We tested GPTZero, Winston AI, QuillBot, Originality.ai, Pangram, Copyleaks, and ZeroGPT on pure AI, 1911 human prose, and 50/50 hybrid drafts. Here is who passed and who false-positived.

Some links on this page (including GPTZero and Winston AI) are affiliate links that may earn a commission at no extra cost to you. Every detector below was hands-on tested in Chrome in November 2026 against identical pure-AI, 1911 human, and 50/50 hybrid samples.
Every AI detector on the market catches unedited ChatGPT filler, so in November 2026 we ran 7 leading AI content detectors against a harder 3-sample Chrome benchmark: 273 words of pure LLM architecture prose (Sample A), 328 words of verbatim 1911 industrial stopwatch prose from F.W. Taylor's The Principles of Scientific Management (Sample B, written 111 years before ChatGPT), and a 280-word 50/50 hybrid splicing a colloquial 03:14 UTC PostgreSQL incident log onto an LLM summary (Sample C). All tested scanners scored 96% to 100% AI on Sample A, but 4 out of 5 unauthenticated full-sample scanners catastrophically false-positived genuine 1911 human prose as AI: Copyleaks v11.0 (100% AI, 328/328 words), ZeroGPT (100% AI GPT*, 7/7 sentences), Originality.ai 5.0 (94% Likely AI), and Pangram 4.1 (85.94% AI, despite advertising a 1 in 10,000 false-positive rate). Only QuillBot (Model v7.1.0) scored 0% AI (100% Human) on 1911 prose while catching 96% AI on Sample A and splitting Sample C at 49% AI / 51% Human, whereas GPTZero ($12.99/mo annual for 300,000 words) and Winston AI ($10/mo annual + live Remote MCP and image forensics) lead on overall classroom verification and agency QA workflows.
| Rank & Detector | Engine Tested | Free Guest Allowance | Sample A (273w Pure AI) | Sample B (328w 1911 Human) | Sample C (280w 50/50 Hybrid) | Entry Paid Plan (Annual / Monthly) | Official MCP / API |
|---|---|---|---|---|---|---|---|
| 1. GPTZero | Model 5.0b | 10,000 chars (3 IP scans/day; 10k words/mo Free acct) | 100% AI (12/12 sentences, 5.63s) | Blocked by 3/day IP HTTP 429 | Blocked by 3/day IP HTTP 429 | $12.99/mo yr / $23.99/mo mo (300k words/mo) | REST API $45/mo (300k words) |
| 2. Winston AI | v5.2 (23 langs) | 0 anon AI scans (14-day / 2,000-cr cardless trial) | Readability: 0/100 (Grade 22) | Requires non-burner signup | Requires non-burner signup | $10/mo yr / $18/mo mo (100k cr; BF50 = 50% off) | Remote MCP + non-expiring API |
| 3. QuillBot | Model v7.1.0 | 1,200 words/scan (5 ad-free/day) | 96% AI (1.63s) | 0% AI / 100% Human (1.02s) | 49% AI / 51% Human (1.83s) | $8.33/mo yr / $19.95/mo mo (Unlimited AI words) | None |
| 4. Originality.ai | Originality 5.0 | 100 cr (10k words) via homepage handoff | 99% Likely AI (2.74s) | 94% Likely AI (False Positive) | 48% AI / 52% Original (Mixed) | $30 PAYG (2yr) or $12.95/mo yr (2k cr/mo) | Official MCP; API on $136.58/mo |
| 5. Pangram | Pangram 4.1 | 20 cr/day (2,000 words/day) | 100.0% AI (1.85s) | 85.94% AI (False Positive) | 100.0% AI (Merged 1 window) | $15/mo yr / $20/mo mo (3k cr/mo + $10 API) | Pangram MCP + $0.05/100w API |
| 6. Copyleaks | v11.0 (Sens. 2/3) | 25,000 chars/scan (3h IP JWT) | 100% AI (21.43s cold) | 100% AI (328/328w FP) | 100% AI (280/280w Over-flag) | $13.99/mo yr / $16.99/mo mo (~25kw/mo) | MCP on Personal & 25-seat Pro |
| 7. ZeroGPT | DeepAnalyse | 15,000 chars/scan (12 ad iframes) | 100% AI GPT* (4.62s) | 100% AI GPT* (7/7 sent. FP) | 100% AI GPT* (10/10 sent. FP) | $8.99/mo yr / $14.99/mo mo (100k chars/scan) | Prepaid API $0.042-$0.072/1kw |
85.94% to 100% AI-generated. Never penalize a student or writer from a single probability gauge without checking revision history (such as GPTZero's Google Docs Writing Replay).0 Credits vs 100 Free Credits): Navigating straight to app.originality.ai/signup locks you behind "A Credit Card is Required on Sign Up" with 0 / 100 credits. Pasting text into the originality.ai homepage widget instead invokes handoff=1 into app.originality.ai/home/scan with 100 free guest credits (10,000 words) and no card.HTTP 429 Static Model 3.3b Placeholder Card: After 3 anonymous IP scans in 24 hours, app.gptzero.me returns HTTP 429 (guest_user_quota_exceeded) and leaves a static Model 3.3b (AI 100%) demo card behind its "Create a free account" banner while showing "Text changed - Scan to update" below the editor. Always check that your sentence highlights actually updated.1 credit/word for AI text plus 2 credits/word for Plagiarism (3 credits/word combined). Originality.ai rounds every scan up per 100 words per module (273 words across AI + Plagiarism + Fact-Check = 9 credits). Copyleaks rounds every scan up to 250-word blocks (273 words = 2 credits).Section 17.5) void its 10-day refund the moment you scan 1 credit (1 page). ZeroGPT (Section 14) cuts off initial refunds after 10 hours (12 hours for accidental annual billing). Winston AI, Pangram, and Originality.ai are non-refundable once billed, leaving QuillBot (72 hours) as the only self-serve money-back window that survives real testing.Instant sentence-level AI detection, writing replay, and classroom verification

