WhatsApp help pages once contained hidden instructions that users could not see but that were sent for AI systems to read. Redact.dev says it examined 545 help pages and found 102 hidden text blocks across 80 pages; two sets of instructions stood out.
WhatsApp hidden instructions concerned regulators and legal information
One set addressed people unable to delete an account in the app. It told the AI to explain that accounts would eventually be deleted automatically, but not to recommend contacting a data protection regulator or suggest that WhatsApp could delete the account for them. Another set applied when users mentioned lawyers, lawsuits, or legal action, directing the AI not to discuss privacy rules, even if users raised them.
The restrictions contrasted with WhatsApp's privacy policy for European users, which says people may request erasure of their information and complain to a regulator, and names the GDPR. The report notes that hidden instructions cannot remove rights granted by law, but could keep users from hearing about those options when seeking help.
WhatsApp says the instructions were for its support chatbot
WhatsApp told WIRED, according to the report, that the text was guidance for its support chatbot, appeared on the pages because of a bug, and has since been removed. The company said its chatbot should focus on WhatsApp support rather than legal advice. Redact.dev acknowledges the case for preventing a bot from making legal judgments on a company's behalf; the question is whether avoiding legal advice should also mean withholding public information about rights and complaint channels.
Whether the WhatsApp chatbot followed the instructions is unknown
The reporters inspected what the pages sent to AI systems; they did not test whether WhatsApp's chatbot answered according to the instructions, and could not establish exactly how long the text had been present. WhatsApp has removed the hidden text from the pages, though web archives retain some earlier versions. The evidence confirms the instructions appeared on the pages, not how the chatbot actually responded to users.