AI-text detection that tells you why.
One HTTP call returns a calibrated probability that text is AI-generated, and the exact signals that fired: buzzword density, stock phrases, uniform sentence rhythm, chat-UI artifacts. Most detectors hand back a bare number. This one hands back the evidence, so you can build a review UI a human can actually trust.
Try the live demo Read the docs
# your first call, free, no key curl -s https://tellcheck-api.munzzyy.workers.dev/v1/detect \ -H "content-type: application/json" \ -d '{"text": "your text here"}'
How it works
No magic, just measurable signals. The detector extracts deterministic features from the text (buzzword and stock-phrase density, per-language word lists, template rhythm, sentence-length uniformity, punctuation and structural patterns), feeds them to a calibrated model, and returns a probability, not a vibe. On text under 20 words, or text it detects as non-English, it does not score at all: it abstains and says why. That abstention is a feature, not a gap. A detector that refuses to guess on a one-line comment is more trustworthy than one that hands back a confident number anyway.
Live demo
Runs the real API. Paste text and score it. 200 detections a day per IP address.
Why it is different
| Most detectors | Tellcheck |
|---|---|
| A bare 0 to 100 score | Probability plus the specific signals that fired |
| Guesses on any input, short or not | Abstains under 20 words or on non-English text, and says so |
| "Trust us" accuracy | Numbers we measured ourselves, on the benchmark page, honest limits included |
| Monthly subscription | Free: 200 detections a day per IP |
| Account and dashboard to start | No account, no key; one HTTP call |
Rows describing other tools are our own characterization, not a line-by-line audit of any one competitor. Every number on this page that is ours is marked as self-measured.
More than AI detection
The /v1/lint endpoint returns every tell it found, with
exact character offsets: buzzwords, stock phrases, hedges, em dashes, bold-label bullets,
chat-UI residue. It works in 16 languages, so you can underline the evidence in an editor
or a review UI. And /v1/scan adds a text-integrity report to every result:
hidden Unicode (zero-width and invisible characters), bidirectional controls (the
Trojan-Source trick), and homoglyph spoofs (a Cyrillic letter hiding in a Latin word). It
can hand back a sanitized copy too. One call defends an LLM pipeline against both AI slop
and hidden-character injection. See the docs.
Integrate in one call
# no key needed curl -s https://tellcheck-api.munzzyy.workers.dev/v1/detect \ -H "content-type: application/json" \ -d '{"text": "your text here"}'
Batch up to 50 texts per call. Full reference in the docs.
Free
The whole API is free. You get 200 detections a day per IP, no account and no key needed. Batches count one detection per text, and the counter resets at UTC midnight. If that ceiling is too low for what you are building, email Munzzyy1@proton.me and ask.