REVIEWED November 20, 2026
The Best AI Detectors in 2026, Compared on Accuracy
An AI detector gives you a probability, not proof, and the gap between the best and worst tools is huge. We compare 16 detectors on what independent studies say about their accuracy and false positives, what you get for free, and how they handle student data, so a flag leads to a fair conversation instead of a false accusation.
Top rated in each category
- Most accurate: Pangram (4.3)
- Free checkers: Grammarly (3.8)
- For teachers: Revision History (3.6)
How we rate detectors
Which AI detector fits what you need?
Start with the job you need done. Click a pick to jump to its full review.
- Fewest false accusations: Pangram (4.3) – The strongest record in independent studies since 2025: almost no human writing flagged, and among the fewest misses.
- Free, no account: QuillBot (3.6) – Six free checks a day of up to 1,200 words, with a separate score for AI-refined text.
- Prove you wrote it yourself: Grammarly (3.8) – Authorship records what you typed, pasted or generated in Google Docs and Word, and you choose who sees the report.
- Replay a student's Google Doc: Revision History (3.6) – Free for up to 150 Docs a month: writing time, typed versus pasted text and a full replay of every edit.
- Inside Canvas or Classroom: Copyleaks (3.6) – AI and plagiarism checks in one report, inside Canvas, Moodle, Blackboard, D2L, Schoology, Sakai and Edsby.
- See why each sentence was flagged: Studera.ai (3.7) – Our own detector: a reason for every flagged sentence and a writing report, in 17 languages, free without an account.
Compare every AI detector
Each cell shows what the tool does, based on the maker's own pages, and the last columns show whether outside researchers have tested it and how it treats student data. Our score is in the first column.
| Detector | Rating | Free, no account | Flags sentences | Spots mixed or edited AI | Explains why | Shows how it was written | Plagiarism check | Upload files | Canvas or Classroom | Other languages | Independently tested | School-grade privacy |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Most accurate | ||||||||||||
| Pangram | 4.3 | Partly – Free plan needs a free account; no card required. | Yes – AI Segment Analysis labels each passage. | Yes – Human, AI-Assisted and AI-Generated labels, plus a humanizer score per segment. | Partly – Highlights phrases common in LLM output; no written reasoning per flag. | No – No typing replay or version-history report. | Partly – Similarity check is included in paid plans. | Yes – File upload and OCR for scanned documents on the free plan. | Yes – Google Classroom, Canvas, Schoology, Moodle, Blackboard and Brightspace through school licenses. | Yes – 24 languages including Spanish, Arabic, Chinese, Vietnamese and Hindi. | Yes – Top performer in Chicago Booth 2025, VUB 2026 and Epoch AI 2026. | Yes – FERPA, SOC 2 Type 2; text never used for training. |
| GPTZero | 4.2 | Yes – Up to 10,000 characters per scan without signing up. | Yes – Advanced Scan highlights sentences; free users get one Advanced Scan. | Yes – AI, Mixed and Human labels, plus an AI Paraphrased label (beta). | Yes – AI Patterns (10 patterns, Aug 2026) and AI Vocabulary explain flags. | Yes – Writing Report and Writing Replay through the Chrome extension. | Partly – Plagiarism Check is included from Premium up. | Yes – Free plan uploads up to 3 files per batch; paid plans up to 250. | Yes – Canvas and Google Classroom on Professional and enterprise plans. | Yes – English plus 25 languages including Spanish; a Premium feature. | Yes – Strong in Chicago Booth 2025, Epoch AI 2026 and RAID; weak on hybrid papers in VUB 2026. | Partly – FERPA, SOC 2 Type II, HECVAT; individual-account content may train AI models unless the user opts out. |
| Originality.ai | 3.7 | Yes – Free scans run without an account, behind a captcha. | Yes – Sentence highlights on every scan. | Yes – Likely AI (81–100%), Mixed (10–80%) and Likely Original (under 10%). | No – Shows scores and highlights but not why a passage was flagged. | Yes – Writer Replay in the Chrome and Firefox extensions and Google Docs. | Yes – Included, also in free credits (2 credits per 100 words). | Partly – .docx, .doc and .pdf upload on paid plans only. | Yes – Moodle integration and Google Classroom writing replay. | Yes – 31 languages including Spanish, detected automatically. | Partly – Good on paraphrased text in RAID 2024 and Walters 2023; higher miss and false-positive rates in Chicago Booth 2025 and Epoch AI 2026. | Partly – FERPA and COPPA terms for school customers; free-checker text may be used for training. |
| Copyleaks | 3.6 | Yes – Up to 25,000 characters per scan without signing up. | Yes – Sentence-by-sentence results with a free account. | Partly – Shows the AI share of mixed text and paraphrasing signals, but no separate AI-assisted label. | Yes – AI Logic lists AI-typical phrases and matching AI sources. | No – No typing replay or pasted-versus-typed report. | Partly – AI and plagiarism in one report on paid and school plans. | Yes – Upload files or paste text, including several files at once on paid plans. | Yes – Canvas, Google Classroom, Moodle, Brightspace, Schoology, Blackboard, Sakai and Edsby. | Yes – 30 languages including Spanish. | Partly – High accuracy in Walters 2023; missed most AI and hybrid papers in VUB 2026. | Partly – FERPA, SOC 2 and SOC 3, GDPR; trains on content unless the customer opts out. |
| Winston AI | 3.3 | Partly – 14-day trial requires a free account. | Yes – AI Prediction Map colors each sentence. | Partly – One Human Score plus paraphrase detection; no separate mixed or AI-assisted label. | Partly – Explain button and a list of the sentences driving the score. | No – No typing replay or version-history report. | Yes – Included, at 2 credits per word. | Yes – Documents, photos and handwriting through OCR, up to 50 files. | Partly – Google Classroom only. | Yes – 23 languages including Spanish. | Partly – Leads the RAID leaderboard (vendor-submitted); early version below 80% accuracy with all tools in Weber-Wulff 2023. | Partly – SOC 2 Type 2, GDPR, no training without opt-in; no FERPA or COPPA statement. |
| Free checkers | ||||||||||||
| Grammarly | 3.8 | Yes – Scan up to 1,400 words without signing in. | Partly – Pro's AI Detector agent marks flagged sections; the free web scan shows only a percentage. | No – Splits text into "resembles AI" and "no AI patterns found". | Partly – Pro's AI Detector agent explains flagged phrases and suggests rewrites. | Yes – Authorship: typed, pasted and AI sources, replay and a shareable report; free. | Partly – Plagiarism checker in Pro. | Yes – Upload a file in the free web tool. | Yes – Authorship works in Canvas and Blackboard assignments; Education plans also integrate with Moodle. | Partly – Grammarly lists Spanish, French, German, Italian and Portuguese; no per-language accuracy published. | Partly – Near the top of the RAID leaderboard (self-submitted results); not part of major academic comparisons. | Partly – Education accounts: no training, FERPA, GDPR and SOC 2 listed. Individual users must opt out of training. |
| Studera.ai | 3.7 | Yes – Paste up to 1,200 words right on the page, no sign-up | Yes – Every sentence is labeled and can be clicked for its reason | Yes – AI-written, AI-rewritten and human, with the share of each | Yes – A reason for each flagged sentence, plus sentence rhythm, AI phrases and tips for writing in your own voice | Yes – Writing report with sessions, time, edits and pasted text when the essay is written in Studera | Partly – Plagiarism and AI in one report with Premium | Yes – Word, PDF and Google Docs; one file at a time free, many at once with Premium | Yes – Google Classroom, Canvas, Moodle and Teams; submissions are checked as they come in | Yes – 17 languages, including Spanish, French, German, Portuguese, Arabic and Japanese | No – Only Studera.ai's own tests so far | Yes – Never used for training, deletable reports, EU hosting for schools and a data processing agreement |
| QuillBot | 3.6 | Yes – Free scans run without signing in; 6 scans a day. | Yes – Flags specific sentences and passages. | Yes – Shows AI, Mixed (AI-refined) and Human percentages. | Yes – Explainer cards: one per scan free, unlimited with Premium. | No – Scores finished text only. | Partly – Separate plagiarism checker; 25,000 words a month with Premium. | Yes – Upload a document in the free web tool. | No – No Canvas, Google Classroom or other LMS integration. | Partly – 20+ languages incl. Spanish; no per-language accuracy published. | Partly – Listed near the top of the public RAID leaderboard (results submitted by QuillBot); not part of the 2025–2026 academic comparisons | Partly – GDPR, EU-US Data Privacy Framework and CASA Tier 2; trains on individual users' inputs by default. |
| Scribbr | 3.3 | Yes – No sign-up required. | Partly – Marks flagged paragraphs and sections rather than scoring each sentence. | Yes – AI-generated, human-written & AI-refined, human-written. | No – Shows percentages only; no reasons for flags. | No – Scores finished text only. | Partly – Paid per document; the free check shows only the top five matching sources. | Partly – Upload through the plagiarism checker; the free check gives an overall AI score, flagged fragments come with the paid report. | No – No LMS integration. | Partly – English, Spanish, German and French. | Partly – Walters (2023) tested an earlier version (500-word limit): 88% correct. | Partly – Says text is not stored or shared; QuillBot's privacy policy applies; no FERPA or COPPA terms. |
| Sapling | 3.0 | Yes – Free up to 5,000 characters per check without signing in. | Yes – Scores every sentence and highlights likely AI passages. | Partly – Shows "Mixed" when only some passages look AI; no AI-refined or paraphrased class. | No – No reasons given for flags. | No – Scores finished text only. | No – No plagiarism check. | Yes – PDF and DOCX with selectable text; scanned files need OCR first. | No – No LMS integration; API available. | Yes – 16 languages incl. Spanish with dedicated models; other languages run through the English model. | Partly – Weak results in 2023 studies (Walters; Liang et al.); the 2026 models have only Sapling's own figures. | Partly – SOC 2 Type II, GDPR and a zero-retention option; no FERPA statement, and content is stored by default. |
| ZeroGPT | 2.8 | Yes – Paste and scan without signing in. | Yes – Highlights the sentences that read as AI. | No – One overall AI percentage; no AI-refined or mixed label. | No – No reasons given for flags. | No – Scores finished text only. | Partly – 300 free words once; 35,000+ words a month on PLUS and higher plans. | Yes – .txt, .docx and text-based PDF; batch uploads on paid plans. | No – LMS connection only through the API. | Partly – Claims all languages; no per-language accuracy published. | Partly – Mixed: no false accusations in Weber-Wulff (2023), but a 16.9% false-positive rate on RAID (2024) with one threshold for all text types. | Partly – No training on text and GDPR hosting in Germany; no FERPA, COPPA or SOC 2 statement. |
| For teachers | ||||||||||||
| Revision History | 3.6 | Partly – Free tier, but you install the extension and sign in with Google. | No – No AI highlighting; it marks where text was pasted or deleted. | No – No AI categories; it reports typed vs pasted share instead. | Partly – Unusual-pattern flags say what stood out (for example a sudden large addition), but there is no AI verdict. | Yes – Replay, timeline, sessions, time spent, typed vs pasted share and paste details. | No – No plagiarism check. | No – Works on live Google Docs and Slides only. | Partly – Works with Google Classroom submissions written in Docs; no Canvas or Schoology integration. | No – No AI classifier in any language; the replay and stats work for any language. | No – Makes no AI call, so there is no accuracy figure; no study has measured it. | Yes – Local, in-browser analysis with no document storage; COPPA statement and custom DPAs for districts. |
| Turnitin | 3.4 | No – No free scans. Access only through a school license. | Yes – The AI report highlights the text segments the model scores as AI. No highlights for scores of 1–19%. | Yes – Separates "AI-generated only" from "AI-generated and AI-paraphrased"; humanizer (bypasser) edits count in the first group. | No – Shows which passages were flagged but not why. | Partly – Only with Turnitin Clarity, a paid add-on where students must write inside Turnitin's editor. | Yes – Similarity Report against web, publications and the student paper repository. | Yes – Students submit files; 300 to 30,000 words of prose for the AI indicator. | Yes – Canvas, Blackboard, Moodle, D2L Brightspace, Schoology, Google Classroom and Microsoft Teams. | Partly – Long-form English, Spanish, Japanese, Arabic, Dutch, German and Turkish. Paraphrase and bypasser detection is English only. | Partly – Tested in several independent studies: very few false positives, but it misses a large share of AI text. | Yes – Acts as a FERPA school official; Turnitin states it has never used student papers to train generative AI models. |
| Brisk Teaching | 3.3 | No – Premium feature; new teachers get a 14-day Premium trial | No – No AI highlighting. It colors typed, pasted and deleted text in the replay instead. | No – No AI categories. Pasted vs typed is the only split. | No – No AI verdict to explain; the teacher interprets the replay. | Yes – Video-style replay of the Google Doc's edit history with paste and pause markers. | No – No plagiarism check. | No – Works only on live Google Docs and OneDrive Word files, not uploads. | Partly – LTI and rostering for Canvas, Schoology and Google Classroom come with Brisk Premium; Inspect Writing itself runs on the Google Doc | No – No AI classifier in any language; the replay works for text in any language. | No – Brisk publishes no accuracy figures because it makes no AI call; no study has measured the replay. | Yes – FERPA, COPPA, SOC 2 Type II, 1EdTech seal, state DPAs; inputs never used to train AI models. |
| Packback | 3.2 | No – Needs a Packback course; no public free scanner. | Yes – Smart highlighter marks the passages most likely to be AI when risk is Medium or High. | No – Three risk levels (Low, Medium, High); no separate class for AI-edited or humanized text. | No – Says the model looks at word choice and sentence structure, but gives no per-text reasons. | Yes – Writing Process Report for Deep Dives and discussions: phases, time, sessions, pastes, AI-feedback use. | Yes – Uniqueness score against the web, 100M+ saved web documents and 50M+ student posts, including paraphrase detection. | Yes – Originality Review accepts PDF, DOCX and TXT; Canvas Sendback scans file uploads. | Yes – LTI 1.3 with Canvas, Blackboard, D2L Brightspace and Moodle, with gradebook sync. | No – AI Risk scoring is documented for English writing | No – Only Packback's own accuracy claims are published. | Partly – FERPA school official and COPPA consent process; AI providers barred from training on its data; no SOC 2 report named. |
| Draftback | 3.0 | No – Only a 30-day trial, and it needs your email. | No – No AI highlighting; it replays every edit. | No – No categories; Draftback does not label text. | No – By design it flags nothing as suspicious and shows only raw playback and statistics. | Yes – Full replay of a Google Doc's revision history plus time and pattern charts. | No – No plagiarism check. | No – Works only on live Google Docs. | No – No LMS integration; it works on Google Docs submitted through Classroom because teachers can edit them. | No – No AI classifier; the replay works for any language. | No – Makes no AI call, so there is nothing to benchmark; no study has measured it. | Yes – No document data leaves the browser; district DPAs available; follows ISO 27000-series practices per the developer. |
“Partly” can mean the feature is only on a paid plan, needs a free account, covers only a few languages, or (for independent tests) that results were mixed. Hover over a mark to see what applies.
Most accurate AI detectors
Detectors built to be trusted on a whole essay, with the strongest outside evidence on catching AI text without flagging human writers. Most of them need a paid plan for regular use.
MOST ACCURATE
1. Pangram
Pangram Labs
Rating: 4.3 out of 5
- Accuracy: 5.0
- Features: 4.0
- Price and free plan: 3.5
- Privacy and schools: 4.0
Why it got this score: Pangram has the strongest independent evidence of any detector, with the lowest false-positive rates across three separate studies. It lacks writing-process tools, its free plan needs an account and its terms require users to be 18+.
Best for: Teachers and schools that want the lowest false-positive risk
Pangram is the detector with the strongest independent test record in 2025–2026. It labels each passage as human, AI-assisted or AI-generated and very rarely flags human writing. The free plan needs an account and covers about 2,000 words a day.
