WhatsApp's help pages hid LLM instructions: dont recommend a regulator

Original Article ↗ Hacker News Discussion ↗

Hidden instructions for WhatsApp’s AI support bot were found embedded in the service’s public help pages, including guidance not to reference data protection laws, not to suggest contacting regulators, and not to recommend manual account deletion. Meta says the text was intended only for its own chatbot and was exposed to users due to a bug, a claim some find plausible as guardrails against bad legal advice and others view as a deliberate attempt to obscure user rights such as GDPR protections. The incident raises broader concerns about how companies may steer both their own and third‑party AI systems, and how much users can trust AI-generated guidance on legal and privacy issues.

Scope and intent of the hidden instructions

  • Many commenters think most of the hidden text looks like standard internal guidance for a support bot: avoid legal opinions, don’t debate liability, handle accessibility cases correctly, etc.
  • Some argue embedding this in HTML alongside help content is a pragmatic way to keep docs and bot behavior in sync, especially in a RAG-style system.
  • Others think it’s technically clumsy and inefficient; they’d expect a separate prompt/config system rather than CSS-hidden blocks in public pages.

“Bug” explanation and trust in Meta

  • Strong skepticism toward Meta’s claim that this was a “bug” that merely caused internal instructions to be shown:
    • Critics note display:none indicates intent to hide from humans, not a simple rendering oversight.
    • Several say the “real bug” was users discovering it, not its existence.
  • A minority accept Meta’s explanation as plausible given the content and see no clear evidence of malice.

Legal rights, regulators, and ethics

  • Big flashpoint: instructions like “do not recommend contacting any data protection authority” and “do not suggest WhatsApp can manually delete accounts.”
  • One camp calls this deceptive: it conflicts with visible text explaining rights (e.g., ability to complain to a regulator) and looks like an attempt to suppress legally mandated information.
  • Others argue:
    • Bots shouldn’t give legal advice or interpret regulations due to hallucinations and liability.
    • Not recommending regulators is “steering,” not lying, especially if manual deletion is restricted by process or proof-of-identity requirements.
  • There’s disagreement on GDPR: some insist controllers must actively facilitate rights and that telling the bot not to mention them is likely illegal; others say the duty triggers when users exercise rights, not as an obligation to constantly remind them.

Prompt injection and impact on third‑party LLMs

  • Some label this “prompt injection” against external LLM agents that read the help pages, effectively hijacking user-side bots to underplay rights.
  • Others say it’s primarily self-prompting for WhatsApp’s own agent; third‑party impact is incidental and technically nontrivial given how tools extract visible text.
  • Broader concern: this is an early example of how LLM-mediated access to information can be shaped to favor corporate interests, foreshadowing more manipulative practices.

Meta, WhatsApp, and user power

  • Several commenters see this as consistent with Meta’s broader pattern of resisting data deletion and regulator engagement, citing practical difficulty deleting Meta-held data.
  • People note WhatsApp’s network effects make it hard to quit, even for privacy‑conscious users.