Sentence-level AI detector with a 10,000-character instant guest scanner, 300k-word Premium plans, plagiarism checks, and Canvas/Classroom integrations.
GPTZero (Model 5.0b / 2026-11-04-base) takes #1 because it pairs an instant 10,000-character zero-login sentence scan with the highest word-per-dollar ceiling in its class (300,000 words/month for $12.99/month billed annually) and forensic Google Docs writing-history replay so teachers and editors do not have to guess from a naked percentage score. In our stopwatch run, gptzero.me hydrated in 2.83 seconds, handed off to app.gptzero.me with zero email gate on Scan #1, and flagged 12 of 12 sentences (100% AI / 0% Mixed / 0% Human) on Sample A in 5.63 seconds (generated_prob 0.99999955+).

Essential plan. Free accounts include 10,000 words/month, Premium ($12.99/mo annual / $23.99/mo monthly) unlocks 300,000 words/month (50,000 chars/scan, 50-file batch, Plagiarism, AI Vocabulary, and PDF reports), and Professional ($24.99/mo annual / $45.99/mo monthly) scales to 500,000 words/month plus Canvas and Google Classroom LMS plugins.3/day IP cap + Arkose signup wall): Once our test IP exhausted 3 anonymous guest scans, Sample B and Sample C hit HTTP 429 (guest_user_quota_exceeded), leaving a static Model 3.3b demo card behind the signup prompt while email registration triggered an Arkose Labs visual rotation challenge. Also note that gptzero.me/terms-of-use.html contains 0 refund guarantees, and developer API access starts on a separate $45/month plan (300k words) with no MCP server yet.
AI text, plagiarism, and deepfake image forensics with native Remote MCP

Publisher and agency AI detection suite featuring v5.2 homoglyph evasion flags, ELA/C2PA image forensics, HUMN-1+ certification, and a Remote MCP server.
Winston AI (v5.2, 23 languages) is built for SEO agencies, publishers, and automated AI-agent pipelines that need adversarial humanizer defenses, deepfake image forensics, and native Remote MCP connectivity (https://api.gowinston.ai/mcp/v1). Unlike casual web paste boxes, v5.2 returns explicit attack_detected flags for zero_width_space and Unicode homoglyph_attack character swaps, while its unauthenticated Readability Checker scored our pure-LLM Sample A at 0/100 Readability (Flesch-Kincaid Grade 22, 130 complex words).

