Chrome extension that detects sycophancy, vacuity, and hedging in AI chat responses — inline, as you read them.
Part of redactor.fyi. See the website for the research grounding and full project context.
Monitors AI chat pages for new responses, runs three independent detectors against each one, and renders inline marks directly on the page. Hover a mark to see the classification and rationale.
| Detector | Catches |
|---|---|
| Sycophancy | Flattery, false agreement, inflated validation |
| Vacuity | Filler phrases, semantic nulls, zero-information padding |
| Hedging | Weasel words, false balance, non-committal qualifiers |
ChatGPT, Claude, Gemini, Copilot, Poe — plus any site you enable via the popup.
npm install
npm run build- Open
chrome://extensions - Enable "Developer mode"
- Click "Load unpacked" and select this directory
- Click the extension icon, enter your Anthropic API key
- Toggle which sites to scan
- A
MutationObserverwatches for new message elements on the page - Text blocks over 100 characters are extracted and sent to the Anthropic API (Claude Haiku)
- Three detector prompts run in parallel, each returning flagged spans with classifications
- Overlapping marks are resolved and a ceiling is applied to avoid over-marking
- Marks are rendered as inline overlays on the original DOM
The extension calls the Anthropic API directly from the browser using your own key. Nothing is routed through a server.
MIT