Research8 min read
AI detection accuracy, explained without the marketing
A single accuracy number hides a lot. What it measures, what it leaves out, and the three questions to ask any detector before you trust its score.
Maya Okafor
Oct 14, 2026
Trusted by 4M+ teachers, writers and editors
Spot the tells of AI writing
Paste any text and Winston AI scores how likely it is that AI wrote it. It checks for ChatGPT, Claude, Gemini, Llama and other models, and paid plans show you which sentences read as machine-made.
No credit card required.
Try a sample
0 words
- 99.4% accurate
- Free to try
- 14 languages
- No sign-up needed
Used by schools, newsrooms and agencies
- TheLedgerline
- kestrelbay
- Verrow Review
- WRENFIELDACADEMY
- Quillstone
- THE MORNINGLantern
- Pike & Fennel
- OakhavenUniversity
01
From draft to evidence in three steps
Make a free account, add your text, and get a score in seconds. Paid plans show it sentence by sentence.
- Step 1
Create a free account
Sign up with your email in under a minute.
- No card needed to start.
- Folders, titles, authors and labels keep every scan easy to find.
- Step 2
Paste text or upload a file
Paste straight into the editor, or drop in a document.
- Upload .docx, .pdf, .png and .jpg files.
- Built-in OCR reads scanned pages and photos, handwriting included.
- Step 3
Get results in seconds
Every scan comes back with the evidence behind its score.
- A human score from 0 to 100: how likely it is that a person wrote the text.
- A prediction map that highlights each sentence reading as AI or mixed.
- An optional plagiarism check against the web.
- A PDF report to share or keep on file.
Start free
No card needed.
02
Built to be right, and to show its work
A score alone doesn't settle anything. Every scan comes with the evidence behind it, from the sentences that tipped it to a report you can hand to someone else.
The prediction map
Every sentence gets its own reading. Highlights show exactly where the tells are, so you can explain a score instead of just quoting it.
Accurate, with few false alarms
- Accuracy in our tests
- 99.4%
- False positive rate on human writing
- 0.6%
So honest work isn't put on trial.
Checks every major model
Trained on writing from each of these, and updated as new models ship.
- ChatGPT
- Claude
- Gemini
- Llama
- Mistral
- DeepSeek
- Grok
- Copilot
Reports you can share
A PDF with the score, the highlighted text and the scan details, ready to attach to a grade or an email.
Private by default
GDPR compliant, with access limited to you and your team. Your text is never used to train our models.
14 languages
Check essays, articles and reports in the languages your students and readers use.
- English
- Spanish
- French
- German
- Portuguese
- Italian
- Dutch
- Polish
- Turkish
- Arabic
- Japanese
- Korean
- Chinese
- Indonesian
03
Built for the way you grade, publish or write
New models ship every few months, and we retrain to keep pace. Pick your workflow to see where a scan fits in.
For education
Check essays before you grade and see exactly which sentences raised a flag, so any conversation with a student starts from evidence.
Learn more about Winston AI for educationFor publishers & SEO
Screen freelance and agency drafts before they go live, and keep thin, machine-written pages from dragging down your rankings.
Learn more about Winston AI for publishers & seoFor writers
Scan your draft before you hand it in, then share a report with a human score so clients can see the work is yours.
Learn more about Winston AI for writersGrading on Canvas, Google Classroom or Blackboard? See which school platforms detect AI, or check a specific type of writing.
04
Everything you need to keep work original
AI detection and plagiarism checks share one account and one report, so a single pass answers both questions.
AI detector
Paste an essay, article or report and see which sentences read as machine-written, with an overall score in seconds.
Check text for AI99.4%
Average accuracy on our public test set of human and AI-written documents.
How we measure accuracyPlagiarism checker
Compare writing against matching pages on the web and every document in your library. Copied and lightly edited passages are highlighted and linked to their source.
Check for plagiarismGrammar checker
Context-aware corrections for spelling, grammar and punctuation, each with a one-line explanation.
Check grammarAPI
Add detection to your LMS, CMS or review queue. One request returns an overall score plus a verdict for every sentence.
Read the API docsRequest200 OKcurl -X POST https://winstonai.ai/api/v1/detect \ -H "Authorization: Bearer $API_KEY" \ -d '{"text": "Paste the draft here…"}' # Response { "human_score": 22, "verdict": "ai" }Writing tools
Check for AI wherever you work
Run a scan from the browser, the classroom or your CMS. When you outgrow buttons, wire detection straight into your own review flow.
05
Tested on the writing that matters
We test Winston AI against rewritten AI text, documents that mix human and AI writing, and real student essays. Then we publish the method, so you can check our work.
- detection accuracy across our test set
- 99.4%
- false positives on human-written essays
- 0.6%
- documents in our test set
- 120k+
What each test measures
Three ways a detector can get it wrong, and how we check for each.
Adversarial and paraphrased text
AI writing run through paraphrasers, synonym swaps and humanizer tools, to check that light edits don't hide the tells.
Mixed human and AI documents
Drafts where a person and a model each wrote part of the text. We check that the right sentences get flagged, not the whole page.