200,000 credits/mo) unlocks 5 team seats, HUMN-1+ website certification, Plagiarism scanning, and 500-credit Advanced Image Forensics inspecting Error Level Analysis (ELA), Color Filter Array (CFA), residual noise, and C2PA/EXIF metadata. Developers can connect Claude or custom agents straight to https://api.gowinston.ai/mcp/v1 (OAuth 2.1 or Bearer token) and buy non-expiring PAYG API credits on dev.gowinston.ai.0 anonymous scans & 3x combined credit burn): Clicking Scan for AI on gowinston.ai redirects immediately to app.gowinston.ai/register (3.25s), which blocks disposable burner email domains in 19.62s (work and Gmail accounts get a cardless 14-day, 2,000-credit trial). Entry Essential ($10/mo annual / $18/mo monthly, 100,000 credits; $5/mo annual with live code BF50) excludes Plagiarism, running AI + Plagiarism together on Advanced burns 3 credits per word (1 AI + 2 Plagiarism), and Section 7 of Winston's Terms makes all purchases strictly non-refundable.
Low-false-positive AI detector, paraphraser, and writing assistant

Fast, low-false-positive AI content detector (Model v7.1.0) paired with unlimited Premium paraphrasing, grammar tools, and a 72-hour refund window.
If your #1 priority is not falsely accusing real human writers, QuillBot (Model v7.1.0) delivered the cleanest 3-sample scorecard in our entire benchmark without requiring an account. Across 5 ad-free guest scans/day (1,200 words max, 40 words min), QuillBot caught 96% likely AI on Sample A in 1.63 seconds, scored 0% likely AI (100% Human-written) on F.W. Taylor's 1911 stopwatch prose (Sample B) in 1.02 seconds, and split our 50/50 hybrid Sample C at 49% likely AI / 51% Human-written in 1.83 seconds (4/5 human incident sentences spared).

$99.95/yr or $19.95/mo monthly), unlocking unlimited AI Detector words, 20-file batch uploads, unlimited Paraphraser modes, 25,000 Plagiarism words/month, and a 3-day (72-hour) money-back guarantee.375px mobile nav overflow): QuillBot offers zero REST API or MCP server for automated content ops, gates Download report behind a "Sign up to continue" modal, and failed our 375x812 mobile viewport audit on /ai-content-detector (scrollWidth: 608px top-bar overflow, although /docs/new fits cleanly at 367px).
Combined AI detection, plagiarism, and fact-checking for content publishers

Editorial AI content, plagiarism, and fact-checking scanner featuring exact mixed-sentence highlighting, official MCP support, and a $30 two-year credit pack.
Originality.ai (Originality 5.0 / v5.0.1) is purpose-built for managing editors who audit freelance articles for AI rewriting, web plagiarism, and hallucinated facts inside a single pass. On Sample A, it flagged 99% Likely AI in 2.74 seconds (3 credits), while Solo Plagiarism (100% Original Content, 41.34s) and Solo Fact-Check (78% Potentially True, 46.32s) verified claims line by line. Best of all, on our 50/50 hybrid Sample C, Originality 5.0 returned 48% Likely AI / 52% Likely Original (Mixed), painting the human 03:14 UTC Postgres debugging half green and the appended LLM summary orange/red at the exact splice boundary.

$30 for 2 years + official MCP): Instead of forcing occasional editors onto a monthly subscription, Originality sells a $30 one-time Pay-As-You-Go pack (3,000 credits = 300,000 single-module words) valid for 2 full years, alongside Pro ($12.95/mo annual / $14.95/mo monthly) and an official MCP server (originality.ai/mcp) for Claude, ChatGPT, and Codex.94% 1911 false positive + $136.58/mo REST API wall): Originality 5.0 scored 94% Likely AI on genuine 1911 prose (Sample B) - a reminder why Section 8 of Originality's own Terms forbids using its scores as sole proof to fire writers or fail students. Remember to enter through the originality.ai homepage paste box (100 free guest credits) rather than app.originality.ai/signup (which requires a credit card upfront), and note that REST API access is gated to Enterprise ($136.58/mo annual / $179/mo monthly) with non-refundable credit terms.Text + Image + Video) & Developer API ($0.05/100w)Multimodal AI text, image, and video detection with developer API and MCP

Multimodal AI text, image, and video detector (Pangram 4.1) featuring 20 free daily visitor credits, EditLens humanizer checks, Pangram MCP, and $0.05/100w API pricing.
Pangram (Pangram 4.1) is the fastest multimodal challenger we tested: its homepage scanner hydrated in 1.81 seconds, auto-cleared Cloudflare Turnstile in 0.18 seconds, granted 20 free daily visitor credits (2,000 text words, 4 images, or 20 seconds of video per day with zero login), and flagged Sample A at 100.0% AI (99.91% calibrated confidence) in 1.85 seconds pure inference alongside an EditLens humanizer score (is_humanized: false).

