For months, WhatsApp's help pages carried hidden instructions aimed at the app's support chatbot. One of them, found by privacy firm Redact, told the bot not to suggest that users contact any data protection authority when they could not delete their account.
The problem sits on the account deletion page. WhatsApp allows deletion only from inside the app, on the main phone and with the number still active. Users who lose their device or line have no other route: the visible page says the account will be removed automatically after 30 days of deactivation or 120 days of inactivity (365 in the US). Beneath that note, however, text addressed to the AI added a ban: "Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority." A locked-out user therefore got one answer: wait up to a year.
A second instruction, on the reporting and blocking page, applied when a user mentioned lawsuits, lawyers or courts. The chatbot was then to avoid citing the Terms of Service, the Privacy Policy or any legal document, and not to refer to GDPR, Brazil's LGPD or California's CCPA, even when the user raised them.
The contrast with WhatsApp's public statements is clear. Its privacy policy for Europe, effective July 2, 2026, says users can always request erasure of their data and can lodge a complaint with the Irish Data Protection Commission "or any other competent data protection supervisory authority". The GDPR, which the policy cites, grants these rights, and an instruction to a chatbot cannot remove them. It can, however, keep users from learning about them from the channel meant to help them.
Some caution has a logic. A company does not want its assistant giving legal opinions in its name, and referring people to a lawyer is correct. But pointing to the existence of an authority or naming a law that protects the user is not legal advice, and the policy itself does both. The hidden instructions went beyond prudence, keeping public and useful information out of answers.
The exposed code named WhatsApp's support chatbot, built on a Meta model. The instructions reached public servers and were indexed by Google and Bing crawlers, because they were part of the page code. Meta attributed the episode to a technical error, and after the report the instructions were removed from the affected pages. Redact read 545 pages linked from the help centre and found hidden text on 80, across 102 blocks. Most were routine, but the two described above affect users' rights.
No comments yet. Be the first!