What it can do
- Free, no account: Partly
- Flags sentences: Yes
- Spots mixed or edited AI: Yes
- Explains why: Partly
- Shows how it was written: No
- Plagiarism check: Partly
- Upload files: Yes
- Canvas or Google Classroom: Yes
- Chrome or Docs add-on: Yes
- Other languages: Yes
- Independently tested: Yes
- School-grade privacy: Yes
Strengths
- Best result in the two largest independent studies: Chicago Booth 2025 (near-zero false positives and misses) and VUB 2026 (97.5% of fully AI papers caught, no human paper flagged)
- Separates human, AI-assisted and AI-generated passages and flags humanized text
- Never trains on submitted text; FERPA compliant and SOC 2 Type 2 certified
- Integrates with Google Classroom, Canvas, Schoology, Moodle, Blackboard and Brightspace
- Supports 24 languages, including Spanish, with per-language false-positive rates published
- Free file upload with OCR for scanned documents
Weaknesses
- No writing-process replay; it judges the finished text only
- Free use requires an account and stops at about 2,000 words a day
- Terms of service require users to be 18 or older
- Still missed about 1 in 10 AI texts written to imitate a real author's style in Epoch AI's 2026 test
Free. What the free version includes: Free account with no card: 20 credits a day (about 2,000 words) plus 3 image scans a day, with file upload and AI-assistance detection. Paid plans start with a 7-day free trial.
Read the full review
How it works
Pangram runs its own deep-learning classifier rather than perplexity or burstiness scores. The current model, Pangram 4.1, was released on October 19, 2026. It splits a document into segments and labels each one Human, AI-Assisted or AI-Generated, with a separate humanizer score. Pangram trains on hard negatives, meaning human texts that earlier models got wrong, which is how it keeps false positives low. It is built for prose of at least 50 words and shows less certainty under 150 words. The same account also checks images and video, and paid plans add a similarity (plagiarism) check, Google Docs scanning and feed scanning in the browser extension.
Accuracy and false positives
Pangram reports a false-positive rate of about 1 in 10,000. Its 4.1 model card lists 70 errors on 781,981 human English texts (0.009%) and a 0.37% miss rate on 520,000 AI texts. Independent tests back this up. Jabarian and Imas (Chicago Booth, NBER w34223, 2025) found near-zero false positives and false negatives, also on short passages and humanized text; Pangram was the only tool that met a 0.5% false-positive cap. The VUB study by Van Vlasselaer et al. (IJEI, June 2026) tested 160 English academic papers: Pangram caught 97.5% of fully AI papers and estimated AI share correctly in 37 of 40 hybrid and 37 of 40 humanized papers, with no human paper flagged. Epoch AI (July 2026) found 0 false positives in 495 human passages but 10.1% missed AI texts that imitated a real author, and 25% in scientific writing.
Languages and non-native writers
Pangram supports 24 languages, including Spanish, Portuguese, French, Chinese, Arabic, Vietnamese, Hindi and Korean. It publishes false-positive rates per language; Spanish is listed at 0.0026%. Pangram reports a 1.99% miss rate on AI text across 18 languages, higher than in English. It also says it catches AI text that was run through a translator and back. For English learners, the VUB 2026 study is useful: its 40 human papers were written by non-native English graduate students, and Pangram flagged none of them. Pangram states that its model does not rely on perplexity, the measure of predictable word choice that made older detectors flag non-native writers in Liang et al. (2023).
Privacy and student data
Pangram states that it never uses submitted text to train its models and never sends it to outside AI providers. It is SOC 2 Type 2 certified (audited by Sensiba) and says it is FERPA compliant. Data is stored in the United States on AWS and encrypted in transit and at rest. Content is kept until the user deletes it, or up to 365 days after a license ends. Users can delete single scans, and organizations can request zero data retention or custom hosting. Pangram's terms of service require users to be at least 18, so classroom use runs through teacher or school accounts rather than students signing up on their own.
Pricing and school licenses
The free plan needs an account but no card and gives 20 credits a day, about 2,000 words, plus 3 image scans. Individual costs $20 a month or $15 a month billed yearly, for 300,000 words a month, plagiarism checks, the Chrome extension's feed scanner and Google Docs. Professional costs $65 a month or $45 billed yearly, for up to 1.5 million words. The API costs $0.05 per 100 words. Paid plans include a 7-day trial. Teachers who sign up through the education page get one month free. School and district licenses are quote-based and include LMS integrations, admin controls and an optional zero-retention setup.
Where others do better
GPTZero and Originality.ai record how a document was written, with typing replay and paste tracking; Pangram only reads the finished text. GPTZero's AI Patterns and Copyleaks' AI Logic explain in more detail why a passage looks AI-written. GPTZero, Copyleaks and Originality.ai let anyone run a scan without an account, while Pangram requires one. Copyleaks connects to more school systems, including Sakai and Edsby, and Originality.ai and Copyleaks cover more languages (31 and 30).
- Price from
- Individual $15/mo billed yearly or $20/mo billed monthly
- For schools
- One free month for educators. School and district licenses by quote, with LMS integrations and a zero-retention option.
- Best evidence
- Near-zero false positives in Chicago Booth's 2025 study; 97.5% of fully AI papers caught in VUB's 2026 study (vendor: 1 in 10,000 false positives)
- Languages
- 24 languages incl. Spanish
- Minimum text
- 50 words (results firmer above 150 words)
- Student data
- FERPA, SOC 2 Type 2; text never used for training; zero-retention option for organizations
- Works in
- Web, Chrome extension, Google Docs
- Age limit
- 18+ under Pangram's terms of service
Sources:
- Pangram – Pricing
- Pangram Knowledge Hub – Subscription tiers
- Pangram Knowledge Hub – What do I get with a free account?
- Pangram – Pangram 4.1 model card, Oct 2026
- Pangram – Multilingual AI detector
- Pangram Knowledge Hub – Does Pangram train on my data?
- Pangram Knowledge Hub – How does Pangram store my data?
- Pangram Knowledge Hub – LMS integrations
- Pangram – Terms of Service
- Jabarian & Imas 2025 – Artificial Writing and Automated Detection, NBER w34223
- Van Vlasselaer, Van Droogenbroeck & Spruyt 2026 – Who wrote this?, IJEI
- Epoch AI 2026 – AI detectors rarely flag human writing, but sometimes miss AI text imitating real authors
MOST ACCURATE
2. GPTZero
Superhuman
Rating: 4.2 out of 5
- Accuracy: 4.0
- Features: 5.0
- Price and free plan: 4.0
- Privacy and schools: 3.5
Why it got this score: GPTZero has the broadest feature set for teachers and a generous no-account free tier, and it rarely flags human writing in independent tests. It scored poorly on mixed human-AI papers in VUB's 2026 study, and the new Superhuman privacy policy allows training on individual users' content by default.
Best for: A second opinion with explanations and writing replay
GPTZero is the most complete detector for classrooms: sentence highlights, explanations of why text looks AI-written, writing replay in Google Docs and Canvas and Google Classroom integrations. It caught almost all plain AI text in recent independent tests but did poorly on mixed human-AI papers in VUB's 2026 study. Since June 2026 it belongs to Superhuman, the company formerly called Grammarly.
What it can do
- Free, no account: Yes
- Flags sentences: Yes
- Spots mixed or edited AI: Yes
- Explains why: Yes
- Shows how it was written: Yes
- Plagiarism check: Partly
- Upload files: Yes
- Canvas or Google Classroom: Yes
- Chrome or Docs add-on: Yes
- Other languages: Yes
- Independently tested: Yes
- School-grade privacy: Partly
Strengths
- Free scans of up to 10,000 characters without an account; 10,000 words a month with a free account
- AI Patterns and AI Vocabulary explain why a passage reads as AI-written
- Writing Report and Writing Replay show how a Google Doc was typed, pasted and edited
- No false positives on 495 human passages in Epoch AI's 2026 test, and under 1% in Chicago Booth's 2025 study
- Labels text as AI, Mixed or Human and flags AI-paraphrased text
- Canvas and Google Classroom integrations; official AI detector partner of the American Federation of Teachers
Weaknesses
- In VUB's 2026 study it estimated AI share correctly in 0 of 40 hybrid papers and missed most fully AI papers
- Since November 16, 2026, Superhuman's privacy policy lets it train on content from individual accounts unless the user opts out
- Multilingual detection, plagiarism checks and full sentence highlighting are paid features
- Terms of use set an 18+ age limit for individual accounts
Free. What the free version includes: Scans of up to 10,000 characters work without an account. A free account gives 10,000 words a month, a Basic scan with an overall score, one Advanced Scan with sentence highlights and batch upload of up to 3 files.
Read the full review
How it works
GPTZero classifies a whole document as AI, Mixed or Human and says how confident it is. Its score is the model's confidence, not the share of AI sentences. The default model since September 2026 is GPTZero 4o, built for mixed and edited AI text. Advanced Scan adds sentence highlights, an AI Paraphrased label for rewritten AI text and, since August 2026, AI Patterns, which names ten recurring AI writing habits in the text. A Chrome extension adds Writing Report and Writing Replay, which show typing, pastes and edits in Google Docs and other sites. Plagiarism, citation and hallucination checks sit in the same app. GPTZero needs at least 250 characters and says document-level results are more reliable than sentence-level ones.
Accuracy and false positives
GPTZero reports over 99% accuracy. On the public RAID leaderboard it catches 95.7% of AI text at a 1% false-positive rate, and 92.9% when RAID's evasion attacks are included. In Jabarian and Imas (Chicago Booth, NBER w34223, 2025), GPTZero kept false positives under 1%, missed only about 0–2% of AI text and stayed 96% accurate on short passages. Epoch AI (July 2026) found 0 false positives on 495 human passages, 2 misses on 297 plain AI texts and 10.8% misses on AI text imitating a real author. The VUB study by Van Vlasselaer et al. (IJEI, June 2026), using versions from May 2025, was much less favorable: 70% of fully AI papers scored under 20% AI, and GPTZero estimated AI share correctly in 0 of 40 hybrid and 1 of 40 humanized papers.
Languages and non-native writers
GPTZero detects AI in English and 25 other languages, including Spanish, Portuguese, French, Chinese, Arabic, Vietnamese and Hindi. Multilingual detection is part of the paid Premium and Professional plans. GPTZero's own help center says most of its training data is English prose by adults and that accuracy is highest for text like that. In Liang et al. (2023), seven popular detectors of that time mislabeled 61.3% of TOEFL essays by non-native writers on average. In October 2026, GPTZero reported 0 false positives on those same 91 TOEFL essays, on 914 essays by English learners and on 1,172 by students with disabilities in the PERSUADE 2.0 set. These are GPTZero's own numbers. In VUB's 2026 study, whose human papers came from non-native English graduate students, GPTZero gave a few low AI scores but labeled every paper human.
Privacy and student data
GPTZero is SOC 2 Type II audited and states compliance with FERPA, GDPR and CCPA. It has completed HECVAT and VPAT reviews. For students under 13, it relies on schools to collect parental consent under COPPA. After Superhuman bought GPTZero in June 2026, GPTZero moved to Superhuman's privacy policy on November 16, 2026. That policy says user content may train Superhuman's AI models, with an opt-out in account settings. Accounts managed by a school follow the school's agreement instead. GPTZero keeps a copy of submitted text but says it does not share it with third parties. Its terms of use set an 18+ age limit for individual accounts.
Pricing and school licenses
The free plan gives 10,000 words a month, up to 10,000 characters per scan, a Basic scan and one Advanced Scan; short scans work without an account. Premium costs $12.99 a month billed yearly or $23.99 monthly, for 300,000 words, Advanced Scan, multilingual detection, plagiarism checks and downloadable reports. Professional costs $24.99 a month billed yearly or $45.99 monthly, adding Canvas and Google Classroom integration, page-by-page scans and 250-file batches. Teams of 2 or more can buy Professional seats online. Schools and districts get quote-based enterprise plans with SSO, bulk seats, admin dashboards and LMS rollout support.
Where others do better
Pangram did far better on hybrid and humanized papers in VUB's 2026 study and had lower false-positive rates in Chicago Booth's 2025 study. Pangram also states that it never trains on submitted text, a firmer promise than Superhuman's opt-out policy for individual accounts. Copyleaks lets anyone scan up to 25,000 characters without an account and connects to more school systems, including Moodle, Brightspace, Schoology and Blackboard. Originality.ai includes plagiarism checks in its free credits, while GPTZero keeps them for paid plans.
- Price from
- Premium $12.99/mo billed yearly or $23.99/mo billed monthly
- For schools
- Team plans from $24.99 per seat a month, billed yearly. District plans by quote with SSO and LMS rollout support.
- Best evidence
- 0 false positives in 495 human texts (Epoch AI 2026) and 95.7% of AI caught at a 1% false-positive rate on RAID; missed most hybrid papers in VUB 2026
- Languages
- 26 languages incl. Spanish (paid plans)
- Minimum text
- 250 characters
- Student data
- FERPA, SOC 2 Type II, GDPR, CCPA; individual accounts can opt out of AI training
- Works in
- Web, Chrome extension, Google Docs, Microsoft Word
- Age limit
- 18+ for individual accounts; younger students use it through a school, which handles parental consent
Sources:
- GPTZero – Pricing
- GPTZero Support – Free and paid plans
- GPTZero Support – What languages does GPTZero support?
- GPTZero – For educators
- GPTZero – Privacy, security and compliance
- Superhuman – Privacy Policy, effective Nov 16, 2026; covers GPTZero
- GPTZero – Introducing GPTZero 4o, Oct 2026
- Business Wire – Superhuman to acquire GPTZero, June 23, 2026
- RAID benchmark – Public leaderboard
- Jabarian & Imas 2025 – Artificial Writing and Automated Detection, NBER w34223
- Van Vlasselaer, Van Droogenbroeck & Spruyt 2026 – Who wrote this?, IJEI
- Epoch AI 2026 – AI detectors rarely flag human writing, but sometimes miss AI text imitating real authors
MOST ACCURATE
3. Originality.ai
Rating: 3.7 out of 5
- Accuracy: 3.5
- Features: 4.0
- Price and free plan: 4.0
- Privacy and schools: 3.0
Why it got this score: Originality.ai offers AI detection, plagiarism checks and writing replay at one of the lowest per-word prices. Its independent record is mixed, with more false positives and misses than Pangram and GPTZero in recent tests, and its free checker may use pasted text for training.
Best for: AI and plagiarism checks plus writing replay on a budget
Originality.ai combines AI detection, plagiarism checks and a writing replay extension at a low per-word price. Independent tests are mixed: it held up well against paraphrasing in the RAID benchmark but missed more AI text than Pangram and GPTZero in Chicago Booth's 2025 study and flagged 3.8% of human texts in Epoch AI's 2026 test. Text pasted into the free checker may be used for training.
What it can do
- Free, no account: Yes
- Flags sentences: Yes
- Spots mixed or edited AI: Yes
- Explains why: No
- Shows how it was written: Yes
- Plagiarism check: Yes
- Upload files: Partly
- Canvas or Google Classroom: Yes
- Chrome or Docs add-on: Yes
- Other languages: Yes
- Independently tested: Partly
- School-grade privacy: Partly
Strengths
- Labels text as Likely AI, Mixed or Likely Original with sentence highlights
- Writer Replay in Chrome, Firefox and Google Docs shows how a draft was written
- Plagiarism, fact-checking and readability tools in the same app
- 31 languages, including Spanish
- Pro plan covers 200,000 words a month for $12.95 billed yearly
- Caught 96.7% of paraphrased AI text in the RAID benchmark paper (ACL 2024)
Weaknesses
- Flagged 19 of 495 human passages (3.8%) in Epoch AI's 2026 test, where Pangram and GPTZero flagged none
- Missed 10–40% of AI text, depending on the model, in Chicago Booth's 2025 study
- Text pasted into the free checker may be used for model training unless you have an account and opt out
- No explanation of why a passage was flagged; the stricter 5.0 Max model uses 10 times as many credits
Free. What the free version includes: A first scan works without an account. Free use is capped at 100 credits per account, about 10,000 words of AI checks with the 5.0 model, and plagiarism checks use 2 credits per 100 words. File upload needs a paid plan.
Read the full review
How it works
Originality.ai released its 5.0 and 5.0 Max models in November 2026. Both give an AI score and sort text into Likely Original (under 10% AI), Mixed (10–80%) or Likely AI (81–100%), with sentence highlights. 5.0 Max is meant for important reviews and costs 10 credits per 100 words instead of 1. The same app checks plagiarism, facts, grammar and readability. The Chrome and Firefox extensions add Writer Replay, which plays back how a Google Doc was drafted. The website needs at least 100 words per scan, and Originality.ai says accuracy drops on shorter texts.