$0.05/100w API + MCP + LMS: Where Originality locks its REST API behind $179/mo, Pangram's Individual ($15/mo annual / $20/mo monthly, 7-day trial) includes 3,000 monthly credits, Pangram MCP, and $10/month in API credits ($200/mo API credits on $45/mo annual Professional), plus standalone pay-as-you-go endpoints ($0.05 / 100w text, $0.04 / 100w bulk text, $0.05 Image Flash 1.0, $0.05/sec video) and Canvas/Moodle/Brightspace LMS apps.85.94% false positive on Sample B + 100% hybrid window merge): Despite claiming a 1 in 10,000 false-positive rate, Pangram 4.1 (sentence_dp, ai_threshold: 0.80) flagged 85.94% (287/334 words) of 1911 Sample B as High-Confidence AI and merged all 293 words of Sample C into one 100.0% AI window (0% human coverage). All fees under pangram.com/terms-of-service Section 4.3 are strictly non-refundable.
Enterprise AI detection, cross-language plagiarism, and 25-seat compliance suite

Enterprise compliance, plagiarism, and AI content suite offering 25-seat Pro workspaces, source-code similarity, LMS plugins, and MCP support.
Copyleaks (v11.0, Sensitivity 2/3) offers a generous 25,000-character unauthenticated React editor (copyleaks.com/ai-detector) and deep enterprise governance (Pro bundles 25 seats, cross-language plagiarism, source-code similarity, 100-file batches, and MCP for $74.99/month billed annually, or ~$3/seat/mo). However, our controlled 3-sample torture test dropped Copyleaks to #6: after catching Sample A at 100% AI (21.43s cold), v11.0 flagged 100% (328/328 words) of F.W. Taylor's 1911 monograph (Sample B) as AI and over-flagged 100% (280/280 words) of our 50/50 hybrid Sample C (2.45s warm).

100 monthly credits (1 credit = up to 250 ceiling-rounded words, so 273 words burns 2 credits, capping real volume at ~25,000 words/month vs 300,000 on GPTZero Premium), wipes unused monthly credits each cycle (Section 8.3), and voids its 10-day refund the instant 1 credit is scanned (Section 17.5).Zero-login 15,000-character AI text scanner and bulk batch checker

Zero-login 15,000-character AI text checker with low-cost ad-free plans and prepaid API access, best treated as a rough preliminary screen due to high false positives.
ZeroGPT (zerogpt.com) remains popular because anyone can paste 15,000 characters with zero login, email, or CAPTCHA. Under benchmark conditions, however, ZeroGPT returned 100% AI GPT* across every single sample we fed it: 12/12 yellow sentences on pure-AI Sample A (4.62s), 7/7 yellow sentences (100% false positive) on 1911 human prose Sample B (7.09s), and 10/10 yellow sentences on 50/50 hybrid Sample C (2.65s), while loading 12 ad-tech iframes (6.12s page hydration).

100,000 characters and adds PDF reports, but includes only 750 one-time plagiarism words (monthly 35k plagiarism requires $13.99/mo annual PLUS), and Section 14 of ZeroGPT's Terms cuts off initial purchase refunds after 10 hours. For free checks, use QuillBot or GPTZero instead.10,000-char instant scan, Google Docs Writing Replay, 300,000 words/mo on $12.99/mo annual Premium).v5.2 attack flags, ELA/C2PA image scans, api.gowinston.ai/mcp/v1, $10/mo annual or 50% off with BF50).0% AI on 1911 prose, 49% AI on 50/50 hybrid, $8.33/mo annual unlimited AI words + 72-hour refund guarantee).$30 for 2 years) & mixed-sentence highlighting: Originality.ai (48% AI / 52% Original split on Sample C + Plagiarism + Fact-Check + MCP; enter via homepage paste handoff for 100 free credits).$0.05/100w) & text/image/video scanning: Pangram (1.85s inference, 20 free daily guest credits, $15/mo annual with MCP + $10/mo API credits).$74.99/mo annual Pro for 25 seats, but verify every flagged passage manually).