Real student essays
Thousands of essays written before AI tools existed, to measure how often honest work is wrongly flagged.
Read our methodology
Datasets, test conditions and error rates behind every number.
06
Updated as fast as new models ship
A detector is only as good as its last training run. Ours learns from every major release, so a brand-new model doesn't become a blind spot.
Model coverage
Text models we detect today
| Model | Maker | Versions | Status |
|---|---|---|---|
| GPTOpenAI · 14 versions | OpenAI | 14 | Updated this weekCovered |
| ClaudeAnthropic · 11 versions | Anthropic | 11 | Updated this weekCovered |
| GeminiGoogle · 9 versions | 9 | Covered | |
| LlamaMeta · 8 versions | Meta | 8 | Covered |
| MistralMistral AI · 7 versions | Mistral AI | 7 | Covered |
| DeepSeekDeepSeek · 5 versions | DeepSeek | 5 | Updated this weekCovered |
| GrokxAI · 4 versions | xAI | 4 | Covered |
Retrained weekly. New releases are usually covered within days.
- Trained on human-reviewed writingA large set of documents checked by people teaches the model what genuine writing looks like, which keeps false positives low.
- Every major model familyClosed and open-weight models alike, including the small ones people run on a laptop.
- Retrained every weekNew releases go into the training mix within days, not at the start of next semester.
- Tested by people who don't work hereIndependent reviewers benchmark our results, and we publish the methods alongside the numbers.
07
What teachers and editors say
Professors, editors and freelancers use the reports to have fairer, more specific conversations about AI in writing.
“I needed a detector I could defend in an integrity hearing. Winston AI shows which sentences drove the score, so I can sit down with a student and talk about the writing instead of arguing over a number.”
Maren OkaforEnglish professor“We review more than two hundred freelance pieces a month. Every draft goes through Winston AI before edit, and the highlights tell my editors exactly where to push back without slowing anyone down.”
Tomas LindqvistEditor-in-chief“Clients keep asking whether their content reads as machine-made. I run each page through Winston AI and share the report. It turns a vague worry into a short list of paragraphs to rewrite.”
Priya RamanSEO consultant“Rolling a detector out across eleven schools meant it had to be simple and fair. The low false-positive rate decided it for us, and teacher training took a single staff meeting.”
Daniel HartSchool district manager“My students know I check for AI, and the reports have made those conversations calmer. When a result comes back mixed, the highlights show me which parts to ask about.”
Ana BeltránHigh school teacher“A client once questioned whether I'd used AI on a feature. I sent a Winston AI report the same afternoon. Now I attach one to every delivery, and the question hasn't come up since.”
Sam WhitfieldFreelance writer08
Questions, answered
Straight answers about how detection works, where it falls short, and what happens to your text.
An AI content detector estimates whether a piece of writing came from a person or from a language model like ChatGPT, Claude or Gemini. Winston AI gives you an overall human score plus a sentence-by-sentence map, so you can see exactly which parts read as machine-written. Think of it as a way to ask better questions, not a final verdict.
Language models write in statistically predictable ways: they favor likely word choices, a steady sentence rhythm and familiar structures. Our model is trained on millions of labelled human and AI documents to recognize those patterns, then scores each sentence and the document as a whole. We retrain it as new models are released.
A plagiarism checker compares your text with existing sources to find copied or closely paraphrased passages. An AI detector looks at how the text itself was written, because AI output is usually new text that won’t match any source. They answer different questions, which is why Winston AI also includes a plagiarism checker.
Often, yes. Teachers notice sudden shifts in voice, generic arguments and confident claims without sources, and many use a detector to check their instincts. The fairest approach pairs a scan with the student’s drafts and a short conversation about the work, rather than relying on any single signal.
On our public benchmark, Winston AI classifies 99.4% of documents correctly and wrongly flags 0.6% of human-written ones as AI. Results are less certain on very short or heavily edited text, so every scan shows a confidence level. The full methodology is on our research page.
Search engines say they rank content on whether it’s helpful and reliable, not on how it was produced. In practice, pages mass-produced with AI and published without editing tend to be thin and generic, and those are the pages that slip. Checking drafts before they go live helps you keep quality, and readers’ trust, high.
Your documents are encrypted in transit and at rest, and only you and the people you share a report with can see them. Your text is never used to train AI models, and you can delete a scan or your whole history at any time. Details are in our privacy policy.
Winston AI detects AI writing in 14 languages, including English, Spanish, French, German, Portuguese, Italian, Dutch and Polish. The language is detected automatically, so there’s nothing to set up, and documents that mix languages are scored sentence by sentence.
09
View all postsFrom the blog
Research notes, classroom guides and product news from the people building the detector.
Know what you’re reading
Paste a draft, upload a file or connect your classroom tools. Get a score in seconds, and on a paid plan see exactly which sentences read as human and which don’t.
Free plan available. No credit card needed.
Q3 summary.docx64% human
The numbers came in late again, so I rewrote the summary on the train. Overall, these results underscore a pivotal moment of robust growth across every segment. Ana will ask where that line came from.