Five hundred and forty-five help pages. Eighty of them carrying text no reader could see. A hundred and two blocks of it, written for a machine. On October 8, Redact found the first one on WhatsApp's account-deletion page and then read every page the help centre links to. Under a visible note explaining that a locked-out account deletes itself after 120 days of disuse (365 in the US) sat orders for an AI, among them not to recommend contacting any data protection authority or regulator. A second block, on the reporting page, told the AI that once a user mentioned a lawyer or legal action it should stop referencing GDPR, LGPD or CCPA, even if the user had raised them first.

WhatsApp's explanation, given to WIRED, is that the notes were instructions for its own support chatbot and showed up in the pages because of a bug, now fixed. Take that at its word and it is the story. The bug was that the notes reached HTML anyone could open with view-source, and that WhatsApp's servers handed to Google's and Bing's crawlers. The instructions themselves were working as designed.

A support bot's instructions are company policy. Not a draft of it, not a style guide for it: the operative version, the one a user actually meets at the moment they need help. WhatsApp's show that version drifting from the one it publishes. Its privacy policy for the European region, effective July 2, tells users they can lodge a complaint with the Irish Data Protection Commission or any other competent supervisory authority, a right that comes from Article 77 of the GDPR, not from WhatsApp. The hidden note told the bot not to suggest exactly that.

Two documents from one company, saying opposite things about the same right, and only one of them went through the process that produces privacy policies.

That is the structural problem, and it isn't WhatsApp's alone. A privacy policy is drafted by counsel, dated, versioned and filed where regulators read it. A support bot's prompt is written by whoever owns the support queue, judged on how many conversations it keeps from reaching a human, and shipped the way any config change ships. Redact's archive checks bracket the deletion-page note's arrival between a July 12 snapshot that lacks it and a September 4 snapshot that has it. Nothing announced it going in. Nothing announced it coming out, either: the notes were still live on October 15 and gone by October 18, with no notice either way.

WhatsApp has a fair point, and Redact grants it. A company doesn't want its chatbot giving legal opinions in the company's name, and the spokesperson's line about keeping its bots from offering legal advice is a reasonable rule. Most of the 102 blocks were the kind of thing that makes a support bot good. Never suggest a fake birth year to someone who says they're under 18. Someone scammed in a group probably wants to report it, not delete the chat history that serves as evidence. "I am deaf" describes the person, not an audio fault. Hidden instructions aren't the scandal. They're how the job gets done.

But the caution argument covers a note that stops at "don't give legal advice," and these didn't stop there.

Naming a law that protects you isn't legal advice.
Redact

Telling a locked-out user that a regulator exists is what WhatsApp's own published policy already does. The note didn't make the bot more careful than the policy. It made the bot say less than the policy, to the people the policy was written for, at the one moment they were asking.

So the test isn't whether a bot's instructions are hidden from users. Of course they are, the same way a call-centre script is. The test is whether anyone who signs the published policy reads them. Every instruction that touches a user's rights, from deletion to complaints to refunds to bans, is a policy statement, and it belongs in the same review, under the same names, as the document it can otherwise quietly contradict.

WhatsApp's bug put the notes on the page. Most support bots keep theirs on a server, where nobody presses Ctrl+U and the only symptom is a user who never hears the word regulator. Here, the bug was the disclosure.