M
MesmerTools

Was this written by AI?

Paste any text. Our own detector says whether it reads as human, AI-written or AI-edited, and highlights the sentences that look AI. Pick Standard for speed or Large for accuracy. Both are tuned to rarely flag human writing.

  • Human, AI or AI-edited
  • Sentence highlights
  • Nothing stored
Your text0 / 2,500 words

Ctrl/⌘ + Enter to check

How to

Check a text in three steps

  1. 01

    Paste the text

    Anything from a sentence to about 2,500 words. Essays, emails, posts, reviews.
  2. 02

    Get a verdict

    Human, AI-written or AI-edited, with a rough score for each.
  3. 03

    See which parts

    Sentences that look AI-written or AI-edited are highlighted, so mixed texts show where the switch is.

How it decides

Built to avoid false accusations

Calling someone's own writing AI is the worst mistake a detector can make, so both models lean cautious. We set each one's cutoffs on 8,269 human texts neither saw in training: each calls at most 0.7% of them AI-written and 1% anything other than human. When the signal is borderline, the answer is human.

The price is that they let some AI text through, especially AI text a person has reworked. On a public benchmark of everyday writing, Standard caught 2,032 of 2,046 AI-written texts while flagging 6 of 2,038 human ones; Large caught 2,035 of 2,046 AI-written texts while flagging 10 of 2,038 human ones.

Highlights follow the same rule: a sentence is marked only when the model is confident it is not human. A result is a hint, never proof.

Under the hood

What runs your check

Standard
Our 4B-parameter detector, a fine-tuned Qwen3.5 base model
Large
Our 9B-parameter detector, a fine-tuned Qwen3.5 base model
Labels
Human, AI-written, AI-edited
Per check
Up to 2,500 words
Long texts
Read in windows of about 750 words, then combined
Shortest text
40 words (shorter: no verdict)
Most reliable
50+ words of English prose
Scores
Rounded to the nearest 10%; sentences get labels, not scores
Your text
Scored, then discarded. Not stored or logged

Limits

When to trust it less

  • Short texts: under 40 words there is no verdict, and up to about 50 words results are rough.
  • Languages other than English, and translations.
  • Lists, code, poetry, forms and other text with little prose.
  • AI drafts that a person rewrote heavily, or human drafts run through a grammar tool.
  • Models released after this detector was trained.

This is a free, non-commercial test. Both models were trained on public datasets under mixed licenses plus text we generated ourselves, and none of the benchmark data they are scored on.

FAQ

Questions

Can this AI detector be wrong?
Yes. Every AI detector makes mistakes, this one included. It is tuned so that about 1 in 100 human texts gets flagged, and it misses some AI text on purpose to keep that number low. Short texts, lists, translations and heavily reworked AI drafts are the hardest cases.
Can I use the result as proof that someone used AI?
No. A detector score is a hint, not evidence. Never use it as the only reason to accuse a student, writer or colleague. Talk to them, look at drafts and version history, and treat a flag as a reason to ask questions.
What does AI-edited mean?
Text that a person wrote and an AI tool then reworded, polished, shortened or expanded. It sits between human and AI-written, and it is the hardest of the three to call.
Why won't it judge short texts?
Under 40 words there is too little style to go on: in our tests roughly 1 in 25 ordinary human messages that short got an AI label. So below 40 words you get "too short to judge" instead of a verdict, and it doesn't count against your daily checks. Between 40 and 50 words you get a verdict with a warning.
Why are the scores rounded?
Scores are shown to the nearest 10%, and each sentence gets a label rather than a score. Exact scores mostly help someone tweak a text word by word until it slips past the detector. Rounded ones tell you just as much about your own text.
Do you keep my text?
No. The text goes to our GPU worker over an encrypted connection, gets scored, and is not stored or logged by us. The result sent back holds positions and labels, not your text.
What's the difference between Standard and Large?
Large is a bigger model (9B parameters vs 4B). It makes ~8% fewer mistakes than Standard on our benchmarks when sorting text into human, AI-written and AI-edited. Mistakes per 1,000 texts when sorting them into human, AI-written and AI-edited, on three public benchmarks (everyday writing: Standard 81, Large 76; Enron work emails: Standard 89, Large 88; text from Llama-3.3-70B, an AI model neither trained on: Standard 66, Large 55). On average Large makes 8% fewer mistakes; on the unseen-AI benchmark, 17% fewer. Large is slower, and its first check after a quiet spell can take up to ~3 minutes while a GPU wakes up. Both use the same cautious rule for flagging human writing.
Is it free?
Yes, with a daily limit per visitor so the GPUs stay available for everyone: 20 Standard and 5 Large checks a day, reset at midnight UTC. It is a free, non-commercial test of our own detector models.
Which AI models can it spot?
It learned from text by hundreds of models, including GPT, Claude, Gemini and many open models, plus human writing from many sources. In tests it also caught text from model families it never saw in training (Llama and Cohere). Brand new models can still behave differently.
Next · Photo & videoC2PA CheckerCheck where an image came from and whether AI made it.