Koji AI Checker
Was this written by AI?
Paste any text and get a quick read on whether it sounds machine-made, plus the exact phrases that give it away.
Turn on the AI model for a much more accurate check. It runs inside your browser, so your text still never leaves your device. One-time download of about 275 MB.
How to check a text
- Paste English textAn essay, post, email or article. 150 words or more gives the best result.
- Press Check textOr press Ctrl + Enter. Nothing is uploaded: the check runs on your device.
- Read the resultYou get a score, the reasons in plain words, and the phrases that sound machine-made.
Not sure what to paste? Try a sample:
0%
AI likelihood
–
HUMANAI
Writing-style signals (0 = human-like, 100 = AI-like)
Most AI-like sentences
Your text, annotated
buzzword
stock transition
personal voice
leans AI (model)
flagged by model
AI detectors can be wrong, especially on short or heavily edited text. Use this as a hint, not as proof.
Technical details
How does this work?
- Writing-style signals. AI tends to write sentences of similar length, lean on buzzwords like "delve" or "seamless", open sentences with "Moreover" or "Furthermore", and avoid contractions and personal voice. These six signals run right in your browser.
- AI model. Press Turn on AI model and two small language models (SmolLM2-135M, base and instruct) are downloaded once and run inside your browser. They score how predictable each word is using the Binoculars method: AI text is unusually predictable. The model counts for 70% of the score and also flags individual sentences. The full app (
python -m aicheck serve) can use larger Qwen2.5 models instead. - How accurate? On a held-out test of 114 answers from the HC3 dataset (people vs. ChatGPT, 2022), the browser model was right 93% of the time (it caught 86% of the AI answers) and never flagged a human answer. Newer AI models write less predictably: on a small set of texts from a 2026 model it caught only 4 out of 10. Treat a "human" verdict as "not obviously AI".
- It is an estimate. Formulaic human writing can look like AI, and AI text edited by hand can pass as human. Short texts under 120 words are unreliable. Never use a single score as proof.
- Second opinion. Compare with GPTZero or Copyleaks. If tools disagree, nobody knows for sure.