Accuracy and false positives
Originality.ai reports 99.95% of student essays correctly labeled human on PERSUADE 2.0 and 89.5–92.4% of humanized AI text caught; these are its own numbers. Independent results are mixed. Walters (Open Information Science, 2023) named it one of three tools, with Copyleaks and Turnitin, that were highly accurate on all three sets of test essays (126 in total). In the RAID benchmark paper (Dugan et al., ACL 2024), it detected 85.0% of AI text at a 5% false-positive rate and 96.7% after paraphrasing, so paraphrase tools did not fool it. Jabarian and Imas (Chicago Booth, NBER w34223, 2025) found false positives under 1% but 10–40% of AI text missed, depending on the model. Epoch AI (July 2026) found it flagged 19 of 495 human passages (3.8%) and missed 17.8% of AI text imitating a real author.
Languages and non-native writers
Originality.ai detects AI in 31 languages, including Spanish, Portuguese, French, Arabic, Chinese, Vietnamese, Hindi and Korean, and picks the language automatically. It reports an AI-class F1 score of 92–93% across 30 languages, below its English results; this is the vendor's own figure. For English learners, Originality.ai reports 98.9% (5.0) and 100% (5.0 Max) of TOEFL essays from Liang et al. (2023) correctly labeled human. The human passages in Epoch AI's 2026 test, where it flagged 3.8%, were blogs, fiction and scientific writing by 99 authors, so polished formal writing can still trigger flags.
Privacy and student data
Text pasted into the free checker may be used to train Originality.ai's models. Account holders can opt out of training and delete scan history in one click. For schools with an institutional agreement, the privacy policy says student data is used only to run the contracted service and is not used for general-purpose AI training, and that it handles such data in line with FERPA, COPPA and state student-privacy laws. Schools are responsible for parental consent for children under 13. Minors may use the service only with permission and supervision from a parent, guardian or school.
Pricing and school licenses
Free use is capped at 100 credits per account; a first scan works without signing up. One credit covers 100 words of AI checks with the 5.0 model. Pro costs $12.95 a month billed yearly ($155.40) or $14.95 monthly, for 2,000 credits (200,000 words) a month, file upload, bulk scans, shareable reports and PDF downloads. With 5.0 Max, the same credits cover about 20,000 words. Enterprise costs $136.58 a month billed yearly or $179 monthly, for 1.5 million words and API access. Schools book a call for quote-based plans with faculty access, Moodle and Google Classroom support.
Where others do better
Pangram and GPTZero flagged no human passages in Epoch AI's 2026 test and missed less AI text in Chicago Booth's 2025 study. GPTZero's AI Patterns and Copyleaks' AI Logic explain why a passage was flagged; Originality.ai shows only scores and highlights. Copyleaks and GPTZero connect to Canvas, which Originality.ai does not list. Pangram never trains on submitted text and Winston AI only does if you opt in, while Originality.ai's free checker may.
- Price from
- Pro $12.95/mo billed yearly or $14.95/mo billed monthly
- For schools
- School plans by quote with a Moodle integration and Google Classroom replay. Extra Pro seats cost about $9–10 a month.
- Best evidence
- 3.8% of human texts flagged and 17.8% of style-imitating AI missed in Epoch AI's 2026 test; 10–40% of AI missed in Chicago Booth's 2025 study
- Languages
- 31 languages incl. Spanish
- Minimum text
- 100 words on the website
- Student data
- Student data under school contracts not used for general AI training; free-checker text may be used for training (opt-out with an account)
- Works in
- Web, Chrome extension, Google Docs
- Age limit
- Minors need permission and supervision from a parent or school; children under 13 only through a school account
Sources:
- Originality.ai – Pricing
- Originality.ai – AI Checker and FAQ
- Originality.ai – Introducing Originality.ai 5.0 and 5.0 Max, Nov 2026
- Originality.ai Help – Multi-language AI detection
- Originality.ai – Education
- Originality.ai – Privacy Policy
- Originality.ai – Terms and Conditions
- Jabarian & Imas 2025 – Artificial Writing and Automated Detection, NBER w34223
- Epoch AI 2026 – AI detectors rarely flag human writing, but sometimes miss AI text imitating real authors
- Dugan et al. 2024 – RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors, ACL 2024
- Walters 2023 – The Effectiveness of Software Designed to Detect AI-Generated Writing, Open Information Science
MOST ACCURATE
4. Copyleaks
Rating: 3.6 out of 5
- Accuracy: 3.0
- Features: 4.5
- Price and free plan: 3.5
- Privacy and schools: 3.5
Why it got this score: Copyleaks leads on school integrations, free no-account scanning and flag explanations, and it rarely flags human writing. It missed most AI and mixed papers in VUB's 2026 study, its personal plans include few words, and it trains on content unless you opt out.
Best for: Schools that want AI and plagiarism checks inside their LMS
Copyleaks has the most generous no-account free scan (25,000 characters) and the widest set of school integrations, and its AI Logic feature explains why text was flagged. It almost never flags human writing, but in VUB's 2026 study it missed most AI-written and mixed papers. Paid personal plans include only about 25,000 words a month.
What it can do
- Free, no account: Yes
- Flags sentences: Yes
- Spots mixed or edited AI: Partly
- Explains why: Yes
- Shows how it was written: No
- Plagiarism check: Partly
- Upload files: Yes
- Canvas or Google Classroom: Yes
- Chrome or Docs add-on: Yes
- Other languages: Yes
- Independently tested: Partly
- School-grade privacy: Partly
Strengths
- Scan up to 25,000 characters free without an account
- AI Logic shows the AI-typical phrases and source matches behind a flag
- AI detection and plagiarism results in one report
- Integrates with Canvas, Google Classroom, Moodle, Brightspace, Schoology, Blackboard, Sakai and Edsby
- 30 languages, with Spanish results published (99.85% of human text and 98.02% of AI text correct, vendor figures)
- No human paper flagged in VUB's 2026 study
Weaknesses
- In VUB's 2026 study, 75% of fully AI papers scored under 20% AI, and AI share was estimated correctly in only 30% of hybrid papers
- Personal plan includes about 25,000 words a month, far less than rivals at a similar price
- Uses submitted content to train its models unless you email support to opt out
- No writing-process replay
Free. What the free version includes: Paste up to 25,000 characters per scan without an account. A free account unlocks sentence-by-sentence detail, highlighted AI phrases and extra monthly credits.
Read the full review
How it works
Copyleaks scores how much of a text looks AI-written and highlights the passages behind the score. AI Logic, launched in 2025, adds two kinds of evidence: AI phrases, which appear far more often in AI text than in human text, and AI Source Match, which looks for matches with AI-generated content already published online. Plagiarism results appear in the same report. The web checker needs at least 255 characters and takes up to 25,000 for free. Schools can tune AI sensitivity per assignment and see analytics by student, course or institution. Copyleaks also checks images and video for AI.
Accuracy and false positives
Copyleaks reports over 99% accuracy and a 0.03% false-positive rate; in English it reports 99.97% of human text and 99.20% of AI text correctly labeled. Independent results are mixed. Walters (Open Information Science, 2023) named Copyleaks one of three tools, with Turnitin and Originality.ai, that stayed accurate on GPT-3.5, GPT-4 and student essays. The VUB study by Van Vlasselaer et al. (IJEI, June 2026), using versions from May 2025, found no false positives on 40 human papers. But 75% of fully AI papers scored under 20% AI and the rest under 40%, and Copyleaks estimated AI share correctly in only 30% of hybrid and 22.5% of humanized papers. Copyleaks was not part of the Chicago Booth 2025 or Epoch AI 2026 tests and has no entry on the public RAID leaderboard.
Languages and non-native writers
Copyleaks detects AI in 30 languages, including Spanish, Portuguese, French, Chinese, Arabic, Hindi, Vietnamese and Korean. It publishes per-language results from its own testing: 99.85% of human and 98.02% of AI text correctly labeled in Spanish, and lower AI catch rates in French (96.18%), German (95.63%) and Portuguese (93.08%). The Pro plan adds detection of AI text translated from another language. Copyleaks says its models keep false positives low for non-native English text. In VUB's 2026 study, it flagged none of the 40 papers written by non-native English graduate students. Copyleaks also warns that AI writing assistants such as Grammarly can lead to text being flagged.
Privacy and student data
Copyleaks states GDPR compliance and SOC 2 and SOC 3 certification. For schools, it acts as a FERPA "school official" and uses education records only as the school directs. Its privacy policy says it uses submitted content to train its models by default. Direct customers can opt out by emailing support@copyleaks.com, and schools can opt out for their users. Schools can also choose whether scans go into a Shared Data Hub that other users' scans are compared against, or a private hub. Users must have reached the age of majority; younger users need a parent's or guardian's permission and supervision.
Pricing and school licenses
The free checker takes 25,000 characters per scan with no account, and a free account adds sentence-level detail and extra monthly credits. Personal costs $16.99 a month, or $13.99 a month billed yearly ($167.88). It includes 100 credits a month, about 25,000 words, because one credit covers 250 words or one image. AI and plagiarism results share those credits. Pro costs $99.99 a month or $74.99 billed yearly, for 25 seats and about 250,000 words a month. Education licenses are quote-based, priced by full-time students, and include LMS integrations, analytics, grammar checks and a private cloud. Canvas schools can buy Copyleaks through Instructure.
Where others do better
Pangram caught 97.5% of fully AI papers in VUB's 2026 study, where Copyleaks missed most of them. GPTZero and Originality.ai include writing replay, which Copyleaks does not offer. For individual users, GPTZero Premium (300,000 words) and Originality.ai Pro (200,000 words) include far more words per month at a similar price. Pangram never uses submitted text for training and Winston AI only does if you opt in, while Copyleaks does so by default.
- Price from
- Personal $13.99/mo billed yearly or $16.99/mo billed monthly
- For schools
- Education licenses by quote, priced per student, with LMS integrations. Canvas schools can buy through Instructure.
- Best evidence
- No false positives but 75% of fully AI papers missed in VUB's 2026 study (vendor: 99%+ accuracy, 0.03% false positives)
- Languages
- 30 languages incl. Spanish
- Minimum text
- 255 characters (350 in the browser extension)
- Student data
- FERPA, SOC 2 and SOC 3, GDPR; trains on content by default, opt-out by request
- Works in
- Web, Chrome extension, Google Docs
- Age limit
- Age of majority; younger users only with a parent's or guardian's permission and supervision
Sources:
- Copyleaks – Pricing
- Copyleaks – AI Detector
- Copyleaks – Academic integrity platform for institutions
- Copyleaks Docs – AI detection supported languages
- Copyleaks – AI detector testing methodology
- Copyleaks – Privacy Policy
- MarketScale – Copyleaks LMS reach and Instructure partnership, Oct 2026
- Van Vlasselaer, Van Droogenbroeck & Spruyt 2026 – Who wrote this?, IJEI
- Walters 2023 – The Effectiveness of Software Designed to Detect AI-Generated Writing, Open Information Science
MOST ACCURATE
5. Winston AI
Rating: 3.3 out of 5
- Accuracy: 3.5
- Features: 3.5
- Price and free plan: 3.0
- Privacy and schools: 3.0
Why it got this score: Winston AI leads the RAID leaderboard and stands out for reading handwriting, but the results for its current model come from its own tests and RAID submission. It has no ongoing free plan, no writing replay and no FERPA statement, which limits its fit for US schools.
Best for: Teachers who grade handwritten or scanned work
Winston AI tops the public RAID benchmark leaderboard and can scan photos of handwritten work with OCR. Its accuracy figures come from its own tests and its RAID leaderboard submission. There is no ongoing free plan, only a 14-day trial with 2,000 words.
What it can do
- Free, no account: Partly
- Flags sentences: Yes
- Spots mixed or edited AI: Partly
- Explains why: Partly
- Shows how it was written: No
- Plagiarism check: Yes
- Upload files: Yes
- Canvas or Google Classroom: Partly
- Chrome or Docs add-on: Yes
- Other languages: Yes
- Independently tested: Partly
- School-grade privacy: Partly
Strengths
- First place on the RAID leaderboard: 99.9% of AI text caught at a 1% false-positive rate, including evasion attacks
- OCR reads photos of handwriting and scanned pages
- Sentence-by-sentence AI Prediction Map and an Explain button that summarizes results in plain language
- Plagiarism, fact-checking and essay-grading tools in the same account
- Does not train on submitted text unless the user opts in; SOC 2 Type 2 certified
- Yearly plans from $10 a month for 100,000 words
Weaknesses
- No free plan after the 14-day, 2,000-word trial, and the trial needs an account
- Results for the current model come from Winston's own testing and its RAID submission, which uses AI models from 2023
- No writing-process replay, and Google Classroom is the only school system it lists
- When text is flagged, the Explain panel also suggests ways to make the writing sound more human
Free. What the free version includes: 14-day free trial with 2,000 credits (one credit per word for AI detection) after creating an account; no card needed. Paste at least 500 characters per scan.
Read the full review
How it works
Winston AI gives each text a Human Score from 0 to 100%. The score reflects how human the whole text reads, not the share of human-written words. An AI Prediction Map colors each sentence, and a list shows the sentences that pushed the score down. The Explain button writes a plain-language summary of the result. Winston runs its 5.0 model, needs at least 500 characters and advises at least 300 words. Its OCR reads photos and scans, including handwriting, which few rivals do. The same account checks plagiarism, facts, AI images and deepfakes, and includes an essay grader.
Accuracy and false positives
Winston AI reports 99.98% accuracy. It holds first place on the public RAID leaderboard, catching 100% of AI text at a 1% false-positive rate without attacks and 99.9% with RAID's evasion attacks included; GPTZero scores 92.9% in the same attacked setting. Vendors run their own detectors on RAID's hidden test set and RAID scores the results. RAID's AI texts come from 2023-era models such as GPT-4, ChatGPT, Llama 2 and Mistral 7B. Winston also reports zero errors on 8,697 texts from the Chicago Booth dataset and 25,996 student essays from PERSUADE 2.0, but it ran those tests itself; it was not one of the tools in the Chicago Booth study. An early version was among 14 tools that all stayed below 80% accuracy in Weber-Wulff et al. (2023).
Languages and non-native writers
Winston AI detects AI in 23 languages: English, Spanish, French, Portuguese, German, Dutch, Italian, Polish, Romanian, Indonesian, Tagalog, Russian, Ukrainian, Bulgarian, Chinese (Simplified), Czech, Slovak, Croatian, Serbian, Bosnian, Macedonian, Belarusian and Azerbaijani. It does not cover Arabic, Vietnamese, Korean or Japanese. Winston says results can vary by language and that text in unlisted languages is not a validated result. For English learners, its main evidence is its own PERSUADE 2.0 run, where it labeled all 25,996 student essays human. Evidence on non-native writers comes only from Winston's own testing.
Privacy and student data
Winston AI is SOC 2 Type 2 certified and GDPR compliant. Its privacy policy says it does not use submitted text or images to train AI models without explicit opt-in consent, and it does not sell or license submitted content. Submitted content is deleted within 30 days of account deletion, with encrypted backups purged after 90 days, and educators can purge data on request. Its privacy pages cite GDPR and SOC 2 rather than FERPA or COPPA. The policy says users must be 18 or have a parent's or guardian's consent.
Pricing and school licenses
The free trial lasts 14 days and includes 2,000 credits; AI detection costs one credit per word, plagiarism two. Essential costs $18 a month or $10 a month billed yearly ($120), for 100,000 credits a month. Advanced costs $29 a month or $16 billed yearly, for 200,000 credits and up to 5 team members. Elite costs $49 a month or $26 billed yearly, for 500,000 credits and unlimited team members. Through Black Friday 2026, the code BF50 cuts these prices in half. Schools and districts can request quote-based institutional plans.
Where others do better
Pangram and GPTZero have results from university and journal studies in 2025 and 2026, while Winston's current numbers come from its own tests and RAID. GPTZero and Originality.ai record how a document was written. Copyleaks and Pangram connect to many more school systems, including Canvas, Moodle and Schoology. GPTZero, Copyleaks and Originality.ai offer free scans without an account, and Pangram has an ongoing free plan, while Winston offers only a 14-day trial. Pangram and Copyleaks also cover more languages, including Arabic and Vietnamese.
- Price from
- Essential $10/mo billed yearly or $18/mo billed monthly
- For schools
- Schools and districts can request quote-based institutional plans. The Advanced and Elite plans add team seats (5 and unlimited) with shared credits.
