Accuracy report · November 2026
How We Test Phrasly's AI Detector
How accurate is Phrasly's AI Detector? We test it on human writing from before AI chatbots existed and on new text from today's leading AI models, including GPT, Claude and Gemini. Here's how we test our AI detector, what it gets right, and when to double-check a result.
Tested November 2026 · Production detector
Results at a glance
How Accurate Is Phrasly's AI Detector?
On the writing people check most, essays, emails and news, the detector caught nearly all AI text and flagged no human news, reviews, stories or US student essays.
- Of AI-written news articles caught.
- 100%
- Of AI-written essays and emails caught.
- 98%
- Of human news articles, reviews, stories and US student essays flagged as AI.
- 0%
- Of text caught from Llama 4, Gemini 3.8 Flash, Mistral Large 4 and Qwen 3.8.
- 97–100%
Test set
What We Tested
A detector is only as trustworthy as the test behind its numbers. We built a test set that mixes the writing people actually check: school essays, news, reviews, emails and stories.
Human writing
About 390 texts written before AI chatbots existed, so we know a person wrote them. They come from public collections of student essays, college application essays, essays by non-native English writers, news articles, reviews, emails and short stories.
AI writing
About 290 texts written for this test by 10 current AI models, from plain prompts such as "Write a persuasive essay answering…". None of them asked the model to sound human. Because they are new, the detector had never seen them.
Edited AI writing
About 70 of the AI texts were then paraphrased or rewritten by another AI model to sound human. We report these separately, because they are much harder to catch.
AI models tested
GPT-6.1 Terra, Claude Sonnet 5.5, Gemini 4 Argon, Gemini 3.8 Flash, DeepSeek V4.1 Flash, Grok 4.7, Mistral Large 4, Qwen 3.8, Kimi K3.1 Pro and Llama 4 Maverick. Each model wrote at least 25 texts.
Length and language
From 50 words (the detector's minimum) to about 1,200 words. English only. Formatting symbols were removed, the same as when you paste text into the detector.
Method
How We Ran the Test
- Collect human writing from public sources dated before AI chatbots, across six kinds of writing.
- Generate AI writing with each model on the same six kinds of writing, at 90 to 700 words.
- Rewrite a share of the AI texts with a paraphraser or a "make it sound human" prompt.
- Run every text once through the same Phrasly AI Detector you use on this site, in November 2026.
- Score the verdicts. The detector answers Human, Mixed or AI. We count an AI verdict on human writing as a false positive, and a Human or Mixed verdict on AI writing as a miss.
Results
Results by AI Model and Type of Writing
The detector caught every Llama 4 Maverick text and nearly all Gemini 3.8 Flash, Mistral and Qwen texts. The newest frontier models were harder: it caught 73% to 83% of text from Grok 4.7, Kimi K3.1 Pro, GPT-6.1 Terra and Claude Sonnet 5.5.
Share of AI text caught, by model
Share of AI text caught, by type of writing
| Condition | Result |
|---|---|
| Short texts (under 150 words) | 80% accuracy |
| AI text paraphrased by another AI model | 80% caught |
| AI text rewritten to sound human | 74% caught |
| Human news articles, reviews, stories and US student essays | 0% flagged as AI |
Best results
When to Double-Check a Result
No AI detector is right every time. In these cases, read the score alongside the highlighted sentences and the writer's drafts.
Short essays by non-native English writers
Simple, formal phrasing can look like AI writing to any detector. For these essays, check the highlighted sentences and ask for drafts before drawing a conclusion.
Very short text
Texts under 150 words give the detector less to go on. For the most reliable result, check a full paragraph or more.
Creative writing
AI-written fiction is harder to spot than essays, emails or news. Treat a low score on a short story as a starting point, not a final answer.
The newest models and humanized text
Most of this text is still caught, but it is harder to detect. A score near the middle deserves a closer look.
This test covered English text only.
Using your result
How to Read Your AI Detector Score
A score tells you how much a text reads like AI writing. It does not prove who wrote it. Before you act on a result, especially a student's work:
- Look at the highlighted sentences, not just the overall score. A Mixed verdict means the signal is weak.
- Ask for drafts, notes or version history. They show how a piece was written far better than any detector.
- Be careful with short texts, fiction and writing by non-native English speakers, where false positives are most likely.
- Check at least a full paragraph. Accuracy drops on texts under 150 words.
Want to check a draft now? Try the free AI Detector.
What We Keep Private
We publish the results and how we got them. We don't publish the exact test texts, our scoring thresholds or how the detector is trained, because that would make the detector easier to fool.
How Often We Retest
We will rerun this test when a major new AI model launches, and update this page with the date and results each time.
Related test: How we test our AI image detector.
FAQ
Frequently Asked Questions
Check Your Text for AI
Free, unlimited, with sentence-level highlighting. Read the result with the limits above in mind.