Modular detection agent combining static analysis with optional editor history (the stronger signal). Classifies likely human, uncertain, likely AI-assisted, or highly likely AI — never a specific model, never absolute proof.
Privacy: Code and optional history are analyzed for the request only and are not stored.
High likelihood with low confidence is a valid outcome — especially on short snippets without editor history.
Paste code
Classifications default to: 0–29 likely human · 30–59 uncertain · 60–79 likely AI-assisted · 80–100 highly likely AI. Absolute “AI generated” labels are intentionally not used.
Large coherent insertions with little follow-up editing are stronger evidence than descriptive variable names. Paste alone is never treated as proof — people paste docs, Stack Overflow, and their own snippets.
Can you tell Claude from ChatGPT? No — model attribution from source alone is unsupported.
Is descriptive naming enough? No — it is a weak linguistic signal only.
What if confidence is low? Prefer interviews, live coding, and code review over the score alone.