- Best evidence
- 99.9% of AI caught at a 1% false-positive rate on the RAID leaderboard (vendor-submitted; vendor: 99.98% accuracy)
- Languages
- 23 languages incl. Spanish
- Minimum text
- 500 characters (300+ words advised)
- Student data
- SOC 2 Type 2, GDPR; text not used for training without opt-in
- Works in
- Web, Chrome extension, Google Docs
- Age limit
- 18+, or a minor with a parent's or guardian's consent
Sources:
- Winston AI – Pricing
- Winston AI – AI detector for education
- Winston AI Help – Which languages does Winston AI work with?
- Winston AI Help – How to use Winston AI for text analysis
- Winston AI Help – Security and compliance
- Winston AI – Privacy Policy
- Winston AI – Winston 5.0 vs popular AI detectors, Oct 2026
- RAID benchmark – Public leaderboard
- Dugan et al. 2024 – RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors, ACL 2024
- Weber-Wulff et al. 2023 – Testing of detection tools for AI-generated text, IJEI
Best free AI detectors for checking your own writing
Quick checks you can run without paying, most of them without an account. Good for a first look at a draft; weaker as the only evidence in a grading decision.
FREE CHECKERS
1. Grammarly
Superhuman
Rating: 3.8 out of 5
- Accuracy: 3.5
- Features: 4.0
- Price and free plan: 4.0
- Privacy and schools: 4.0
Why it got this score: Grammarly's free Authorship report is the most practical way for a student to prove their own writing, and it works in Canvas and Blackboard. The free detector itself is bare, and its accuracy evidence rests on a self-submitted RAID leaderboard entry.
Best for: Students who want to show how they wrote, not just pass a scan
Grammarly's free AI Detector gives one overall AI percentage for up to 1,400 words, with no account. Its stronger tool for students is Grammarly Authorship: a free, opt-in record of what was typed, pasted or AI-generated, with a replay and a report the student can share, including in Canvas and Blackboard. The detector ranks near the top of the public RAID benchmark leaderboard.
What it can do
- Free, no account: Yes
- Flags sentences: Partly
- Spots mixed or edited AI: No
- Explains why: Partly
- Shows how it was written: Yes
- Plagiarism check: Partly
- Upload files: Yes
- Canvas or Google Classroom: Yes
- Chrome or Docs add-on: Yes
- Other languages: Partly
- Independently tested: Partly
- School-grade privacy: Partly
Strengths
- Authorship is free: it labels typed, pasted and AI text in Google Docs, Word and Grammarly's editor, with a replay and a report the student chooses to share
- Authorship works inside Canvas and Blackboard assignments at Grammarly for Education schools
- Catches 99.7% of AI text at a 1% false-positive setting on the RAID leaderboard (98.2% with adversarial attacks)
- Free scans of up to 1,400 words with no account, plus file upload
- Grammarly says Education customers' text is never sold or used to train models
- Runs where students already write: Google Docs, Word and Grammarly's desktop apps
Weaknesses
- The free web scan shows only one percentage; highlighted sections and explanations need Pro
- No mixed or AI-refined category, and text rewritten with Grammarly's own AI agents raises the AI score
- Accuracy evidence comes from Grammarly's RAID leaderboard entry; it was not part of the major 2025–2026 academic comparisons
- For individual accounts, keeping text out of model training requires an opt-out in settings
Free. What the free version includes: Up to 1,400 words per scan with no account; the result is one overall percentage. Grammarly's own comparison says more scans require a free sign-up. Authorship is free with a free account.
Read the full review
How it works
Grammarly splits text into sections, checks each for patterns linked to AI writing and reports what share of the text resembles AI. The free web tool shows only that percentage. In Pro, the AI Detector agent in Grammarly's docs, Google Docs and Word marks flagged sections, explains why, offers rewrites and creates citations for AI use. Authorship is a separate, opt-in feature: while a student writes in Google Docs, Word or Grammarly's editor, it labels passages as typed, pasted from a website, AI-generated or suggested by Grammarly, and records a replay. The student decides whether to share the report. Grammarly's parent company, Superhuman, also bought GPTZero in June 2026.
Accuracy and false positives
Grammarly says its detector reaches 99% accuracy and ranks #1 on RAID, a public benchmark from the University of Pennsylvania (Dugan et al., ACL 2024). On the current RAID leaderboard, Grammarly catches 99.7% of AI text when the threshold is set for a 1% false-positive rate, and 98.2% when adversarial attacks such as paraphrasing are included; a few entries, including Winston AI, score as high or higher. Vendors submit their own results and RAID sets the threshold, so the numbers show how well the tool separates AI from human text, not its error rate at the default setting. In an October 2024 Decrypt test on the Declaration of Independence, Grammarly did best of four detectors. Grammarly's help center warns that its rewrite, paraphrase and humanizer agents use an LLM, so text edited with them will raise the AI score.
Languages and non-native writers
Grammarly's comparison page says its detector handles Spanish, French, German, Italian and Portuguese as well as English, the same languages its writing suggestions cover. It publishes no per-language accuracy and no tests on essays by English learners. Grammarly says the model is tuned to minimize false positives because wrongly flagging human writing is worse than missing AI text. For ESL students worried about a false flag, Authorship is the more useful tool: it shows the text was typed, which a percentage cannot.
Privacy and student data
Grammarly's trust center says it does not sell user content and that individual users can opt out of model training in account settings. For Grammarly for Education, Grammarly states that no user content is used to train its models and that partners may not store or train on it; its education brochure lists GDPR, FERPA and SOC 2. Authorship is opt-in and tracks text sources only with consent. A school sees a report only when the student shares it. Pasted-text data is used briefly to identify its source and then anonymized. Users must be old enough to consent on their own: 13 under most US law.
Pricing and school licenses
The web detector and Authorship are free. Grammarly Pro costs $30 a month or $144 a year ($12 a month) and adds the AI Detector agent with highlighted sections and explanations, the plagiarism checker and more generative AI features. Grammarly for Education is sold to institutions by quote, with admin controls and LMS integrations; Authorship in Canvas requires an institution-wide plan. Grammarly also offers an AI Humanizer agent.
Where others do better
GPTZero, now owned by the same company, adds mixed-text categories and AI Patterns explanations, and it has stronger independent results (Chicago Booth 2025, Epoch AI 2026). QuillBot and Scribbr separate AI-refined text in their free tools. Copyleaks and Sapling publish longer lists of supported languages. Turnitin Clarity offers a similar writing-process record built into the LMS, but only through institutions.
- Price from
- Pro $12/mo billed yearly or $30/mo billed monthly
- For schools
- Grammarly for Education licenses by quote, with Authorship in Canvas and Blackboard and no training on student text.
- Best evidence
- 99.7% of AI text caught at a 1% false-positive setting on the RAID leaderboard (vendor: 99% accuracy, #1 on RAID)
- Languages
- English plus Spanish, French, German, Italian and Portuguese (vendor list)
- Minimum text
- Longer passages score more reliably (Grammarly)
- Student data
- Education: FERPA, SOC 2, GDPR, no training on text; individuals can opt out of training
- Works in
- Web, Chrome extension, Google Docs, Microsoft Word, Windows, Mac
- Age limit
- 13+ under most US law (users must be old enough to consent on their own; 16+ in the EU)
Sources:
- Grammarly – Free AI Detector
- Grammarly Support – AI Detector user guide
- Grammarly – Authorship
- Grammarly – Authorship is now available in Canvas LMS
- Grammarly – Grammarly for Education
- Grammarly for Education – Secure, responsible AI brochure
- Grammarly – Trust Center
- Grammarly – Terms of Service
- Grammarly – Best AI detectors, plan prices and free limits, Oct 2026
- RAID benchmark – Public leaderboard, TPR at 1% FPR
- Decrypt, Oct 2024 – AI detectors claim the Declaration of Independence was 98% AI-generated
FREE CHECKERS
2. Studera.ai
· Our own tool
Rating: 3.7 out of 5
- Accuracy: 3.0
- Features: 4.0
- Price and free plan: 4.5
- Privacy and schools: 4.0
Why it got this score: A reason for every flagged sentence, a writing report, file upload and LMS integrations, free for up to 1,200 words without an account and in 17 languages. Its accuracy rests on Studera.ai's own tests so far.
Best for: Seeing why each sentence was flagged, and how it was written
Studera.ai's AI detector labels every sentence as AI-written, AI-rewritten or human and explains each flag. You can paste up to 1,200 words with no account, upload Word, PDF and Google Docs files, and see a writing report of how the essay was written. Teachers get a class overview, a conversation guide and a plagiarism check on Premium, in 17 languages including Spanish.
What it can do
- Free, no account: Yes
- Flags sentences: Yes
- Spots mixed or edited AI: Yes
- Explains why: Yes
- Shows how it was written: Yes
- Plagiarism check: Partly
- Upload files: Yes
- Canvas or Google Classroom: Yes
- Chrome or Docs add-on: Yes
- Other languages: Yes
- Independently tested: No
- School-grade privacy: Yes
Strengths
- Every sentence gets a label (AI-written, AI-rewritten or human) and a plain-language reason, plus a confidence level
- Free with no account for up to 1,200 words; a free account adds file upload for Word, PDF and Google Docs
- Writing report: when students write in Studera, sessions, edits and pasted text are saved and can be shared
- For teachers: check a whole class at once, compare with a student's earlier writing and get questions for a follow-up conversation
- Works where essays are written: Chrome and Edge extension, Google Docs, Word, Google Classroom, Canvas, Moodle and Teams
- Judges Spanish and 15 other languages on their own terms; no humanizer, and text is never used to train AI models
Weaknesses
- Accuracy figures come only from Studera.ai's own testing; no published study or benchmark covers it yet
- Plagiarism checks, PDF reports, class overviews and full-length papers need Premium
- Free checks have a daily limit
Free. What the free version includes: Up to 1,200 words per check with no account, and a daily number of checks. A free account adds file upload (one file at a time), saved reports and the writing report.
Read the full review
How it works
The detector splits the text into sentences and measures how much sentence length varies and which stock AI phrases show up in that language. A language model then labels each sentence as AI-written, AI-rewritten or human, with a short reason, and the overall percentage is calculated from those labels, weighted by length. Every flagged sentence comes with a tip for rewriting it in your own words. When students write in Studera, a writing report also shows how the essay came together: sessions, time spent, edits and pasted text.
Accuracy and false positives
Studera.ai hasn't been included in any published study or public benchmark, so there are no independent accuracy or false-positive numbers for it. That puts it behind Pangram, GPTZero and Originality.ai, which have been tested by outside researchers.
To limit the damage of a wrong call, results come with a confidence level, texts under 150 words are never rated with high confidence, and the score is shown in three bands rather than as an exact verdict. Teachers can also compare an essay with the student's earlier writing and look at the writing report, which says more than any percentage.
Languages and non-native writers
The detector works in 17 languages, including Spanish, French, German, Portuguese, Arabic, Hindi and Japanese, and looks for AI phrases specific to each one instead of treating other languages as broken English. Phrases that students commonly use on their own are left off the lists on purpose. That matters for English learners, whose plain and predictable writing is what older detectors most often mistook for AI.
Privacy and student data
Text is used only for the check you ask for and is never used to train AI models, and you can delete checks and reports whenever you want. The school profile runs in the EU, with hosting, storage and AI processing under GDPR, and schools and districts get a data processing agreement. Teachers decide whether students can use the detector in their classes. There's no published FERPA or SOC 2 statement, which many US districts ask for.
Pricing and school licenses
Checks of up to 1,200 words are free, with a daily limit and no account needed; a free account adds file upload, saved reports and the writing report. Premium costs $5 a month or $45 a year and adds unlimited checks, full-length papers and dissertations, uploading many files at once, PDF reports and share links, plagiarism checks, and the class overview with a conversation guide. Schools and districts can get a license with the EU school profile.
Where others do better
Pangram and GPTZero have independent evidence behind their accuracy, which Studera.ai doesn't have yet. Turnitin is already built into many districts' grading workflow, and Grammarly Authorship tracks writing in Google Docs and Word even when students don't write in a separate tool. Copyleaks and Originality.ai cover more languages for plagiarism checks.
- Price from
- Premium $5/mo or $45/yr
- For schools
- School and district licenses with the class overview, an EU-hosted school profile and a data processing agreement.
- Best evidence
- Studera.ai's own tests only
- Languages
- 17 languages, incl. Spanish
- Minimum text
- 80 words (1,200 free; full papers on Premium)
- Student data
- Never used for training; EU school profile; DPA
- Works in
- Web, Chrome extension, Google Docs, Microsoft Word, LMS
Sources:
FREE CHECKERS
3. QuillBot
Learneo
Rating: 3.6 out of 5
- Accuracy: 3.5
- Features: 3.5
- Price and free plan: 4.5
- Privacy and schools: 3.0
Why it got this score: QuillBot pairs a strong RAID leaderboard result with the most useful free breakdown for students: AI, AI-refined and human, plus explanations, with no account. It lacks school features, and individual users' text trains its models by default.
Best for: Free draft checks with a separate score for AI-refined text
QuillBot's free AI Detector scores text from 0 to 100% and splits it into AI, Mixed (human writing refined with AI tools) and Human. It highlights flagged sentences and adds an explainer card on why a passage reads as AI. On the public RAID benchmark leaderboard it sits near the top.
What it can do
- Free, no account: Yes
- Flags sentences: Yes
- Spots mixed or edited AI: Yes
- Explains why: Yes
- Shows how it was written: No
- Plagiarism check: Partly
- Upload files: Yes
- Canvas or Google Classroom: No
- Chrome or Docs add-on: No
- Other languages: Partly
- Independently tested: Partly
- School-grade privacy: Partly
Strengths
- Separates fully AI-written text from human text that was polished with AI tools
- Catches 99.2% of AI text at a 1% false-positive setting on the public RAID leaderboard (98.1% with adversarial attacks)
- Free scans of up to 1,200 words, 6 a day, with no account
- Explainer cards say why a passage was flagged (one per scan free, unlimited with Premium)
- QuillBot says the model leans toward "human" when a result is uncertain
- Document upload in the free tool; Premium adds downloadable, shareable reports
Weaknesses
- Inputs from individual users are used to train QuillBot's models by default; the opt-out exists only in the browser extension settings
- No writing-process replay, LMS integration or school license with student-data terms
- Humanized AI text gets through more often: accuracy fell to 58% on humanized DeepSeek text in a 2025 study
- The AI detector runs on QuillBot's website and app, not in the Chrome extension
Free. What the free version includes: Up to 1,200 words per scan and 6 scans a day, no account needed. Minimum 80 words; one explainer card per scan.
Read the full review
How it works
Paste or upload text and QuillBot returns a 0–100% score split into AI, Mixed and Human. Mixed covers human writing that was refined with tools such as paraphrasers or grammar checkers. Flagged sentences are highlighted, and explainer cards describe why a passage reads as AI. The current model is version 8.4. QuillBot says the model leans toward "human" when a result is uncertain, to reduce false positives, and it tells educators never to act on a score alone. The same engine runs Scribbr's free AI Detector. QuillBot also offers a separate AI image detector.
Accuracy and false positives
QuillBot cites a 99% detection rate based on RAID, a public benchmark from the University of Pennsylvania (Dugan et al., ACL 2024). On the RAID leaderboard, QuillBot catches 99.2% of AI text when the threshold is set for a 1% false-positive rate, and 98.1% when adversarial attacks are included. Vendors submit their own results and RAID sets the threshold, so the figure shows how well the tool separates AI from human text, not its error rate at the default setting. In Paullet et al. (Issues in Information Systems, 2025), the paid detector labeled 35 of 36 human samples correctly (the 36th was too short to scan) and caught all 34 AI samples. A 2025 preprint by Alshammari and Rao found near-perfect detection of DeepSeek text, but accuracy fell to 58% on humanized text. In an October 2024 Decrypt test, QuillBot scored the Declaration of Independence as 100% human.
Languages and non-native writers
QuillBot says the detector supports 20+ languages, including English, Spanish, French and Portuguese, and lets you pick the language or detect it automatically. It publishes no accuracy figures per language and no tests on essays by English learners. Its FAQ warns that formulaic human writing, such as academic definitions, legal copy or templates, can score higher than expected, and that heavily paraphrased text is harder for any detector. For ESL students who use grammar or paraphrasing tools, the Mixed category helps separate light AI editing from AI-written text, but scores on short or formulaic essays deserve extra caution.
