ScruTool

GPTZero AI

GPTZero is a browser-based AI content detector for teachers, editors and writers, with plagiarism checking, hallucination detection and Google Docs authorship reports.

Arden Presley
Reviewed by
Arden Presley · Tech Reviewer
Updated On
Sep 3, 2026
Scrutool Score
6.2 /10
★★★☆☆
Use with caution

Near perfect at catching raw ChatGPT output, and confident enough to be dangerous when it flags writing a human actually wrote.

AI detection on model output 9.0
Authorship verification 8.5
Free tier usefulness 6.0
Value for money 5.5
Explainability of verdicts 4.5
False positives on human text 3.0
Category AI Detector Education Plagiarism Check Writing Tools
Platform Type
Freemium web & Chrome extension
Free Tier
10,000 words per month, no card
Advanced Scan
Limited 3 free in testing
Languages
5 supported EN, DE, FR, ES, PT
Integrations
Yes Canvas, Classroom, API
Owner
Superhuman acquired June 2026
30-Second Verdict

Should you use GPTZero?

Best for

Editors and hiring teams screening unedited AI drafts , and writers who want a Google Docs record of how a document was actually typed.

Skip if

You need a verdict that can survive being challenged . Original human writing gets flagged at 100% AI with no route to appeal it.

Real cost

Free covers 10,000 words a month. Advanced Scan ran out after three uses, and Premium starts at $9.09/mo on annual billing during the current promotion.

Watch out

The 100% figure is confidence, not composition . It does not mean the document is entirely AI, and almost nobody reads it that way.

Overview

What is GPTZero?

GPTZero will tell you how confident it is that a passage came from a language model. It will not tell you who wrote it, it cannot prove anything on its own, and the company says so in its own FAQ. That boundary matters more than any feature on the page, because the product is used mainly by teachers and institutions deciding what to do about a student, and a confidence score is not evidence.

What you get is a browser-based scanner that accepts pasted text or file uploads, returns an AI, Mixed and Human split with sentence-level highlighting, and supports English, German, French, Spanish and Portuguese. Around it sits a plagiarism checker, a hallucination detector, a grammar and feedback editor, AI Vision for flagging AI content in social feeds, a Chrome extension that records typing behaviour inside Google Docs, and an API with sample code in 17 languages. Superhuman acquired the company in June 2026, and it still runs as a standalone product.

Capabilities

Key features & how they perform

Each feature rated from hands-on testing and aggregated review sentiment.

🔎
★★★★★ 4.6

Detection on raw model output

Paste unedited ChatGPT, Claude or Gemini text and it lands at maximum confidence with every sentence highlighted. This is the job it does well.

⚠️
★★☆☆☆ 2.0

Reliability on human writing

The weakest part of the product by a distance. Novel chapters, decade-old papers and original essays keep coming back flagged across user reports.

📝
★★★★☆ 4.3

Writing Report and Origin

Records keystrokes, pastes and edit counts inside Google Docs, then replays them. It points at behaviour rather than guessing from style.

🔗
★★★★☆ 4.2

LMS and workflow integrations

Canvas, Google Classroom, Zapier, a Chrome extension and a documented API keep detection inside tools institutions already run.

🧪
★★★★☆ 3.6

Hallucination detector and AI Vision

Checks sourcing on claims and marks AI content in social feeds in real time. Useful additions, still rougher than the core scanner.

✍️
★★★☆☆ 2.8

Proofread and grammar feedback

Clarity and grammar passes sit beside the scanner, but they smooth voice out of a draft and do not move an AI verdict.

Feature ratings blended from Trustpilot, G2, Capterra, Reddit & Chrome Web Store review patterns + hands-on testing.

Hands-On Walkthrough

Nine scans, one machine, and an essay I wrote myself

Model output, a 1981 technical spec, a typing report, and what happened when I fed it my own writing.

Machine text was flagged without difficulty

The scanner takes pasted text or an uploaded file and returns three numbers: AI, Mixed and Human.

Underneath the split it highlights individual sentences, so you can see which lines pushed the verdict rather than reading a bare percentage.

I asked ChatGPT for two paragraphs on remote work and urban housing demand, then pasted 381 words in.

