no instructor dashboard (Grammarly guide).
What to use alongside a score. Process records — Turnitin Clarity, Grammarly Authorship, GPTZero’s Writing Replay, or ordinary document version history — show how a piece was written. They only exist if students use them from the start, so they belong in the assignment brief. For classroom tools beyond detection, see best AI for teachers.
The false-positive problem: why detectors can’t be proof
This is the section that should shape how you use every tool above. AI detectors do not just occasionally miss AI text; they occasionally flag genuine human writing as AI, and they do so unevenly.
The landmark evidence is a 2023 Stanford study (published in the journal Patterns) that ran essays through seven detectors. It classified 61.22% of TOEFL essays written by non-native English speakers as AI-generated, while judging essays by US-born eighth-graders almost perfectly (Stanford HAI). Across the sample, 97% of the non-native essays were flagged by at least one detector and 19% were flagged unanimously. The mechanism is the perplexity problem described earlier: non-native writers use more common words and simpler structures, which reads as “predictable”, which reads as “machine”. The bias is baked into how the tools work.
Newer evidence is more reassuring on this point. The June 2026 peer-reviewed study used human papers written by non-native English-speaking graduate students, and all four tools it tested classified them correctly, with its authors describing false positives as “rare across all tools, suggesting improvement compared to earlier studies” (VUB). Older studies found much higher rates, and the risk compounds at scale: at a 10% false-positive rate, an instructor running 300 essays will wrongly flag around 30 innocent students. That maths — not any single dramatic case — is why even the vendors now caveat that their output is not proof. A detector can start a conversation; it cannot end one.
The humaniser arms race
If false positives are the risk to the honest, humanisers are the loophole for the dishonest. These tools rewrite AI output — swapping synonyms, restructuring sentences, adjusting rhythm — specifically to lower perplexity-based detection scores (we list them in our best AI humanizer guide).
They work more often than not. Independent testing found that after about three passes through a good humaniser, no detector reliably caught the text, and GPTZero’s detection fell to roughly 18% (GradPilot). Undetectable.ai reports 85–88% bypass rates across major detectors, with one test cutting an essay’s Turnitin AI score from 98% to about 18% (Phrasly). Detectors are fighting back — GPTZero’s Paraphraser Shield and Pangram’s paraphrase training close some of the gap — but this is an arms race the defender is structurally losing, because the attacker only has to win once per document. The practical consequence: a confident “not AI” result is weak evidence of human authorship, because the most determined AI users are precisely the ones who have laundered their text.
Watermarking: the alternative that sidesteps detection
Because statistical detection is beatable and biased, the industry’s more durable bet is watermarking — embedding an invisible, statistically detectable signal into AI output at generation time. Google’s SynthID is the leading effort. Google opened its SynthID Detector portal to early testers at I/O 2025 (20 May 2025) — it scans images, audio, video and text for the watermark and highlights which parts carry it (Google) — and has expanded access since, alongside open-sourcing its text implementation on Hugging Face (SynthID docs).
Watermarking has one decisive advantage over detection — near-zero false positives, because it looks for a known signal rather than guessing from style — and two decisive limits. First, it only works on content from models that add the watermark: SynthID covers Google’s Gemini, but text from ChatGPT is not watermarked, and OpenAI has never released a public text watermark. Second, text watermarks can be weakened by heavy editing or paraphrasing, and short snippets often return “uncertain”. The cautionary precedent is OpenAI’s own AI Text Classifier, launched in January 2023 and withdrawn just six months later in July 2023 for a “low rate of accuracy” — it correctly identified only about a quarter of AI text and mislabelled human writing, especially from non-native speakers (TechCrunch). Until watermarking is universal and robust, it complements detection rather than replacing it.
Pricing comparison: what you’ll actually pay
| Tool | Free tier | Entry paid | Higher tier | Model |
|---|---|---|---|---|
| Pangram | 2,000 words/day | $20/mo Individual | $65/mo Professional + API credit | Word allowance |
| Originality.ai | Signup credits only | $14.95/mo (2,000 credits) | $179/mo Enterprise + API | Credit-based |
| GPTZero | Basic checker, no account | Premium (300k words/mo) | Professional (500k words/mo), Team, API | Word allowance; local-currency pricing |
| Copyleaks | 25,000 characters a scan | ~$10/mo Essential | ~$16/mo Advanced + enterprise | Page/credit |
| Winston AI | Trial | ~$18/mo (80k words) | ~$29/mo (200k words) | Word/credit |
| Turnitin | None | Institution licence | Clarity add-on | Institutional |
| Scribbr / QuillBot | Yes | Scribbr premium | — | Free + premium |
Cost strategy: for individual or classroom use, Pangram’s free 2,000 words a day or GPTZero’s free checker cover most needs. For publishing at scale, Originality.ai’s credits or Copyleaks’ page-based plans are the value picks. Do not pay enterprise pricing for a number you then have to corroborate by hand anyway.
