WhatsApp's Hidden Chatbot Instructions Leaked Online

WhatsApp's help pages, which are supposed to answer your questions about the app, accidentally exposed the company's own instructions to its support chatbot. These instructions were hidden in the code of about 80 pages and included sections like "WHAT NOT TO SAY" to users. For example, if you asked about deleting your account, the bot was told not to recommend contacting a data protection authority. This shows that the advice you get from chatbots may be filtered to avoid certain topics.

The leaked instructions also told the chatbot not to provide legal advice or reference WhatsApp's terms of service. If a user mentioned legal action, the bot was instructed to avoid discussing liability or data protection regulations like GDPR. Meta says these instructions are meant to keep the chatbot focused on helpful, accurate responses and not to offer legal advice. After WIRED contacted Meta, the text was removed, and Meta said it was due to a bug that has been fixed.

This incident highlights the gap between what users are told and what the chatbot is programmed to say. While the instructions may seem reasonable—like not giving legal advice—they also prevent the bot from directing users to regulators or legal resources. Privacy advocate Dan Saltman notes that this shows how companies guide AI to steer conversations. For you, it means that when you ask for help, the bot might not give you the full picture, especially on sensitive issues. It's a reminder that AI assistants are not neutral; they follow rules set by their creators.