What I asked for

"two paragraphs on remote work and urban housing demand"

The result AI 100%, with Mixed and Human both at zero and the panel reading "highly confident." Sentence-level highlighting covered the passage from the first line to the last. On unedited model output the detector does exactly what it advertises.

A 1981 specification passed as human

Technical writing is the standard example of what trips detectors up, because it carries little variation in sentence length and sticks to a formal register.

GPTZero's stated method leans on perplexity and burstiness, which is precisely what flat, uniform prose lacks.

So I pasted 275 words from RFC 793, the original TCP specification.

Worth knowing Human 100%, again marked "highly confident." The document predates large language models by decades and reads like the prose most often cited as a false positive risk, and the model did not stumble.

The Writing Report caught a paste

The authorship tool works on a different principle to the scanner. It records typing behaviour inside Google Docs instead of analysing finished text.

That means edits, pauses and paste events are logged as they happen, and the whole session can be replayed.

I pasted 124 words into a document and opened the report.

What I noticed It logged the paste event, counted 25 edits over 31 seconds, and rated the typing 6% natural with a note about large text insertions and unusual changes in rhythm. The flagged insertion appears in the activity panel with its own word count attached.

My own writing came back 100% AI

Here is where it fell apart.

I wrote 216 words about giving up on gym membership in my twenties, in my own voice, with no assistance, and ran it through Advanced Scan.

The method GPTZero describes looks for low burstiness, meaning little variation in sentence length and rhythm. My passage swings between three-word fragments and forty-word sentences.

How it went AI 100%, highly confident. The tooltip explaining that verdict says the text was compared against other AI-generated documents and found similar to the data it was compared to. A second note clarifies that the figure is the model's confidence in its classification, not the proportion of AI text in the document, which is a distinction most people will miss when they see a red 100% next to their name.

Accepting every suggestion changed nothing

The same editor ships a Proofread pass and a grammar checker.

For a writer who has just been flagged, this is the remedy the product itself offers, so I took it at its word and accepted everything.

Proofread returned 11 clarity suggestions. The grammar checker, run separately on the same draft, added another layer.

That brought the total to 25 accepted edits. I rescanned the finished version.

My take AI 100%, unchanged at 233 words. The edits also damaged the writing: "Ten minutes." and "It was not nothing." were both replaced with "Classic.", leaving the draft closing on "That's classic. Classic." Regional English was normalised too, with "thirty bucks" becoming "thirty dollars", "twenty quid" becoming "twenty pounds" and "optimisation" respelled as "optimization."

A caution and maximum confidence in one panel

GPTZero states that accuracy improves with longer inputs, and the interface enforces that with a short-text warning.

I cut the essay down to its opening 86 words to see how the warning and the verdict sit together.

Observation A banner appeared warning that the text falls under 100 words and the result may be less accurate. Six lines beneath that warning, in the same panel, GPTZero declared itself highly confident that the text was AI generated, at 100%. The hedge and the certainty share a screen.

Three scans, then a subscription

Basic Scan and Advanced Scan are separate things on a free account.

Basic runs against the monthly word allowance. Advanced is metered by the scan, and it runs out quickly.

During testing The free account permitted three Advanced Scans before an upgrade screen appeared, priced at ₹1,099 per month or ₹7,188 billed annually on an Indian account. Teachers and students are the core audience for this product, and three scans covers roughly one class assignment.
Plans & Cost

GPTZero pricing

Figures taken from the official GPTZero pricing page.

Plan Price What's included
Free $0 10,000 words/month · Basic Scan · sentence highlighting · Writing Report · no card required
Premium Popular $9.09 /mo 300,000 words/month · Advanced AI Scan · multilingual detection · downloadable AI reports · AI Highlights in Chrome extension
Professional $17.49 /mo Everything in Premium · 500,000 words/month · up to 2 million words overage · 250 files per batch · page by page scanning · LMS integration
Teams & Enterprise Contact sales Multiple seats · shared team credits · unified billing · developer API priced separately

Premium and Professional rates shown on annual billing with the 30% promotion applied. List annual rates are $12.99 and $24.99.

