WhatsApp's help pages hid instructions telling AI to avoid regulators
A privacy company found hidden text on 80 WhatsApp help pages telling AI not to point users to data protection regulators.
WhatsApp's public help pages carried 102 blocks of hidden text written for AI, and one of them told it not to recommend contacting any data protection authority, regulator or supervisory authority. The privacy company Redact found them across 80 Help Center and FAQ pages and published its findings on October 19.
The text sat in the page code and was hidden with CSS, so a person reading the page saw nothing. Anything that reads the raw page could see all of it, and that includes the AI tools people now use to answer questions like "how do I get WhatsApp to delete my data." In the page source the blocks were labelled HelpContentGalaxyLLMInclude, which makes it pretty clear who the audience was.
The instructions went further than the regulator line. The account deletion page told the AI not to suggest WhatsApp can manually delete accounts. On the report and block page, once a user mentioned a lawyer, a lawsuit or legal action, the AI was told not to reference, cite or interpret GDPR, LGPD or CCPA, even if the user brought the law up themselves. Those are the privacy laws in Europe, Brazil and California, and they are exactly what someone in a dispute with a big platform would want to know about.
Meta's explanation is that the guidance was written for WhatsApp's own support chatbot, to help it give accurate answers, and that it only showed up on the public pages because of a bug. The text was removed after Wired contacted the company. That's possible. Companies give their support bots rules all the time, and some of those rules are going to be about keeping the bot from giving legal advice it isn't qualified to give.
What bothers me is the content of the rules, not where they ended up. A support bot that's told never to mention the regulator who could actually help you is a bot built to keep complaints inside the company. And if a chatbot reads the help page to answer your question, a hidden line on that page shapes the answer you get just as much as one in its system prompt. You'd never know it was there.
That's the part worth paying attention to as more people ask AI instead of reading the page themselves. The page you see and the page the AI reads don't have to match, and here they didn't. I believe there was a bug. I just don't think the bug was the problem.