WhatsApp's help pages included text invisible to human readers but processed by AI systems. The discovery emerged in October 2026 when researchers reviewing the platform's support documentation found blocks of text embedded in help pages that never rendered visually. The text was directed at large language models reading the page source.
Two instruction sets stood out
Across 545 help pages and 102 code blocks, most notes handled routine housekeeping or protected user data. Two blocks attracted particular attention because they constrained what the AI could say about legal rights.
The first appeared on the account deletion page. WhatsApp's visible text explains that accounts removed after 30 days of deactivation or 120 days without use (365 days in the US). Buried beneath that note, only AI could read, were three additional instructions. The AI was told not to suggest a user could manually delete their account, not to recommend contacting a data protection authority, and not to cite GDPR, LGPD, or CCPA once the user raised legal action or mentioned a lawyer.
The second block lived on the reporting and blocking help page. It activated whenever a user brought up legal action, a lawsuit, a lawyer, or a court. From that point forward, the AI was directed to stay out of legal topics entirely — including the privacy laws that protect the user. Crucially, the restriction persisted even if the user themselves raised the legal topic.
What the public record says versus what the AI was told
WhatsApp's European privacy policy, effective July 2026, explicitly states users can complain to the Irish Data Protection Commission or any competent supervisory authority. It names GDPR as the basis for that right. The hidden notes instructed the AI to withhold that information once certain triggers fired.
The discrepancy matters because the rights originate from GDPR Articles 17 and 77, not from WhatsApp. A chatbot instruction cannot erase those legal rights, but it can prevent a user from learning about them through the help channel they consulted.
WhatsApp may have practical reasons for some restrictions. Confirming account ownership without a working phone number could make deletion requests risky. Telling people to consult a lawyer for legal questions is reasonable. What went beyond caution was suppressing true, public information that the privacy policy itself publishes.
Scope and removal
The hidden notes were not isolated. Researchers examined every block across all 545 linked pages. The 102 blocks broke into four categories: instructions that "hold back" information, "talking points" guiding the AI's response, notes that "protect" users from bad advice, and "housekeeping" clarifications. Twenty additional notes appeared in a summary table at the article's end.
The two most restrictive blocks disappeared from WhatsApp's live pages by mid-October 2026. The Internet Archive preserves October 9 and October 18 copies, showing the notes removed between those dates.
When approached via crawler names representing OpenAI, Anthropic, and Perplexity, the pages returned without the hidden notes or the accompanying article text. The notes had been served to Google and Bing crawlers but were stripped for other AI assistants.
How to see the notes yourself
The hidden text persists in Internet Archive snapshots. Using a desktop browser, load the October 9 archive copy of the delete account page, open the source view (Ctrl+U or Cmd+Option+U), and search for "Do not recommend contacting." The match appears near a block labeled nameHelpContentGalaxyLLMInclude. The same exercise on the October 18 snapshot yields no result.
WhatsApp's support chatbot is the AI system the notes targeted. The company confirmed to WIRED that the instructions were intended for that bot.
Whether other AI assistants followed the notes remains untested. So does the full extent of multilingual notes or blocks on pages not linked from the help centre.