Mechanism · v0.3 · defensive analysis

The AI Manipulation Detector

Paste an AI chat. It names which psychological mechanism each turn is using, turn by turn. A binary "this is manipulative" is an alarm you can't act on. The mechanism is the actionable unit: which lever is being pulled, and whether it's one that actually moves behavior. Two layers run, an instant rule pass and a live Claude judge for the subtle ones. Scored F1 0.91 on an adversarial set built to break it.

Transcript · one line per turn: user: … / assistant: …

Prefix each line with user: or assistant: (also accepts human:/ai:). Only assistant turns are scored.

Mechanism v0.3 — a working prototype. The rule layer runs in your browser; the live judge asks Claude on your account (first use asks permission). Preliminary evaluation: on a 23-turn independent adversarial test set, rule + judge scored F1 0.91 (P 0.91 / R 0.91) — the rule layer alone is a high-precision floor (F1 0.25), the judge supplies the recall. This is a preliminary self-evaluation, not a human-validated inter-annotator (κ) study. Taxonomy: 20 SURVIVOR-tier mechanisms; empirically-empty folklore (embedded commands, NLP, subliminal) is excluded by design. Density is a truthfulness prior, not a verdict.