Privacy and student data
QuillBot's help center says user inputs are used both to run the service and to train its models. Browser-extension users can switch off text storage and training in settings; Team Plan inputs are never used for training. QuillBot states GDPR and CCPA compliance, certification under the EU-US Data Privacy Framework and a CASA Tier 2 security assessment. It does not publish FERPA or COPPA commitments or a student data privacy agreement. Learneo's terms require users to be of legal age where they live and at least 13, so most high school students fall outside the terms.
Pricing and school licenses
The detector is free with daily limits. Premium costs $19.95 a month or $99.95 a year ($8.33 a month). It removes the word and scan limits, unlocks unlimited explainer cards and report downloads, and includes the paraphraser, grammar checker, AI Humanizer and 25,000 plagiarism-check words a month. The Team Plan costs $69.95–$89.95 per seat per year depending on seat count and adds an admin dashboard. There is no district license and no LMS add-on.
Where others do better
GPTZero and Grammarly Authorship show how a text was written, with typing replay and pasted-text tracking; QuillBot only scores finished text. Turnitin Clarity and Copyleaks plug into school LMSs and come with institutional privacy terms. GPTZero and Pangram were tested in larger independent academic studies, such as Chicago Booth's 2025 study, with false-positive rates under 1%. Sapling and Copyleaks publish longer lists of supported languages with dedicated models.
- Price from
- Premium $8.33/mo billed yearly or $19.95/mo billed monthly
- For schools
- No school license. Team Plan seats cost $69.95–89.95 a year, and QuillBot doesn't train on Team inputs.
- Best evidence
- 99.2% of AI text caught at a 1% false-positive setting on the RAID leaderboard; 0 of 35 human samples flagged in a 2025 study (vendor: 99% detection rate)
- Languages
- 20+ languages incl. Spanish
- Minimum text
- 80 words (300+ recommended)
- Student data
- GDPR, EU-US Data Privacy Framework; inputs used for training by default (never on Team plans)
- Works in
- Web, iPhone
- Age limit
- Learneo's terms require users to be of legal age where they live (18 in most US states) and at least 13
Sources:
- QuillBot – AI Detector
- QuillBot – Premium plans
- QuillBot – Team Plan
- QuillBot Help Center – How does QuillBot use my data?
- QuillBot – Trust Center
- Learneo – Terms of Service
- RAID benchmark – Public leaderboard, TPR at 1% FPR
- Paullet et al., 2025 – Precision check: A critical look at the reliability of AI detection tools, Issues in Information Systems 26, 3
- Alshammari & Rao, 2025 – Evaluating the Performance of AI Text Detectors Using DeepSeek Generated Text, arXiv
- Decrypt, Oct 2024 – AI detectors claim the Declaration of Independence was 98% AI-generated
FREE CHECKERS
4. Scribbr
Learneo
Rating: 3.3 out of 5
- Accuracy: 3.5
- Features: 2.5
- Price and free plan: 4.0
- Privacy and schools: 3.0
Why it got this score: Scribbr offers the simplest unlimited free check with an AI-refined category, on an engine that scores well on RAID. It has few other features, only four languages, and its privacy claims conflict with QuillBot's training policy.
Best for: Students who want unlimited quick checks without an account
Scribbr's free AI Detector is QuillBot's detection engine running inside Scribbr's site; both belong to Learneo. It labels text as AI-generated, human-written & AI-refined, or human-written, with unlimited free checks of up to 1,200 words and no account. Scribbr's paid plagiarism checks include AI detection on the uploaded paper.
What it can do
- Free, no account: Yes
- Flags sentences: Partly
- Spots mixed or edited AI: Yes
- Explains why: No
- Shows how it was written: No
- Plagiarism check: Partly
- Upload files: Partly
- Canvas or Google Classroom: No
- Chrome or Docs add-on: No
- Other languages: Partly
- Independently tested: Partly
- School-grade privacy: Partly
Strengths
- Unlimited free checks of up to 1,200 words, no sign-up
- Separates AI-generated, AI-refined and human-written text
- Runs on QuillBot's detector, which ranks near the top of the public RAID leaderboard
- Pay-per-paper plagiarism check ($19.95 up to 7,500 words) includes AI detection with flagged fragments; no subscription
- Plagiarism submissions are never added to a shared database
Weaknesses
- Detects AI in only four languages: English, Spanish, German and French
- No explanations, writing-process replay, LMS integration or browser extension
- Scribbr's own FAQ says the free detector handles GPT-3.5-era text with "average accuracy" and reserves GPT-4 detection for the paid version
- Privacy messages conflict: Scribbr says it does not store text, while the embedded tool runs under QuillBot's privacy policy, which allows training on inputs
Free. What the free version includes: Unlimited checks of up to 1,200 words each, no account needed.
Read the full review
How it works
The detector on scribbr.com loads from quillbot.scribbr.com and states that QuillBot's privacy policy and terms apply; Scribbr and QuillBot are both owned by Learneo. Paste up to 1,200 words and the tool shows what share of the text looks AI-generated, human-written & AI-refined, or human-written, and marks the flagged passages. Scribbr's paid plagiarism check runs the same AI detection on an uploaded paper, shows the flagged fragments and allows unlimited rechecks. Scribbr frames the tool as a pre-submission check for students.
Accuracy and false positives
Scribbr publishes no accuracy figure for its own detector. Its FAQ cites Scribbr's comparison of detectors, where the best premium tool reached 84% and the best free tool 68%. Walters (Open Information Science, 2023) tested an earlier Scribbr version with a 500-word limit on 126 papers: 88% correct overall, 95% of human papers labeled correctly with 2% false positives, but only 69% of GPT-4 papers caught. Because the current engine is QuillBot's, the closest current evidence is QuillBot's RAID leaderboard entry (99.2% of AI text caught at a 1% false-positive setting) and a small 2025 study in which QuillBot flagged none of 35 human samples.
Languages and non-native writers
Scribbr detects AI in English, Spanish, German and French, so Spanish-speaking students can check work in either language. It publishes no per-language accuracy and no tests on essays by English learners. The AI-refined category matters for ESL students who run drafts through grammar or paraphrasing tools: light AI editing shows up as refined rather than AI-generated. Short or formulaic texts are the most likely to be misjudged by any detector.
Privacy and student data
Scribbr's detector page says submissions stay private and that Scribbr does not store or share the data. The embedded tool's own FAQ says text is used only to generate the report and is not sold or shared with third parties, and that QuillBot's privacy policy applies. QuillBot's help center says user inputs are used to train its models. For plagiarism checks, Scribbr states that submissions are never added to its content database, so other checkers will not match them later. Scribbr publishes no FERPA or COPPA commitments. Learneo's terms require users to be of legal age where they live and at least 13.
Pricing and school licenses
The AI Detector is free. Paid AI detection comes with Scribbr's premium plagiarism check, priced per paper: $19.95 up to 7,500 words, $29.95 for 7,500–50,000 words and $39.95 above that. Each check includes a similarity report, the AI Detector with flagged fragments and the AI Proofreader. There is no subscription and no school or district license. Scribbr also sells an AI Humanizer.
Where others do better
QuillBot uses the same engine but adds explainer cards, sentence highlights and 20+ languages. GPTZero and Grammarly Authorship record how a text was written. Copyleaks and Turnitin integrate with school LMSs. Sapling covers 16 languages with dedicated models. Pangram and GPTZero have stronger results in recent independent studies.
- Price from
- Plagiarism check with AI Detector $19.95 one-time
- Best evidence
- 88% of 126 papers classified correctly in Walters (2023), earlier version; current engine is QuillBot's (99.2% of AI caught at 1% false positives on RAID)
- Languages
- English, Spanish, German, French
- Minimum text
- Longer text works best; 1,200 words per free check
- Student data
- Says it does not store or share text; QuillBot's privacy policy applies to the detector
- Works in
- Web
- Age limit
- Learneo's terms require users to be of legal age where they live (18 in most US states) and at least 13
Sources:
- Scribbr – Free AI Detector
- Scribbr – Plagiarism Checker
- Scribbr – Embedded AI Detector
- Learneo – Terms of Service
- QuillBot Help Center – How does QuillBot use my data?
- Walters, 2023 – The Effectiveness of Software Designed to Detect AI-Generated Writing: A Comparison of 16 AI Text Detectors, Open Information Science 7, 1
- RAID benchmark – Public leaderboard
FREE CHECKERS
5. Sapling
Rating: 3.0 out of 5
- Accuracy: 2.5
- Features: 3.0
- Price and free plan: 3.5
- Privacy and schools: 3.5
Why it got this score: Sapling stands out for its 16 dedicated language models, sentence scores and enterprise-grade data options. Independent evidence is old and weak, including a high false-positive rate on non-native writers, and its 2026 improvements rest on its own testing.
Best for: Texts in other languages, and teams that need an API
Sapling's free AI Detector checks up to 5,000 characters with no account and scores every sentence, not just the whole text. It has dedicated models for 16 languages, including Spanish, all updated in November 2026. The independent studies that tested it date from 2023 and rated it weak, especially on essays by non-native writers.
What it can do
- Free, no account: Yes
- Flags sentences: Yes
- Spots mixed or edited AI: Partly
- Explains why: No
- Shows how it was written: No
- Plagiarism check: No
- Upload files: Yes
- Canvas or Google Classroom: No
- Chrome or Docs add-on: Yes
- Other languages: Yes
- Independently tested: Partly
- School-grade privacy: Partly
Strengths
- Dedicated models for 16 languages, including Spanish, updated in November 2026
- Free with no account: 5,000 characters per check, PDF and DOCX upload, URL checks
- Scores each sentence on the same scale as the overall score, plus a "Mixed" result
- SOC 2 Type II, GDPR, per-request zero data retention and self-hosted options
- Chrome extension that checks text on any web page
- Public changelog of model updates
Weaknesses
- Lowest group of 16 tools in Walters (2023): 65% correct overall and 33% of GPT-4 papers caught
- Scored 62 of 91 TOEFL essays by non-native writers (68%) as AI in Liang et al. (2023)
- No explanations, AI-refined category, plagiarism check, LMS or writing-process features
- Its privacy policy says user content is stored and used to improve the service unless a no-retention setup is used
Free. What the free version includes: 5,000 characters per check with no account, up to 5 files per batch; a free account adds higher rate limits and history. Most reliable after about 50 words.
Read the full review
How it works
Sapling's detector is a Transformer model that estimates, token by token, how likely the text is AI-generated. It returns an overall score, highlights likely AI passages and scores each sentence on the same scale, so a partly AI text can have a low overall score while some sentences score high. It detects the language automatically and uses a dedicated model for each supported language. You can paste text, upload PDF or DOCX files, or check a web page by URL. Shared report links expire after three days. Sapling also runs separate detectors for resumes and product reviews.
Accuracy and false positives
Sapling publishes an 11-page technical report and a model changelog. By its own figures, the September 2026 model cut false positives by 60% compared with the previous one, and a new English model on November 19, 2026 flags less human text in academic abstracts, personal essays and essays by English learners. Independent tests cover only older versions. Walters (Open Information Science, 2023) placed Sapling in the lowest group of 16 detectors: 65% correct overall, 69% of human papers correct with no false accusations (the rest uncertain), and only 33% of GPT-4 papers caught. In Liang et al. (Patterns, 2023), Sapling scored 62 of 91 TOEFL essays (68%) as AI, against 4 of 88 essays by US eighth graders. Sapling warns that short, generic and essay-like texts are the most likely to be false positives.
Languages and non-native writers
Sapling has dedicated models for 16 languages, including English, Spanish, Portuguese, French, German, Italian, Dutch, Polish, Russian, Arabic, Chinese, Japanese and Korean. Text in any other language runs through the English model, which Sapling says is less reliable. The non-English models were updated on November 13, 2026 to flag less human text. Sapling lists fewer false positives on college application essays and short essays by English learners as still in development. Given the 2023 TOEFL results, ESL students should treat high scores on short essays with caution.
Privacy and student data
Sapling holds a SOC 2 Type II certification, states GDPR compliance and supports HIPAA and PCI requirements. Its data documentation says numbers, email addresses and URLs are discarded before processing, data is encrypted, and clients can set retention periods or a no-retention mode; the detector API has a per-request zero-retention flag. Self-hosted and private-cloud deployments are available. The general privacy policy says user content is stored and used only to improve the service, and it does not mention FERPA or COPPA. Sapling says it does not knowingly collect data from children under 13.
Pricing and school licenses
The detector is free up to 5,000 characters per check. Pro costs $25 a month, or $12 a month billed yearly, and raises the limit to 100,000 characters per check and 10 files per batch; it also includes Sapling's writing assistant. Enterprise starts at 10 seats for $15 per seat per month, with custom data retention. The API is metered, with a $5 minimum and up to 200,000 characters per request. Students can ask for a discount. Sapling does not sell a humanizer.
Where others do better
QuillBot and Scribbr separate AI-refined text, and QuillBot and Copyleaks explain why text was flagged. GPTZero, Pangram and Copyleaks have stronger and more recent independent results. GPTZero and Grammarly Authorship show how a text was written. Copyleaks and Turnitin integrate with LMSs and add plagiarism checks.
- Price from
- Pro $12/mo billed yearly or $25/mo billed monthly
- For schools
- No school license is listed. Sapling offers a discounted student plan on request, and Enterprise starts at 10 seats for $15 per seat per month.
- Best evidence
- 65% correct in Walters (2023), lowest group of 16 tools; for its 2026 models Sapling reports 60% fewer false positives than the previous model
- Languages
- 16 languages incl. Spanish
- Minimum text
- About 50 words
- Student data
- SOC 2 Type II, GDPR; zero-retention option; content stored to improve the service by default
- Works in
- Web, Chrome extension
- Age limit
- Not for children under 13 (privacy policy)
Sources:
- Sapling – AI Detector
- Sapling – Pricing
- Sapling – Security
- Sapling – Privacy Policy
- Sapling – Data retention and processing
- Walters, 2023 – A Comparison of 16 AI Text Detectors, Open Information Science 7, 1
- Liang et al., 2023 – GPT detectors are biased against non-native English writers, Patterns 4, 7
- Liang et al., 2023 – Published data and detector results, GitHub
FREE CHECKERS
6. ZeroGPT
Olive Works LLC
Rating: 2.8 out of 5
- Accuracy: 2.0
- Features: 3.0
- Price and free plan: 4.0
- Privacy and schools: 3.0
Why it got this score: ZeroGPT is generous and cheap, and it does not train on submitted text. But independent tests repeatedly caught it flagging human writing, from RAID's 16.9% false-positive floor to famous historical documents, so its scores should not be used to judge students.
Best for: Quick free scans of long texts, not grading decisions
ZeroGPT is one of the most-used free AI checkers: 15,000 characters per scan, no account, an overall AI percentage and highlighted sentences. Independent tests show a high false-positive risk. It has flagged the US Constitution, a passage of Genesis and the Declaration of Independence as AI-written.
What it can do
- Free, no account: Yes
- Flags sentences: Yes
- Spots mixed or edited AI: No
- Explains why: No
- Shows how it was written: No
- Plagiarism check: Partly
- Upload files: Yes
- Canvas or Google Classroom: No
- Chrome or Docs add-on: No
- Other languages: Partly
- Independently tested: Partly
- School-grade privacy: Partly
Strengths
- 15,000 characters per free scan with no account
- Highlights suspect sentences and creates a PDF report for each scan
- Terms state that submitted text is never used to train its models; account data is hosted in Germany under GDPR
- Low-cost paid plans with 100,000+ characters per scan and batch checks of 50–100 files
- API plus WhatsApp and Telegram access
Weaknesses
- On the RAID benchmark (2024), one threshold across all kinds of text gave it a 16.9% false-positive rate, flagging 48% of human-written Wikipedia texts
- Flagged part of the US Constitution and a Genesis passage (88.2% AI) in a 2023 Ars Technica test, and the Declaration of Independence (97.93%) in a 2024 Decrypt test
- No mixed or AI-refined category and no explanation of why text was flagged
- No LMS integration, school privacy terms or official browser extension
Free. What the free version includes: 15,000 characters per scan with no account. ZeroGPT recommends at least 150–200 words. The free version shows ads.
