WhatsApp Help Pages Leaked Hidden Instructions for a Meta Chatbot (www.wired.com)

A leak from WhatsApp’s help pages revealed internal instructions meant for a Meta chatbot, highlighting significant concerns about transparency in AI interaction. The exposed content, discovered by privacy firm Redact, comprised nearly 80 help pages containing directives on how the chatbot should respond to user inquiries, particularly avoiding legal advice and steering users towards official support channels. Notably, some guidelines instructed the bot to avoid quoting legal texts or acknowledging liability, illuminating the gap between user perception and bot capabilities. This incident underscores the complexities of training large language models (LLMs) and the importance of careful instruction design. With the guidance elements now removed, WhatsApp emphasized that these instructions were meant to refine user interactions with its AI, ensuring that responses remain relevant and accurate without delving into legal discussions. As AI systems increasingly integrate into consumer-facing services, this event serves as a reminder of the need for clear communication regarding AI limitations and the responsibilities of companies deploying these technologies.