What educators and publishers actually think
Educators are losing faith faster than they are adopting. The clearest signal of 2026 is institutions switching detection off: Curtin, Cape Town and Queensland have each switched Turnitin’s AI detection off (Curtin; UCT; UQ). The concern is rarely that detection never works; it is that the false-positive cost falls hardest on the students least able to contest it.
The field is moving from detection to process. Turnitin’s Clarity and Pangram’s segment-level scoring both point the same way — towards showing how a document was written (drafts, timelines, revision history) rather than issuing a verdict on whether a machine touched it. Many educators now favour “AI-resilient” assessment — oral defences, in-class writing, process portfolios — over any scanner.
Publishers use detection as an editorial gate, not an SEO shield. The savvier content teams have internalised that Google judges quality, not origin, so they run detectors like Originality.ai to enforce a house standard and vet freelance work, not to avoid an imaginary ranking penalty.
Recent developments
GPTZero 4o (24 September 2026). GPTZero’s new model adds AI Patterns and Masked Detection and claims one false positive in 10,402 student essays (GPTZero).
Pangram 4 (29 July 2026). Aimed at mixed human-AI text and humaniser output, released alongside a research preview of image detection (Pangram).
The peer-reviewed study that changed the picture (June 2026). Van Vlasselaer, Van Droogenbroeck and Spruyt found Turnitin, GPTZero and Copyleaks gave fully AI-written academic papers average AI scores of 14.3% or less, against 85.7% for Pangram, and the authors report every tool correctly identified the human papers (IJEI).
Superhuman agrees to buy GPTZero (23 June 2026). The company formerly named Grammarly announced the deal; no completion had been announced by 27 September 2026 (BusinessWire).
Anthropic starts watermarking Claude text (11 August 2026). See our AI watermarking guide.
SynthID Detector opened and kept expanding (May 2025 onwards). Google’s watermark-detection portal reached early testers at I/O 2025, covering text, image, audio and video from participating models, and access has widened since — a step towards watermarking as the long-run answer (Google).
Pangram 3.0 ships segment-level detection (December 2025). A four-tier classification separating light AI assistance from full generation, well suited to the nuance education actually needs (Pangram).
University of Chicago Booth publishes the strongest independent test yet (August 2025). “Artificial Writing and Automated Detection” found only Pangram met a ≤0.5% false-positive cap while detecting reliably (Chicago Booth Review).
Universities keep switching detection off (2025–2026). Curtin disabled Turnitin’s AI writing detection from 1 January 2026 (Curtin); Cape Town stopped using AI detection tools from 1 October 2025 (UCT); Queensland made Turnitin’s AI Writing Indicator unavailable from Semester 2 2025 (UQ).
Turnitin Clarity reframes the category. The move to policy profiles and draft timelines signals that even the incumbent is shifting from ”% AI” verdicts towards writing-process transparency (Turnitin Clarity).
Frequently asked questions
What is the best AI detector in 2026?
For accuracy with the fewest wrong accusations, Pangram — it was the only detector to meet a strict false-positive cap (≤0.5%) in University of Chicago Booth testing while still catching AI text reliably, and the highest-scoring of four tools on fully AI-written papers in a June 2026 peer-reviewed study. For publishers and edited or paraphrased content, Originality.ai leads, with the highest accuracy of the commercial detectors in the independent RAID benchmark and 96.7% on paraphrased text. For free checks, Pangram’s free plan gives 2,000 words a day and GPTZero has a no-account checker. No detector is accurate enough to be treated as proof.
What is the most accurate AI detector?
Pangram, on current independent evidence. It met the Chicago Booth study’s ≤0.5% false-positive cap while still detecting AI text, and in the June 2026 International Journal for Educational Integrity study it averaged 85.7% AI on fully AI-written papers, where Turnitin, GPTZero and Copyleaks averaged 14.3% or less. For paraphrased text, Originality.ai has the best RAID result (96.7%). Vendor claims of 99% or more, Pangram’s included, are measured on clean text.
Are AI detectors accurate?
Sometimes, within limits. On clean, unedited AI text from a known model, the best detectors exceed 95%. On edited, paraphrased or mixed text — what people actually submit — independent testing shows accuracy falling to roughly 60–90% depending on the tool and test. Vendors advertise 98–99%+ figures measured under ideal conditions; treat those as a ceiling, not a real-world expectation.
Can AI detectors be wrong and flag human writing as AI?