Read the full review
How it works
ZeroGPT, run by Olive Works LLC, analyzes linguistic and statistical signals such as token patterns, burstiness and entropy, then returns an overall AI percentage with the suspect sentences highlighted. Each scan can produce a PDF report. ZeroGPT says it covers 35+ model families and retrains regularly. The site bundles many other tools: a plagiarism checker, paraphraser, summarizer, translator, AI image and video detectors, a chatbot and an AI Humanizer. Paid plans add batch checks and longer inputs, and an API, WhatsApp and Telegram bots offer other ways in.
Accuracy and false positives
ZeroGPT reports 98.4% accuracy and a false-positive rate under 1%, based on its own testing. Independent results vary widely. In Weber-Wulff et al. (IJEI, 2023), ZeroGPT made no false accusations on 18 human or machine-translated texts but missed 36.1% of AI texts, ranking 4th to 6th of 14 tools (59–74% accuracy, depending on the scoring method). Walters (2023) found 87% of 126 papers correct and 79% of human papers correctly labeled. The RAID benchmark (Dugan et al., ACL 2024) found that with one threshold across all kinds of text, ZeroGPT's false-positive rate was 16.9%, and it flagged 48% of human Wikipedia texts, 29% of news articles and 52% of recipes. Ars Technica (2023) showed it flagging part of the US Constitution and a Genesis passage (88.2% AI); Decrypt (2024) got 97.93% AI for the Declaration of Independence.
Languages and non-native writers
ZeroGPT says all its tools support all languages, but it publishes no accuracy figures for any language other than its overall claim. Liang et al. (Patterns, 2023) included ZeroGPT among seven detectors that wrongly flagged 61.3% of TOEFL essays by non-native writers on average; in the study's published results, ZeroGPT scored about half of the 91 TOEFL essays as likely AI, against almost none of 88 essays by US eighth graders. At Universidad Isabel I, Álvaro Bayón (2025) cited a test in which ZeroGPT flagged six of ten human texts. ZeroGPT's own FAQ says polished, formulaic writing can look like AI.
Privacy and student data
ZeroGPT's terms state that private user content is never used to train its own or third-party models. The FAQ says checked text is processed only to produce the result and then discarded, and is not saved, shared or published. Account data (email, name and a hashed password) is encrypted and hosted in Germany under the GDPR, and no-retention modes exist for API and enterprise customers. ZeroGPT publishes no FERPA, COPPA or SOC 2 commitments. Users must be at least 13, and minors need a parent's permission and supervision.
Pricing and school licenses
The free detector covers 15,000 characters per scan. PRO costs $8.99 a month billed yearly ($107.88) or $14.99 monthly, with 100,000 characters per scan, 50-file batches, PDF reports and no ads. PLUS ($13.99/$21.99) and MAX ($19.99/$30.99) add plagiarism words and longer inputs. The EDU plan costs $21.99 a month, billed yearly only, and EXPERT costs $38.99/$60.99. Enterprise is by quote, and the API costs $0.042 per 1,000 words. Every paid plan includes the AI Humanizer. Refunds must be requested within 6 hours (monthly) or 10 hours (yearly) of purchase.
Where others do better
QuillBot and Scribbr offer similar free limits but lean toward "human" when unsure and separate AI-refined text. GPTZero and Pangram have far stronger independent records, with false-positive rates under 1% in Chicago Booth's 2025 study. Grammarly Authorship and GPTZero show how a text was written. Copyleaks and Turnitin integrate with LMSs and offer institutional privacy terms.
- Price from
- PRO $8.99/mo billed yearly or $14.99/mo billed monthly
- For schools
- EDU plan at $21.99 a month, billed yearly, with longer scans and batch uploads. Enterprise plans by quote.
- Best evidence
- 16.9% false positives on the RAID benchmark with one threshold for all text types (Dugan et al., 2024) (vendor: 98.4% accuracy, under 1% false positives)
- Languages
- Claims all languages; interface in 6 incl. Spanish
- Minimum text
- 150–200 words recommended
- Student data
- Text not stored or used for training; GDPR, hosted in Germany
- Works in
- Web
- Age limit
- 13+; minors need a parent's or guardian's permission and supervision
Sources:
- ZeroGPT – AI Detector
- ZeroGPT – Pricing
- ZeroGPT – FAQ
- ZeroGPT – Terms of Use
- Dugan et al., 2024 – RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors, ACL 2024
- Weber-Wulff et al., 2023 – Testing of detection tools for AI-generated text, International Journal for Educational Integrity 19:26
- Walters, 2023 – A Comparison of 16 AI Text Detectors, Open Information Science 7, 1
- Liang et al., 2023 – GPT detectors are biased against non-native English writers, Patterns 4, 7
- Ars Technica, July 2023 – Why AI detectors think the US Constitution was written by AI
- Decrypt, Oct 2024 – AI detectors claim the Declaration of Independence was 98% AI-generated
- Universidad Isabel I, Feb 2025 – Álvaro Bayón on false positives in AI text detection
Best AI detection tools for teachers and schools
What US schools actually use: the detector built into the district's plagiarism system, and tools that replay how a Google Doc was written. Process evidence is harder to argue with than any score.
FOR TEACHERS
1. Revision History
TabWrite LLC
Rating: 3.6 out of 5
- Accuracy: 3.0
- Features: 3.5
- Price and free plan: 4.5
- Privacy and schools: 4.0
Why it got this score: The most informative free writing-process report for Google Docs, with a generous free tier and local-only processing. It flags patterns but makes no AI call, so it supports a conversation rather than a verdict.
Best for: A quick writing-process report for each Google Docs essay
Revision History is a Chrome extension that turns a Google Doc's edit history into a short report and a replay. It shows total writing time, number of sessions, how much text was typed versus pasted, and flags unusual patterns such as a large block appearing at once. It gives no AI score and analyzes documents in the browser without storing their content.
What it can do
- Free, no account: Partly
- Flags sentences: No
- Spots mixed or edited AI: No
- Explains why: Partly
- Shows how it was written: Yes
- Plagiarism check: No
- Upload files: No
- Canvas or Google Classroom: Partly
- Chrome or Docs add-on: Yes
- Other languages: No
- Independently tested: No
- School-grade privacy: Yes
Strengths
- Summary bar at a glance: total editing time, sessions across days, word count and the share of text typed vs pasted
- Replay with large pastes and deletions marked by icons, plus paste size, time and location
- Flags unusual writing patterns (sudden large additions, uneven writing speed, sections with almost no revision) while stating they are not evidence on their own
- Works on Google Docs and Google Slides, including Classroom submissions, and can export a PDF writing report
- Generous free tier (150 documents a month) and a low-cost Plus plan at $32 a year
- Processes documents locally in the browser; the company says it collects no document or student data
Weaknesses
- No AI classifier: it cannot judge whether typed text was copied from a chatbot by hand
- Only sees work written inside that Google Doc or Slides file; Microsoft Word and other editors are not covered
- Unusual-pattern flags can catch normal habits, such as pasting from one's own outline or notes
- No published SOC 2 report; privacy rests on local processing and district DPAs
Free. What the free version includes: Free plan analyzes up to 150 unique Google Docs a month with full replay and analytics; Google Slides analysis does not count toward the limit. A free account (Google sign-in) is needed.
Read the full review
How it works
Google Docs keeps a fine-grained record of edits so collaborators can see each other type. Revision History reads that record in the teacher's browser. Opening a student's Doc (for example from a Google Classroom submission) shows a summary bar with total editing time, number of sessions, word count and the percentage of text typed versus pasted. The replay plays the document from blank to final, with icons on large pastes and deletions. A paste log lists each paste with its time, size and place, and the tool flags unusual patterns. Teachers typically scan the summary first: a 1,200-word essay with four sessions and 95% typed text looks very different from one written in eight minutes with 80% pasted.
Accuracy and false positives
Revision History does not label text as AI-written, so it has no accuracy or false-positive rate to report. Its unusual-pattern flags are only prompts; the Chrome Web Store listing says these patterns "are NOT evidence of any specific behavior and may have perfectly reasonable explanations." Common innocent causes are pasting from a student's own notes, outline or another Doc, or writing offline and pasting later. The report also cannot show the source of typed words, so a student who retypes text from another screen leaves a normal-looking history. The company recommends using the report as a starting point for a conversation with the student.
Languages and non-native writers
The report counts edits and pastes, not word choice, so it works the same for any language. That avoids the bias that text classifiers show against non-native English writers (Liang et al. 2023 found an average 61.3% false-positive rate on TOEFL essays across seven detectors). Teachers of English learners should still expect some pasted text from translation or dictionary tools and ask about it before assuming AI use.
Privacy and student data
The privacy policy (effective July 23, 2026) says the teacher extension analyzes documents entirely in the browser and does not collect, store or transmit document content or metadata. It collects the teacher's name and email at sign-up. The company states the extension is designed to comply with COPPA and offers custom student data privacy agreements to districts; a Student Data Privacy Consortium addendum describes the same local processing. Its newer student product, Byline, does handle student writing under a school DPA, sends it to Google Cloud Vertex AI for coaching, keeps records up to 12 months by default and states that student data is not used to train AI models.
Pricing and school licenses
The free plan covers 150 unique Google Docs a month with full replay and analytics, and Google Slides checks don't count toward it. Plus costs $32 a year billed annually (or $2.67 a month on that plan) and removes the limit. Schools and districts get custom pricing with volume discounts, custom DPAs and implementation support, and IT can install it for every teacher from the Google Workspace admin console. The Chrome Web Store listing shows 300,000 users.
Where others do better
Brisk Inspect Writing adds OneDrive Word support (beta), but it needs Brisk Premium and gives fewer summary numbers. Draftback offers the most detailed, uninterpreted playback but costs $40 a year for teachers after a trial. Turnitin Clarity and Packback combine a process record with an AI score and plagiarism check, but only for writing done inside their own platforms. Revision History cannot score text for AI; teachers who want that need a classifier such as GPTZero or Pangram.
- Price from
- Plus $32/yr
- For schools
- Schools & Districts plan by quote with student data privacy agreements; admins can install it for every teacher.
- Best evidence
- No AI score; reports the Google Docs edit record and flags unusual patterns
- Languages
- Any language (no AI classifier)
- Minimum text
- No minimum; the student must write in the Google Doc or Slides file
- Student data
- Analysis runs locally in the browser; no document content stored; district DPAs available
- Works in
- Chrome extension, Google Docs
- Age limit
- Teacher extension intended for adults (18+); under 18 needs a parent's, guardian's or teacher's permission. Its student coaching product, Byline, runs only through school deployments.
Sources:
- Revision History – Home, features and pricing
- Revision History – Privacy Policy, effective July 23, 2026
- Chrome Web Store – Revision History: Writing Process Visibility for Google Docs & Slides
- Revision History – Getting started
- Student Data Privacy Consortium – Revision History addendum
- School district approved-tools page – Revision History
- Liang et al. 2023 – GPT detectors are biased against non-native English writers, Patterns
FOR TEACHERS
2. Turnitin
Turnitin LLC
Rating: 3.4 out of 5
- Accuracy: 3.5
- Features: 4.0
- Price and free plan: 2.0
- Privacy and schools: 4.0
Why it got this score: Independent studies confirm very few false positives but show it misses much AI text, especially from the newest models. It fits school workflows well, but only institutions can buy it and Clarity costs extra.
Best for: Schools that already license Turnitin
Turnitin adds an AI writing score to the Similarity Report that schools already use for plagiarism checks. Only instructors and admins see it, and only if the school licenses Turnitin Originality and leaves the feature on. Turnitin Clarity is a separate paid add-on where students write inside Turnitin so teachers can replay the draft.
What it can do
- Free, no account: No
- Flags sentences: Yes
- Spots mixed or edited AI: Yes
- Explains why: No
- Shows how it was written: Partly
- Plagiarism check: Yes
- Upload files: Yes
- Canvas or Google Classroom: Yes
- Chrome or Docs add-on: No
- Other languages: Partly
- Independently tested: Partly
- School-grade privacy: Yes
Strengths
- Built into the Similarity Report in Canvas, Blackboard, Moodle, D2L Brightspace, Schoology and Google Classroom, so teachers need no extra tool
- Highlights the passages it scores as AI and splits them into AI-generated and AI-paraphrased text; it also looks for text run through humanizer (bypasser) tools
- Built to avoid false accusations: scores from 1% to 19% show only as "*%" with no highlights, and Turnitin targets under 1% false positives on documents with more than 20% AI
- Very low false positives in independent tests: 0 of 40 human papers flagged in a 2026 VUB study, and 0% false positives in Weber-Wulff et al. 2023
- Clarity adds a writing space with replay, pasted-text flags and an optional AI assistant whose chat history the teacher can read
- Students cannot buy or run it themselves, so a school controls access and settings
Weaknesses
- No individual or teacher plan: AI detection needs a Turnitin Originality license, and Clarity is another paid add-on
- Misses a lot of AI text by design; in a 2026 VUB study it scored essays written fully by ChatGPT Deep Research at 0% AI or below the 20% cutoff
- Students never see the AI report unless the instructor downloads and shares it, which makes disputes hard
- Paraphrase and bypasser detection works only on English; sentence highlights did not line up with the real AI passages in Temple University's test
No free version
Read the full review
How it works
When a school licenses Turnitin Originality, every submission gets an AI writing indicator next to the similarity score. The model drops lists, headings, references, tables and code, then scores each prose sentence. The percentage is the share of qualifying sentences it predicts were AI-generated or AI-paraphrased. Scores of 1–19% show as "*%" with no highlights. Only instructors and admins can open the report, and admins can switch the feature off for the whole school. Turnitin Clarity is a different product: students draft inside a Turnitin writing space, and the teacher gets a Writing Report with playback, a timeline, pasted text matched against the Similarity Report and, if allowed, the student's chats with the built-in AI assistant.
Accuracy and false positives
Turnitin says it keeps false positives under 1% for documents with more than 20% AI writing, and that it accepts missing some AI text to get there: a document flagged at 50% AI "could contain as much as 65%." Independent tests back the low false positives and show the misses. In Weber-Wulff et al. (2023), Turnitin was the most accurate of 14 tools (76–81% depending on scoring) with no false positives, but it missed 10 of 18 manually edited or machine-paraphrased AI texts. Perkins et al. (2024) found it flagged 91% of disguised GPT-4 papers but only 54.8% of their AI content. Temple University's test scored 93% of human texts, 77% of fully AI texts and 43% of hybrid texts correctly, and its highlights did not match the real AI passages. In the 2026 VUB study, it flagged none of 40 human papers, but essays written fully by ChatGPT Deep Research scored 0% or under the 20% cutoff.
Languages and non-native writers
Turnitin's FAQ lists long-form English, Spanish, Japanese, Modern Arabic, Dutch, German and Turkish. Spanish has had its own model since September 2024, updated in May 2026. Detection of AI paraphrasing and humanizer tools is English only. Papers in other languages get no AI report at all. Turnitin says its training data included second-language writers and that it tests each release on more than 700,000 papers written before ChatGPT. It also warns that short, formulaic or repetitive writing is more likely to be flagged, which is common in ESL classes. Clarity's writing space and replay work in any language.
Privacy and student data
Turnitin says it acts as a FERPA "school official" and that students keep ownership of their work. Schools decide whether submissions are added to Turnitin's student paper repository, which is used to check later papers for plagiarism. The AI indicator has no separate repository. In 2026 Turnitin proposed new user-agreement terms on data processing for generative AI; after pushback from universities it postponed the new agreement to September 2027 and stated that it has never used student papers to train generative AI models. Clarity's AI assistant runs on Anthropic's Claude Haiku 4.5, and students and teachers must accept a separate license agreement to use it.
Pricing and school licenses
Turnitin publishes no prices and sells nothing to individual teachers or students. Schools and districts buy licenses by quote. The AI writing indicator comes only with Turnitin Originality; Feedback Studio customers without it must talk to their account manager. Clarity is a paid add-on for Feedback Studio customers, generally available since July 2025, and AI detection inside Clarity also needs Originality. Several US universities, including the University of Arkansas, ran Clarity pilots in spring 2026.
Where others do better
Several US universities have chosen not to use the AI indicator: Vanderbilt disabled it in August 2023, the University of Pittsburgh's Teaching Center turned it off in 2023, Yale's Poorvu Center lists it as disabled, and UT Arlington chose not to enable it because of false-positive risk. Teachers who want a free, instant check get more from GPTZero or Pangram, which also catch more fully AI-written text in recent studies (in VUB 2026, Pangram scored fully AI-written papers at 85.7% AI on average). For writing-process evidence in Google Docs, Revision History (free for 150 Docs a month) and Brisk Inspect Writing (Brisk Premium) work on the documents students already use, while Clarity requires students to write inside Turnitin.