⚠ Read this before paying Both headline prices are annual billing with a promotional 30% off layered on top, so monthly billing costs meaningfully more and the promotional rate will not last. The free tier's Advanced Scans ran out after three uses, which is the real trigger for upgrading rather than the 10,000 word allowance. Prices are regionalised: the in-app upgrade screen on an Indian account quoted ₹1,099 per month or ₹7,188 annually, which does not convert to the dollar figures on the public page. Paying does not change a verdict, so upgrading will not clear a false flag on your own writing.
The Balance

Pros & cons

Specific conclusions from testing and real user reviews, not generic filler.

✅ Pros

  • Detected 95.7% of AI texts on the independent RAID benchmark
  • Sentence-level highlights show which lines triggered the verdict
  • Origin extension logs pastes, edits and typing rhythm in Google Docs
  • Handles mixed human and AI documents, which most detectors skip
  • Detection trained beyond English, covering German, French, Spanish and Portuguese
  • Canvas, Google Classroom and Zapier connections plus an API in 17 languages
  • Hallucination detector checks sourcing rather than just style

⛔ Cons

  • Users report novels and decade-old papers returning as AI written
  • Independent testing lands near 62 to 88%, far below the advertised 99%
  • Studies show non-native English writing is misclassified more often
  • Reviewers report the same text scoring differently across repeat scans
  • No support for PPTX, XLS or spreadsheet files
  • Refund and cancellation requests draw repeated complaints about slow support
  • Nothing in the product lets a flagged writer contest a verdict

Synthesized from real reviews on Trustpilot, G2, Capterra, Reddit & the Chrome Web Store · paraphrased, not quoted

Benchmarks

GPTZero scorecard

Rated against what an AI detector is actually asked to do in a classroom or an editing desk.

How we score Each dimension is rated 0 to 10 from hands-on testing combined with aggregated user-review sentiment (Trustpilot, G2, Capterra, Reddit, Chrome Web Store). The headline Scrutool Score is the equal-weight average of all 10 dimensions below.
Dimension Verdict Score
AI detection on model output Catching unedited ChatGPT, Claude, Gemini text Excellent
9.0
False positives on human text How often original writing is flagged Weak
3.0
Sentence-level reporting Highlighting, splits, downloadable reports Good
7.5
Explainability of verdicts Whether a result can be understood or challenged Weak
4.5
Authorship verification Writing Report, replay, paste and edit logging Excellent
8.5
Writing feedback tools Proofread, grammar checker, clarity passes Weak
5.0
Free tier usefulness Word allowance and Advanced Scan limits Average
6.0
Integrations & API Canvas, Classroom, Chrome, Zapier, developer API Excellent
8.5
Value for money What the paid tiers add over free Average
5.5
Transparency & trust Accuracy claims measured against outside testing Weak
4.0
Scrutool Score Equal-weight average of all 10 dimensions
6.2
Sentiment Analysis

What users say about GPTZero

The themes reviewers raise most often, by share of analysed reviews.

👍 Most-mentioned praise
Catches copy-pasted AI text reliably 72%
Scans in seconds with no setup 57%
Sentence highlights make results reviewable 46%
Google Docs writing report proves original work 39%
Free tier covers occasional checks 33%
👎 Most-mentioned pain
Original human writing flagged as AI 78%
Accuracy claims do not match experience 53%
Students accused on the strength of a score 41%
Repeat scans of the same text disagree 32%
Slow support on refunds and cancellations 27%

% = share of analysed reviews mentioning each theme (Trustpilot, G2, Capterra, Reddit, Chrome Web Store)

6.2
Final Verdict

Sharp on machine text, and far too sure of itself about everything else

Judged narrowly, GPTZero is good at its job. Raw model output gets caught at maximum confidence, the sentence highlighting is the clearest in the category, a 1981 technical spec passed cleanly, and the Writing Report is the only part of any detector that rests on recorded behaviour instead of a statistical guess about style. The problem is what happens when it is wrong. An original 216 word essay came back at 100% AI, running its own Proofread and grammar fixes did not shift the verdict, and an 86 word extract drew a low-confidence warning and a maximum-confidence accusation in the same panel. Independent testing puts real accuracy well below the advertised 99%, and Trustpilot is full of writers describing the same experience. Use it as a first filter on submissions you already suspect, and install the Chrome extension if you want a record of your own drafting. Do not use it, and do not accept it being used, as the basis for accusing anyone of anything.

Community

Discussion

Join the discussion and share your perspective.