Tool / 01
AI content detector
Check text, code or an image for AI origin. You get a calibrated verdict with a confidence interval — and an honest «insufficient data» when the input is too short or the signal is ambiguous. Read the methodology for how we treat numbers on this site.
0 words
How it works
Provenance & metadata
For images we first look for cryptographic provenance (a C2PA content manifest, checked in your browser — the file never leaves your device for this step) and for generator fingerprints in file metadata: PNG text chunks written by Stable Diffusion UIs (A1111 'parameters', ComfyUI workflow), and camera EXIF. A valid C2PA signature is a definitive answer; metadata hints are strong but can be stripped.
Zero-shot perplexity scoring
For text and code we measure how 'surprising' the content is to open language models (a Binoculars-style approximation): AI-generated text is usually more predictable to LLMs than human writing. The score is computed per sentence (or per code line), so you can see which segments look machine-generated, and a bootstrap confidence interval shows how stable the estimate is.
Model-family attribution
When there is enough text, we score it under proxy models for several families (GPT-style, LLaMA, Qwen, Mistral, DeepSeek) and report the most likely source family. This is experimental: families that share training data are easy to confuse, and we say 'unknown' instead of guessing when the margin is small.
External second opinion
Optionally, a commercial detector (Sapling for text, Sightengine for images) is asked for an independent score when configured on the server. It is used as a second vote, never as the only signal.
Honest limits
- · Scores are calibrated zones (confident / likely / abstain), not exact percentages — treat the confidence interval as part of the answer.
- · Short inputs are refused on purpose: under 150 words of text or 40 lines of code, every known method is unreliable.
- · Human writers — especially non-native English speakers — can be flagged as AI. Edited or paraphrased AI text can pass as human.
- · Attribution identifies a model family, not a specific model, and only for English text/code today.