- Price from
- School and district licenses only, by quote
- For schools
- Institution and district licenses only, by quote. Clarity is a separate paid add-on.
- Best evidence
- 0 false positives on human essays in Weber-Wulff 2023 and VUB 2026, but it scored fully AI-written Deep Research essays at 0% in VUB 2026 (vendor: under 1% false positives above 20% AI)
- Languages
- 7 for long-form AI detection, incl. Spanish; paraphrase/bypasser detection English only
- Minimum text
- 300 words of prose (max 30,000 words)
- Student data
- FERPA school official; no generative-AI training on student papers per Turnitin; schools choose whether papers go into the repository
- Works in
- Web
- Age limit
- School-managed accounts only; students reach it through their school's LMS or Turnitin class
Sources:
- Turnitin – AI writing detection capabilities FAQs
- Turnitin – AI writing detection model release notes
- Turnitin – FAQs for administrators using Turnitin Clarity
- Turnitin – FAQs for instructors using Turnitin Clarity
- Vanderbilt University – Why we're disabling Turnitin's AI detector, Aug 2023
- University of Pittsburgh Teaching Center – Using Turnitin with Canvas
- Yale Poorvu Center – Turnitin
- UT Arlington – Turnitin
- Weber-Wulff et al. 2023 – Testing of detection tools for AI-generated text, IJEI
- Perkins et al. 2024 – Detection of GPT-4 generated text in higher education, Journal of Academic Ethics
- Van Vlasselaer, Van Droogenbroeck & Spruyt 2026 – Who wrote this?, IJEI
- Temple University – Evaluating the effectiveness of Turnitin's AI writing indicator
FOR TEACHERS
3. Brisk Teaching
Rating: 3.3 out of 5
- Accuracy: 3.0
- Features: 3.5
- Price and free plan: 2.5
- Privacy and schools: 4.5
Why it got this score: A well-vetted replay that many teachers already have installed, with strong school privacy credentials. It shows how a Doc was written but makes no AI call, and after a 14-day trial it needs a Brisk Premium school license.
Best for: Schools with Brisk Premium that want a Google Docs replay
Inspect Writing is a Premium tool inside Brisk, a Chrome and Edge extension for teachers. It replays how a student built a Google Doc, colors typed, pasted and deleted text, and marks where big pastes happened. It gives no AI score; the teacher reads the replay and decides what to ask the student.
What it can do
- Free, no account: No
- Flags sentences: No
- Spots mixed or edited AI: No
- Explains why: No
- Shows how it was written: Yes
- Plagiarism check: No
- Upload files: No
- Canvas or Google Classroom: Partly
- Chrome or Docs add-on: Yes
- Other languages: No
- Independently tested: No
- School-grade privacy: Yes
Strengths
- Part of the most-installed teacher extension in this group (1,000,000 Chrome Web Store users), so many teachers already have it
- Color-coded replay: green for typed text, yellow for large pastes, red for deletions, with icons that jump to each paste and to long typing pauses
- Works on any Google Doc the teacher can open, including work submitted through Google Classroom; OneDrive Word documents are in beta
- Strong school privacy paperwork: FERPA, COPPA and SOC 2 Type II statements, 1EdTech TrustEd Apps seal, signed state data privacy agreements, and no training on teacher or student inputs
- Sits next to Brisk's feedback and grading tools, so a teacher can check the process and write feedback in one place
Weaknesses
- Inspect Writing is a Premium feature: after a 14-day trial, teachers need a school or district Premium license
- No AI classifier: it cannot tell whether the typed text itself came from a chatbot
- Only covers writing done in that Google Doc or OneDrive Word file; work drafted elsewhere and pasted in shows as one big paste with no history
- A replay cannot show whether a student retyped text from another screen
- The extension asks for broad browser permissions, which some district IT reviews flag
No free version
Read the full review
How it works
Google Docs stores nearly every edit to power real-time collaboration. Inspect Writing reads that history and plays it back like a video. Typed text appears in green, large pastes in yellow and deletions in red. A clipboard icon marks each paste, a speech bubble marks teacher comments, and a fast-forward icon marks pauses in typing. A teacher opens the student's Doc (for example from a Google Classroom submission), clicks the Brisk icon and picks Inspect Writing. What the teacher looks for is a pattern: an essay that grew over several sessions with edits, or a finished paragraph that appeared in one paste a few minutes before the deadline.
Accuracy and false positives
Brisk states that Inspect Writing "doesn't produce an accuracy score." It shows what happened in the document and leaves the judgment to the teacher, so there is no false-positive rate to measure. The record is only as complete as the writing done in that file. A student who drafts in another app and pastes the result, or who uses a fresh copy of the Doc, leaves a short or empty history. That is a reason to ask questions, not proof of AI use. A replay also cannot show where typed words came from. Brisk's own guidance tells teachers not to treat any tool as the final verdict and to start with a conversation.
Languages and non-native writers
Because there is no language model scoring the text, the replay works the same for an essay in English, Spanish or any other language. That makes it fairer to English learners than text classifiers, which studies such as Liang et al. (2023) found flag non-native writing more often. Brisk's blog makes the same point and tells teachers to consider a student's language background before drawing conclusions. One caution: some English learners draft in a translation tool and paste the result, which looks like a large paste in the replay.
Privacy and student data
Brisk says it is fully compliant with FERPA, COPPA and GDPR and holds SOC 2 Type II. It has the 1EdTech Data Privacy Certification and TrustEd Apps seal and has signed the National Data Privacy Agreement in more than 20 states, with district agreements available on request. Brisk states that student and teacher inputs are never used to train, fine-tune or improve AI models, and that it does not store student work after a task is done. Brisk offers an IT approval packet with its privacy policy, state DPA and required configurations. Some district reviews point out that the extension can read pages the teacher visits, which Brisk says it does only when the teacher asks it to act.
Pricing and school licenses
Inspect Writing is part of Brisk Premium. New teachers get a 14-day Premium trial and then move to the Free Forever plan, which keeps more than 20 other AI tools but not Inspect Writing. Schools and districts buy Premium by quote; it adds rostering through ClassLink or SFTP, LTI setup, an administrator dashboard, custom data privacy agreements and professional learning. The Intelligence tier adds curriculum alignment and district guidance features.
Where others do better
Revision History gives a fuller report for the same Google Docs history, and its free plan covers 150 Docs a month: summary stats for time, sessions and typed vs pasted share, flags for unusual patterns, Google Slides support and downloadable PDF reports. Turnitin Clarity and Packback record the process inside their own editors and pair it with an AI score. Teachers who want a direct AI-likelihood score next to the replay need a classifier such as GPTZero or Pangram, since Brisk deliberately gives none.
- Price from
- Brisk Premium for schools and districts, by quote; 14-day trial for new teachers
- For schools
- Brisk Premium for schools and districts (by quote) adds rostering, LMS integration and custom privacy agreements.
- Best evidence
- No AI score; shows the edit record Google Docs already keeps
- Languages
- Replay works in any language; no AI classifier
- Minimum text
- No minimum; the student must write in the Google Doc or OneDrive Word file
- Student data
- FERPA, COPPA, SOC 2 Type II; inputs never used to train AI models
- Works in
- Chrome extension, Google Docs, Microsoft Word
- Age limit
- Teacher tool; student-facing Brisk Boost is offered through schools
Sources:
- Brisk Teaching – Inspect Writing
- Brisk Help Center – How to: Inspect Writing
- Brisk Help Center – Is Brisk free for educators?
- Brisk Teaching – Plans and pricing
- Brisk Teaching – Privacy Center
- Chrome Web Store – Brisk Teaching
- Brisk Teaching blog – How to detect AI writing in student work, Nov 2025
- Brisk Help Center – Getting Brisk approved by your IT department
- Region 16 ESC – Brisk Teaching district review
- Brisk Teaching – AI tools for teachers
FOR TEACHERS
4. Packback
Packback Inc.
Rating: 3.2 out of 5
- Accuracy: 3.0
- Features: 4.0
- Price and free plan: 2.5
- Privacy and schools: 3.5
Why it got this score: Packback pairs an AI score with plagiarism checks and a writing-process report inside the LMS, which few tools do. Its accuracy rests only on its own claims, and students often pay per course.
Best for: College courses that already run writing through Packback
Packback is a college discussion and essay platform with built-in AI coaching, plagiarism checks and an AI Risk indicator. For essays written in Packback, teachers also get a Writing Process Report showing planning, drafting (including pastes), revising and time spent. Through Canvas "Sendback," it can also scan regular file uploads for AI and plagiarism.
What it can do
- Free, no account: No
- Flags sentences: Yes
- Spots mixed or edited AI: No
- Explains why: No
- Shows how it was written: Yes
- Plagiarism check: Yes
- Upload files: Yes
- Canvas or Google Classroom: Yes
- Chrome or Docs add-on: No
- Other languages: No
- Independently tested: No
- School-grade privacy: Partly
Strengths
- Combines three signals in one place: AI Risk (Low/Medium/High), a Uniqueness plagiarism score and a Writing Process Report
- Smart highlighter shows which passages drove a Medium or High AI Risk rating
- Writing Process Report counts sessions, time per phase and paste events for essays drafted in Packback
- Integrates with Canvas, Blackboard, D2L Brightspace and Moodle via LTI 1.3; Canvas Sendback scans file uploads inside SpeedGrader
- Students can preview their originality report before submitting, which turns the check into feedback
- States that no AI provider may train on data processed for Packback and that data from minors is not used for its AI
Weaknesses
- Accuracy figures come only from Packback's own testing (it claims under 0.01% false positives)
- Students usually pay a per-course access fee unless the college or program covers it
- AI Risk gives three levels only and no reason for the rating
- The writing-process record exists only for essays written inside Packback's editor, not for Google Docs or Word drafts
No free version
Read the full review
How it works
Packback runs inside a college course through the LMS. Students post in discussions or write "Deep Dives" essays in Packback's editor, where an AI coach gives feedback but does not write text. Each submission gets an Originality Fingerprint with two parts: a Uniqueness score for matched text and an AI Risk level of Low, Medium or High, based on word choice and sentence patterns. For Medium or High, a smart highlighter shows the passages behind the rating. For essays written in Packback, the Writing Process Report shows time in planning, drafting, revising and reflecting, the number of sessions and paste events. In Canvas, "Sendback" adds the AI and plagiarism checks to ordinary file-upload assignments in SpeedGrader.
Accuracy and false positives
Packback says its AI detection has "the lowest false positive rate of any commercially available tool (<0.01%)," compared with the 1% other vendors advertise. It has since moved to a new detection model and asks instructors to regenerate older reports. Accuracy figures come only from Packback's own testing. Packback's help center warns that concise, formulaic or repetitive human writing can still get a Medium or High rating, and that the indicator is a starting point for a conversation, not a judgment. Its own AI writing coach is designed so that feedback students act on should not raise their AI Risk.
Languages and non-native writers
Packback's help pages describe AI Risk for English-language course writing and name no other supported languages. Its warning that formulaic or repetitive writing is more likely to be flagged matters for English learners, whose writing classifiers misread more often (Liang et al. 2023). The Writing Process Report is a fairer second check for these students because it shows drafting and revision over time rather than judging word choice.
Privacy and student data
Packback's privacy policy says it acts as a FERPA "school official" and keeps subscriber information separated from student posts at the server level. It does not let children under 13 register without parent or school consent. Packback uses OpenAI's APIs and states that no AI provider it uses may train models on data processed for Packback, and that it does not use data from minors for its AI. OpenAI keeps API data up to 30 days for abuse monitoring. Account data is kept until deletion or eight years of inactivity. The policy describes AES-256 encryption at rest but does not cite a SOC 2 report.
Pricing and school licenses
Instructors use Packback free, while students normally buy access per course and per term, with no recurring charge. Many colleges negotiate institutional or inclusive-access deals so students pay nothing. In K-12, Packback partners with AVID, so AVID Elective and Excel teachers can set their course price to $0.
Where others do better
Turnitin is already licensed at many more US campuses and its AI indicator has been tested independently, with very low false positives. GPTZero and Pangram give free, instant checks with more detailed categories and explanations. For K-12 teachers working in Google Docs, Revision History shows the writing process for free, and Brisk Inspect Writing does with Brisk Premium, without moving students to a new platform.
- Price from
- Students usually pay a fee per course; free for AVID K-12 classes
- For schools
- Colleges can cover student fees with an institutional license, and AVID K-12 teachers get it free.
- Best evidence
- Packback reports under 0.01% false positives (vendor figure only)
- Languages
- English-language courses
- Minimum text
- Checks discussion posts and essays; short, formulaic text is more likely to be misflagged
- Student data
- FERPA school official; AI providers may not train on Packback data; minors' data not used for its AI
- Works in
- Web
- Age limit
- Under 13 only with parent or school consent; minors who cannot sign contracts need consent
Sources:
- Packback – Originality and AI detection
- Packback Help – How do I use the AI Risk indicator?
- Packback Help – New AI writing detection model
- Packback Help – How to read and use the Writing Process Report
- Packback Help – Packback Sendback in Canvas
- Packback Help – Submitting an Originality Review assignment
- Packback Help – Is payment required to use Packback?
- Packback Help – K-12 instructors: setting community price to $0
- Packback – Privacy Policy
- Packback – University of North Texas partnership
FOR TEACHERS
5. Draftback
James Somers
Rating: 3.0 out of 5
- Accuracy: 3.0
- Features: 2.5
- Price and free plan: 3.0
- Privacy and schools: 4.0
Why it got this score: A trusted, privacy-friendly replay that many US teachers know, but it now costs money and offers fewer summary stats than free alternatives. Like the other replay tools, it shows process rather than judging AI use.
Best for: A detailed replay of every edit in a Google Doc
Draftback is one of the first Google Docs replay extensions, made by writer James Somers. It plays back a Doc's full revision history like a movie and adds charts of writing time and patterns. It deliberately flags nothing as suspicious and gives no AI score; it is now a paid extension with a 30-day free trial.
What it can do
- Free, no account: No
- Flags sentences: No
- Spots mixed or edited AI: No
- Explains why: No
- Shows how it was written: Yes
- Plagiarism check: No
- Upload files: No
- Canvas or Google Classroom: No
- Chrome or Docs add-on: Yes
- Other languages: No
- Independently tested: No
- School-grade privacy: Yes
Strengths
- Long-established Docs replay tool, listed with 400,000 Chrome Web Store users (500,000+ per the developer)
- Full playback controls: play, pause, rewind, skip ahead or watch at real speed
- A summary page with charts of time spent and writing patterns
- Works on old documents too, because it reads the history Google Docs already keeps
- Privacy-first design: replay data stays in the teacher's browser, and only the user's email is collected for the subscription
- District data privacy agreements and volume pricing available
Weaknesses
- No free plan after the 30-day trial; teachers pay $40 a year or $6.99 a month
- No flags, paste counts or AI score by design, so the teacher must watch and interpret the replay
- Only works on Google Docs where the user has edit permission
- Some teachers in Chrome Web Store reviews complain about the switch to a paid plan
No free version
Read the full review
How it works
After installing the extension, a Draftback button appears in the Google Docs toolbar. Clicking it builds a replay from the edit updates Google Docs already records for real-time collaboration; Draftback itself records nothing new, so it also works on documents written before it was installed. The teacher can play the document from the first keystroke, pause, rewind and jump ahead, and open a page of charts showing when and how long the student wrote. A teacher looking at an essay checks whether it grew in many small steps over days, or whether long, polished paragraphs appeared all at once. Only users with edit permission on the Doc can open the replay; Google Classroom gives teachers that on student Docs.
Accuracy and false positives
Draftback makes no claim about detecting AI. Its developer writes that AI detectors are known for false positives and false negatives, and that Draftback deliberately avoids "flagging" anything, leaving interpretation to the teacher who knows the student. That means there is no error rate to test, but also no shortcut: a teacher has to watch or skim the replay. The replay shows only what happened in that Doc. A student who writes elsewhere and pastes in, or who copies the Doc into a new file, leaves little history, and a replay cannot show the source of words a student typed in by hand.
