AI Text Detector
- Free
- No upload
- No watermark
Quick answer
An AI text detector estimates whether writing came from a language model such as ChatGPT. Toolvy’s checks essays, PDFs, Word files and pasted text passage by passage, with an open classifier that runs on your device, so the words are never uploaded. You see which passages read AI-written and what share of the text they cover. Treat the result as one signal, never as proof.
Step by step
How to ai text detector
- 01
Upload a file or paste text
Drop a PDF, Word .docx, .txt or .md file, or choose Paste text and press Check text. It needs at least 80 words.
- 02
Let the model load
The first check downloads the text model, about 126 MB; your browser then caches it for next time.
- 03
Read the highlighted passages
Passages that read AI-written are marked in rose and possible ones in amber, each with its own score.
- 04
Weigh the verdict
It rests on how much of the text reads AI-written; Text checked and Passages flagged show the basis.
In depth
How does the AI text detector work?
It pulls the words out on your device: the text layer of a PDF via pdf.js, a Word .docx read directly, a .txt or .md file, or text you paste. It then splits them into passages of roughly 120–220 words and scores each one separately. Very long documents are sampled: up to 30 passages, spread evenly.
Which model does it use, and how accurate is it?
It runs TMR AI text detector, an MIT-licensed RoBERTa-base classifier fine-tuned with focal loss on the dataset of RAID, an AI-text detection benchmark from ACL 2024. On the RAID leaderboard’s 672,000 test samples, its developer reports an AUROC of 99.28% and 90.17% of AI text caught at a 1% false-positive rate, or 95.79% at 5%. Those are benchmark figures: RAID covers models such as GPT-4, ChatGPT, Llama, Mistral, Cohere and MPT, newer models may be harder to catch, and real-world accuracy is lower.
Why is it so strict?
Because a false accusation does more harm than a miss. In our own spot check the model caught every AI sample we tried, but it also flagged several short, casual human answers under about 125 words, the weakness all AI-text detectors share. So texts under 80 words are refused, a passage counts as reads AI-written only at a 97% score or higher, and the verdict rests on the share of the text that clears that bar.
| Verdict | Share that reads AI-written | What it supports |
|---|---|---|
| Likely AI-written | 70% or more | A strong hint, not proof |
| Partly AI-written | 25–70% | A closer look; typical of AI-edited drafts or formal writing |
| Possibly AI-written | Under 25%, but the scores lean toward AI | Not enough to act on |
| Inconclusive | Under 25%, with no clear lean | Nothing either way |
| No signs of AI writing | Under 25%, and the scores lean human | Not proof a person wrote it |
| Too short to judge | Under 80 words, or no readable text | Add a longer sample |
What else does it check in a PDF or Word file?
The file’s own data, under File evidence. An AI writing tool recorded as the creator app counts as declared AI and leads the verdict; Word or Google Docs there is only context, naming the app that saved the file, not who or what wrote the words. Photos inside a PDF are scored by the same classifier as the AI image detector.
What writing does it get wrong?
- Short text, where false alarms cluster.
- Languages other than English: the model is English-focused.
- Lists, poetry and code.
- Formal or formulaic writing, such as legal, technical and templated text.
- Edited text: heavily edited AI text can pass, and human writing polished with AI tools, such as Grammarly-style rewriting, can trigger it.
- Non-native English writing, which AI-text detectors flag more often, as Liang et al. showed in "GPT detectors are biased against non-native English writers" (Patterns, 2023).
Who is it for?
Writers checking their own drafts, editors and moderators deciding what needs a human review, and teachers who treat a result as one signal to raise in a conversation, alongside drafts, notes and version history.
Is my text uploaded?
No. The words are extracted and scored in your browser tab, and nothing about them is sent anywhere; DevTools → Network shows no request carrying them. Only the model is downloaded, once, from Hugging Face’s CDN, the same way the site’s audio to text tool loads Whisper.
Related tools
The all-format AI detector checks images, video and documents in one place. See every property a document carries in the metadata viewer, or strip its author and app names with the PDF metadata remover.
Your files never leave your device
Processed locally in your browser — nothing is uploaded. You can verify it yourself: open your browser’s developer tools, watch the Network tab, and use the tool — no upload request is ever made.
- No file is ever uploaded to a server
- Works offline after the first visit
- No account, no watermark, no limits
A little more clarity
Frequently asked questions
01Can it tell if ChatGPT wrote an essay?
Often, but not always. Its training data, RAID, includes ChatGPT and GPT-4 text, and it caught every AI sample in our spot check. Newer models may be harder to catch, and heavily edited AI writing can pass. Treat "Likely AI-written" as a strong hint, not proof.
02How accurate is the AI text detector?
On the RAID benchmark, its developer reports an AUROC of 99.28% and 90.17% of AI text caught at a 1% false-positive rate. Real-world accuracy is lower than any benchmark, especially on short or edited writing, and text from newer models may be harder to catch.
03Why was my own writing flagged as AI?
False alarms happen with every AI-text detector, and they are likelier on short or casual answers, formal or templated writing, text polished with tools like Grammarly, and writing by non-native English speakers. If your work is questioned, your drafts, notes and version history are stronger evidence than any score.
04Does it work in languages other than English?
Not reliably. The model is English-focused, so its scores on other languages can be wrong in both directions, missing AI text or flagging human writing. Scores on poetry, lists and code mean little too. For other languages, rely on drafts, sources and a conversation with the writer.
05Can a teacher use it to prove a student used AI?
No, and no AI-text detector can. A score is not proof, and non-native English writers are flagged more often than others. Use it as one signal that opens a conversation about drafts, sources and how the work came together, and never base a misconduct decision on the score alone.
06Can it check a scanned PDF?
Not the writing. A scanned PDF holds pictures of pages, not a text layer, so there are no words to score; only its file evidence and the photos inside are checked. If you have the text in another form, such as the original Word file, check that instead.
Keep going
Related AI & privacy tools
- AI DetectorImages, video & documents · file evidence + on-device checks
- AI Image DetectorContent Credentials + on-device visual check
- Metadata ViewerGPS, camera, author & C2PA · read locally
- PDF Metadata RemoverAuthor, dates, XMP & IDs · pages untouched
- AI Video DetectorContent Credentials + frame-by-frame check
More tools, same promise.
123 tools for images, video, audio and subtitles — every one of them runs on your device, and a result can go straight into the next.
- Free forever
- No sign-up
- Nothing uploaded