Yes. A 2023 Stanford study found detectors flagged 61% of essays by non-native English speakers as AI-generated while judging native writing almost perfectly (Stanford HAI); a June 2026 peer-reviewed study found all four tools it tested classified non-native graduate students’ papers correctly. At a 10% false-positive rate, roughly one in ten innocent people is wrongly flagged — which is why detector output should never be the sole basis for a penalty.
What is the best free AI detector?
Pangram’s free plan covers 2,000 words a day and has the strongest independent results. GPTZero runs a basic check with no account. Scribbr and QuillBot (same engine) scored 78% in Scribbr’s own test, and Copyleaks scans up to 25,000 characters free without an account. All free tools degrade on edited or humanised text, so use them for a signal, not a verdict. For every free checker ranked, see best free AI detector.
Can Turnitin detect ChatGPT, Claude and Gemini?
Turnitin said at its 2023 launch that it identified 97% of ChatGPT and GPT-3 writing, and says its false-positive rate is below 1% on documents more than 20% AI-written. Independent results are lower: a June 2026 peer-reviewed study found it reported 0% on 40 fully AI-written academic papers produced with ChatGPT Deep Research, and Turnitin itself says its results should not be the sole basis for action against a student. It is also institution-only, so you cannot pre-check your work against it directly.
Can AI detectors detect paraphrased or “humanised” text?
Poorly. Independent testing found that after about three passes through a quality humaniser, no detector reliably identified the text, and GPTZero’s detection dropped to roughly 18% (GradPilot). Originality.ai scored 96.7% on RAID’s paraphrase set in 2024, and In September 2026 Winston AI ranked first on RAID’s leaderboard with attacks included, and GPTZero’s Paraphraser Shield helps, but a determined user with a humaniser can beat almost any detector. A clean “human” result is therefore weak proof of human authorship.
Do AI detectors work fairly for non-native English speakers?
The evidence is mixed. The 2023 Stanford study measured a 61% false-positive rate on non-native essays across seven detectors. The June 2026 peer-reviewed study, whose human papers were written by non-native English-speaking graduate students, found no misclassified human papers across Turnitin, GPTZero, Copyleaks and Pangram. If English is your second language, check your own work before submitting and keep drafts and version history as evidence in case you are wrongly flagged.
Is it cheating to check my own writing with an AI detector?
No. Running your own draft through a detector to see whether it might be wrongly flagged is a sensible precaution, especially for non-native English writers or anyone writing clean, formulaic prose. Detectors flag plenty of genuine human writing, so knowing your risk in advance — and keeping drafts, notes and revision history — protects you against an unreliable tool.
Does Google penalise AI content, and do I need a detector for SEO?
No. Google’s stated position is that it does not penalise content for being AI-generated; it rewards helpful, high-quality, original content demonstrating experience and expertise, regardless of how it was produced (Google Search Central; independent analysis). Use a detector like Originality.ai to enforce your own editorial standard or to vet freelance work — not to avoid a ranking penalty that does not exist.
Which AI detector do most universities use?
Turnitin is by far the most embedded in higher education, integrated into thousands of institutions’ assessment workflows. But its dominance is being questioned from the inside: MIT, Yale, Vanderbilt, Curtin and others have disabled or discouraged AI detection over false-positive concerns, and several are shifting towards process-based approaches such as Turnitin’s own Clarity or tools like Pangram with lower false-positive rates.
Can AI text be watermarked instead of detected?
Increasingly, yes — but only for some models. Google’s SynthID embeds an invisible watermark into Gemini output that its SynthID Detector (opened to early testers at I/O 2025) can identify with near-zero false positives (SynthID). The limits are real: it only works on content from models that add the watermark — ChatGPT text is not watermarked — and heavy editing can weaken it. Watermarking is the more durable long-term approach, but today it complements detection rather than replacing it.
Conclusion: how to choose
AI detection is genuinely useful and genuinely unreliable, and holding both ideas at once is the whole skill.
- Most accurate, lowest false positives: Pangram — the best independent results on both false positives and detection.
- Best for publishers and edited text: Originality.ai — top paraphrase robustness plus bulk scanning and an API.
- Best free tool: Pangram’s free plan (2,000 words a day); GPTZero for a no-account check; every free option in best free AI detector.
- Best for enterprise and multilingual: Copyleaks — low false positives across 30-plus languages.
- Education incumbent: Turnitin — ubiquitous, but institution-only and best used as one input among several.
- Reporting-heavy option: Winston AI — good documentation, and first on the RAID leaderboard in September 2026, though those results were run by Winston itself.
The single rule that should govern all of it: a detector produces a probability, not a verdict. Use the tool with the lowest false-positive rate you can, corroborate every consequential flag with human judgement and process evidence, and never let a percentage make a decision about a person on its own. For related reading, see our guides to the best AI for writing, AI for students, AI for essays and the wider guides hub.