Languages and non-native writers
Draftback reads edits, not word choice, so it treats every language the same. It avoids the bias text classifiers show against non-native English writers, documented by Liang et al. (2023). Teachers should still ask before assuming anything when an English learner pastes text, since translation and dictionary tools often explain large pastes.
Privacy and student data
The developer states that no document data is collected by or sent to any remote service. Replays are cached only in the user's Chrome browser so they load faster the next time. To manage subscriptions and trials, the extension collects the active user's email and a browser ID after the user consents; student email is stored only if a student uses Draftback themselves. Email addresses are not sold or shared. Draftback says it follows ISO 27000-series security practices and offers district privacy agreements. The terms require users to be 13 or older unless a school or district authorizes and supervises use.
Pricing and school licenses
Draftback lists teacher pricing of $40 a year or $6.99 a month, against a regular price of $100 a year or $9.99 a month. Every new user gets a 30-day free trial without a credit card. Schools and districts can get volume pricing by emailing the developer, with payment by purchase order, card or ACH and custom student data privacy agreements.
Where others do better
Revision History reads the same Google Docs history and adds a summary bar (time, sessions, typed vs pasted share), paste logs, unusual-pattern flags, Google Slides support and PDF reports, with 150 free documents a month. Revision History is free for 150 Docs a month and adds summary stats, and Brisk Inspect Writing (with Brisk Premium) color-codes pastes and covers OneDrive Word documents in beta. Draftback remains the choice for teachers who prefer a plain, detailed replay without any automated flags.
- Price from
- Teacher $40/yr or $6.99/mo billed monthly
- For schools
- Volume pricing for schools and districts by quote (PO, card or ACH), with custom student data privacy agreements and simple onboarding.
- Best evidence
- No AI score; plays back the edit record Google Docs keeps
- Languages
- Any language (no AI classifier)
- Minimum text
- No minimum; needs edit access to the Google Doc
- Student data
- Replays stored only in the user's browser; email collected for subscriptions; district DPAs
- Works in
- Chrome extension, Google Docs
- Age limit
- 13+, unless used under a school's or district's authorization and supervision
Sources:
Can you trust an AI detector?
Detectors have improved a lot since 2023, but every study points the same way: a score is a reason to look closer, not proof. These are the numbers worth knowing before you act on one.
- 26% OpenAI's own AI text classifier caught 26% of AI-written text and wrongly flagged 9% of human text. OpenAI shut it down in July 2023. OpenAI, 2023
- 61% Seven popular detectors wrongly flagged 61% of TOEFL essays by non-native English writers on average, while essays by US eighth graders were classified accurately. Liang et al., Patterns, 2023
- 39% → 22% Simple tricks like adding typos and varying sentence length cut seven detectors' average accuracy on AI text from 39.5% to 22.1%. Turnitin dropped the most. Perkins et al., International Journal of Educational Technology in Higher Education, 2024
- <1% On about 2,000 human-written texts, Pangram, GPTZero and Originality.ai each wrongly flagged fewer than 1%. They differed far more on how much AI text they missed. Jabarian & Imas, NBER working paper, 2025
- 40 of 40 On 160 long English academic papers, Turnitin missed all 40 papers written entirely by ChatGPT, Copyleaks missed 30 and GPTZero 28, while Pangram flagged 39 of them at least in part. None of the four flagged a human paper. Van Vlasselaer et al., International Journal for Educational Integrity, 2026
- 1 in 100 Turnitin aims for fewer than 1 false positive per 100 human-written documents, and shows scores under 20% only as an asterisk because they are the least reliable. Turnitin, AI writing detection FAQ
What this means in practice
- Use detectors on whole essays, not short answers. Accuracy drops fast below a few hundred words.
- Treat a flag as a reason to talk with the student, never as the only evidence for a grade or a misconduct report.
- Be most careful with English learners and very formal writing, which older detectors mistook for AI far too often.
- Ask for the writing process: version history, drafts and notes are harder to fake than a score is to dispute.
Which AI detectors are free?
What each detector lets you check without paying, what the paid plan costs, and what schools and districts can get.
| Detector | Free version includes | Price from | For schools |
|---|---|---|---|
| Pangram | Free account with no card: 20 credits a day (about 2,000 words) plus 3 image scans a day, with file upload and AI-assistance detection. Paid plans start with a 7-day free trial. | Individual $15/mo billed yearly or $20/mo billed monthly | One free month for educators. School and district licenses by quote, with LMS integrations and a zero-retention option. |
| GPTZero | Scans of up to 10,000 characters work without an account. A free account gives 10,000 words a month, a Basic scan with an overall score, one Advanced Scan with sentence highlights and batch upload of up to 3 files. | Premium $12.99/mo billed yearly or $23.99/mo billed monthly | Team plans from $24.99 per seat a month, billed yearly. District plans by quote with SSO and LMS rollout support. |
| Originality.ai | A first scan works without an account. Free use is capped at 100 credits per account, about 10,000 words of AI checks with the 5.0 model, and plagiarism checks use 2 credits per 100 words. File upload needs a paid plan. | Pro $12.95/mo billed yearly or $14.95/mo billed monthly | School plans by quote with a Moodle integration and Google Classroom replay. Extra Pro seats cost about $9–10 a month. |
| Copyleaks | Paste up to 25,000 characters per scan without an account. A free account unlocks sentence-by-sentence detail, highlighted AI phrases and extra monthly credits. | Personal $13.99/mo billed yearly or $16.99/mo billed monthly | Education licenses by quote, priced per student, with LMS integrations. Canvas schools can buy through Instructure. |
| Winston AI | 14-day free trial with 2,000 credits (one credit per word for AI detection) after creating an account; no card needed. Paste at least 500 characters per scan. | Essential $10/mo billed yearly or $18/mo billed monthly | Schools and districts can request quote-based institutional plans. The Advanced and Elite plans add team seats (5 and unlimited) with shared credits. |
| Grammarly | Up to 1,400 words per scan with no account; the result is one overall percentage. Grammarly's own comparison says more scans require a free sign-up. Authorship is free with a free account. | Pro $12/mo billed yearly or $30/mo billed monthly | Grammarly for Education licenses by quote, with Authorship in Canvas and Blackboard and no training on student text. |
| Studera.ai | Up to 1,200 words per check with no account, and a daily number of checks. A free account adds file upload (one file at a time), saved reports and the writing report. | Premium $5/mo or $45/yr | School and district licenses with the class overview, an EU-hosted school profile and a data processing agreement. |
| QuillBot | Up to 1,200 words per scan and 6 scans a day, no account needed. Minimum 80 words; one explainer card per scan. | Premium $8.33/mo billed yearly or $19.95/mo billed monthly | No school license. Team Plan seats cost $69.95–89.95 a year, and QuillBot doesn't train on Team inputs. |
| Scribbr | Unlimited checks of up to 1,200 words each, no account needed. | Plagiarism check with AI Detector $19.95 one-time | – |
| Sapling | 5,000 characters per check with no account, up to 5 files per batch; a free account adds higher rate limits and history. Most reliable after about 50 words. | Pro $12/mo billed yearly or $25/mo billed monthly | No school license is listed. Sapling offers a discounted student plan on request, and Enterprise starts at 10 seats for $15 per seat per month. |
| ZeroGPT | 15,000 characters per scan with no account. ZeroGPT recommends at least 150–200 words. The free version shows ads. | PRO $8.99/mo billed yearly or $14.99/mo billed monthly | EDU plan at $21.99 a month, billed yearly, with longer scans and batch uploads. Enterprise plans by quote. |
| Revision History | Free plan analyzes up to 150 unique Google Docs a month with full replay and analytics; Google Slides analysis does not count toward the limit. A free account (Google sign-in) is needed. | Plus $32/yr | Schools & Districts plan by quote with student data privacy agreements; admins can install it for every teacher. |
| Turnitin | No free version | School and district licenses only, by quote | Institution and district licenses only, by quote. Clarity is a separate paid add-on. |
| Brisk Teaching | No free version | Brisk Premium for schools and districts, by quote; 14-day trial for new teachers | Brisk Premium for schools and districts (by quote) adds rostering, LMS integration and custom privacy agreements. |
| Packback | No free version | Students usually pay a fee per course; free for AVID K-12 classes | Colleges can cover student fees with an institutional license, and AVID K-12 teachers get it free. |
| Draftback | No free version | Teacher $40/yr or $6.99/mo billed monthly | Volume pricing for schools and districts by quote (PO, card or ACH), with custom student data privacy agreements and simple onboarding. |
How we picked and rated the detectors
Every detector gets a score from 1 to 5 on four criteria. Accuracy weighs the most, because a detector that flags honest students does more harm than no detector at all.
- 40 % Accuracy: How well the detector catches AI text without flagging human writing. Independent studies count most; a company's own numbers alone can't earn more than 3.5.
- 25 % Features: Sentence highlights, separate labels for mixed or edited AI text, explanations, writing-process evidence, file upload, plagiarism checks and Canvas or Google Classroom.
- 20 % Price and free plan: What you can check without paying, what the paid plan costs, and whether schools can get a license.
- 15 % Privacy and schools: How the tool treats student text (storage, model training, FERPA, COPPA, SOC 2), the age limit, and how well it fits a US school or district.
We picked the detectors US teachers and students search for and use: the big stand-alone detectors, the free checkers that rank for “AI detector”, and the tools schools run inside Canvas, Google Classroom and Google Docs.
Accuracy scores lean on published, independent research: peer-reviewed studies, university tests and public benchmarks. A company's own accuracy claims are reported as theirs and can't lift the accuracy score above 3.5 on their own. Facts about features, free limits, prices and privacy come from each company's website, help center and policies, and every source is listed under the tool.
Brisk, Revision History and Draftback don't score text for AI at all; they replay how a Google Doc was written. Their accuracy score reflects how far that evidence goes: it shows exactly what happened in the document, but not whether typed text was copied from another screen, so all three get 3.0.
We left out tools whose main product is a “humanizer” for getting AI text past detectors, such as Undetectable AI, even when they also offer a free detector.
Detectors change their models often, so we go back through the list every three to four weeks. Spot a mistake? Let us know and we'll fix it.
Last reviewed:
Report an errorWhen a detector flags a student's essay
Start with the student, not the score. Ask them to walk you through how they wrote the essay, explain a paragraph in their own words, or show their notes and earlier drafts.
Look at the writing process. If the essay was written in Google Docs, Brisk, Revision History or Draftback can replay it; Grammarly Authorship and GPTZero do the same for students who turn on tracking. A long paste with no edits says more than a percentage.
Never base a grade or a misconduct report on a detector alone. Turnitin says its own score shouldn't be the sole basis for action, and the US Department of Education's Office for Civil Rights has warned that false flags can fall hardest on English learners.
Check a text sentence by sentenceTools that help students learn, not just turn things in
Free to try, for teachers and students.
- AI-assisted grading – Grade writing against your rubric, with feedback you can edit.
- Quiz – Check what students actually understood, with written answers graded by AI.
- Study set – Turn class material into flashcards, a quiz and a summary.
- Summarize text – Cut a long reading down to what matters.
AI detectors: common questions
What is the best AI detector?
Pangram has the strongest independent record: in every published test since 2025 it flagged almost no human writing, and it was among the best at catching AI text, especially in mixed and edited papers. GPTZero comes close on false positives and has more features for teachers, such as writing replay and explanations. For free checks without an account, QuillBot and Scribbr are the easiest to start with.
What is the best AI detector for teachers?
It depends on what your school already has. If your district licenses Turnitin, its AI indicator is already in your grading workflow. For a second opinion on a whole essay, Pangram and GPTZero have the best evidence on false positives. And if students write in Google Docs, a replay tool like Revision History (free for 150 Docs a month) or Brisk Inspect Writing (with Brisk Premium) shows how the essay was written, which is stronger evidence than any score.
What AI detector do most teachers use?
In a 2023–24 survey by the Center for Democracy & Technology, 68% of US middle and high school teachers said they regularly use an AI detection tool. Most use the Turnitin license their school already pays for, or a Chrome extension that replays Google Docs: Brisk, whose teacher extension includes Inspect Writing, has about a million users, Draftback about 400,000 and Revision History about 300,000. GPTZero is the most-searched stand-alone detector.
Is GPTZero accurate?
On false positives, yes: it flagged none of 495 human passages in Epoch AI's 2026 test and stayed under 1% in the 2025 Chicago Booth study. It is weaker at catching AI text that has been edited, humanized or written by the newest models. In a 2026 study of long academic papers, it missed 28 of 40 papers written entirely by ChatGPT and got none of the 40 partly AI-written papers right.
Pangram or GPTZero: which is better?
Pangram catches more AI text in mixed and edited writing, and it never trains on what you submit, though neither tool reliably catches text run through a humanizer. GPTZero has more tools for teachers, including writing replay, explanations of why a passage was flagged and a free scan without an account. Both rarely flag human writing. Note that both require users to be 18 or older for individual accounts.
What is the best free AI detector?
QuillBot gives six free checks a day of up to 1,200 words without an account, with separate scores for AI-written and AI-refined text. Copyleaks checks up to 25,000 characters without an account, GPTZero 10,000 characters, and Scribbr has unlimited checks of up to 1,200 words. Grammarly Authorship is free and shows how a document was written. ZeroGPT is free too, but independent tests found it flags a lot of human writing.
Can AI detectors be wrong?
Yes, in both directions. Older tools flagged human writing far too often, and even good detectors miss AI text that has been edited, paraphrased or run through a humanizer. In a 2026 Notre Dame study, Pangram and GPTZero missed over 96% of AI text run through a humanizer, and flagged many abstracts that people had only lightly polished with AI. That is why a score should start a conversation, not end one.
Does Turnitin detect AI writing?
Yes, if your school licenses Turnitin Originality. The AI indicator is visible to instructors, not students, needs at least 300 words of prose, and shows scores under 20% only as an asterisk. It covers seven languages, including Spanish, though paraphrase and humanizer detection works only in English. Independent studies find it rarely flags human writing but misses a lot of AI text from newer models.
Are AI detectors biased against English learners?
Older ones were. In a 2023 Stanford study, seven detectors flagged 61% of TOEFL essays by non-native English writers as AI on average. Newer detectors do much better: in a 2026 study, Pangram, GPTZero, Copyleaks and Turnitin flagged none of 40 papers by non-native English graduate students. Even so, the US Department of Education's civil-rights office has warned schools about relying on detectors that misjudge English learners.
Do AI detectors work in Spanish and other languages?
Several do. Originality.ai covers 31 languages, Copyleaks 30, GPTZero 26 on paid plans, Pangram 24 and Winston 23, all including Spanish. Turnitin detects AI in seven languages, including Spanish. Studera.ai checks 17 languages with phrase lists for each one. Accuracy is usually best in English, and most independent studies have only tested English.
How can a student prove they didn't use AI?
Show the writing process. Google Docs keeps a full edit history, and teachers can replay it with Brisk, Revision History or Draftback. Grammarly Authorship records what was typed, pasted or generated in Google Docs and Word, and the student decides who sees the report. Outlines, notes and earlier drafts help too, and so does being able to explain the essay in your own words.
Can AI detectors spot ChatGPT, Claude and Gemini?
The leading detectors are trained on text from all the major chatbots, and they catch plain, unedited output from them well. They struggle more with the newest models and with text that has been rewritten: in a 2026 study of long academic papers, Turnitin missed every paper written with ChatGPT's Deep Research mode, while Pangram caught almost all of them.
Is ZeroGPT accurate?
Not reliably. In the RAID benchmark (2024), using one threshold for all kinds of text gave ZeroGPT a 16.9% false-positive rate, and it flagged about half of human-written recipes and Wikipedia texts. It has also flagged famous human texts like the US Constitution as AI. It is free, handles long texts and doesn't train on what you paste, but it shouldn't be used to judge a student's work.
How long does a text need to be for an AI detector?
Longer is better. Most detectors need at least 50 to 100 words, Turnitin needs 300, and results get firmer above a few hundred words. Short answers and single paragraphs give unreliable scores with every tool, so check whole essays instead.
How did you rate the detectors, and how often is this updated?
Accuracy counts for 40% of the score, features 25%, price and free plan 20%, and privacy and fit for US schools 15%. Accuracy leans on published independent studies, and a company's own numbers alone can't lift it above 3.5. We recheck every detector every three to four weeks, because models and prices change often.