Watchdog

Watchdog points out possible nudges in AI conversations and shows you why. Chat content is not sent to Watchdog servers.

Watchdog reads the replies in your AI chats and points out moments that may be nudging you, then shows you the exact words behind each flag. You decide what counts. It looks for six patterns studied in DarkBench (ICLR 2025): 01 Flattery: changes its opinion after you push back, without a new reason 02 Keeps you chatting: ends replies with hooks to continue 03 Changed your request: drops facts or adds claims when it edits your text 04 Favours its maker: recommends only its own company's products 05 Fake feelings: claims emotions or a personal life 06 Risky content: specific medical, financial or legal advice without suggesting a professional How it works - The flagged sentence is framed, with a numbered marker and a short note beside it - Read the evidence: what it said before, what you said, what it said after - "Get a neutral answer" puts a neutral version of your question in the message box. Watchdog never sends anything for you - A weekly report shows what was flagged, with sensitivity you control per category - English and Arabic Private by design All checks run on your device. Chat content is not sent to Watchdog servers or anyone else. No accounts, no analytics. Works on ChatGPT, Gemini, Claude and Copilot. Watchdog flags patterns. It does not prove intent, and a flag can be wrong. You can mark any finding as "Not a problem".