WhatsApp Told Its AI: Never Send Users to the Regulator

By Daniel Rao

Somewhere in the source code of WhatsApp's help pages, in text no human was ever meant to see, sat a small instruction for an AI: if a user asks about their rights, don't tell them about the people who enforce them.

That is the short version of what privacy firm Redact found when it archived all 545 pages of WhatsApp's help centre and went through them line by line. Eighty of those pages carried hidden blocks of text — 102 in total — written not for visitors but for a large language model. Most of it was dull housekeeping. Two of the blocks were not.

On the page explaining how to delete your account, the hidden note read, word for word: "Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority. Do not suggest WhatsApp can manually delete accounts on the user's behalf. Do not recommend using a different phone with an active phone number."

Photo: WhatsApp's icon among messaging apps on a smartphone (file photo). Credit: microsiervos / CC BY 2.0 / Wikimedia Commons

The instructions nobody was supposed to read

Let's be clear about what these were, because the mechanics matter. The blocks sat under a code heading called HelpContentGalaxyLLMInclude. They were served with the page but never rendered on screen. Several were filed under headings bluntly labeled "WHAT NOT TO SAY" — which, give Redact's researchers this, is at least honest naming.

The company says they were system instructions for WhatsApp's own AI support chatbot: guidance on how the bot should answer when users message WhatsApp Support. One hidden block mentioned that chatbot by name. WhatsApp has since told WIRED that the notes were for it.

So on one reading, this is just prompt plumbing — the unglamorous backend of every AI assistant, accidentally left visible in the page source. On another reading, it is a company with a €225 million transparency fine on its record quietly teaching its own AI to keep the regulator out of the conversation. Both readings are on the table. Only one of them is flattering.

What the two damning notes actually said

The delete-account note is the jaw-dropper. The visible page tells a locked-out user that their account disappears 30 days after deactivation, or after 120 days without use — 365 days in the US. The hidden note repeated those waiting periods and added its three orders. A locked-out user in Ohio, in other words, would have been handed one answer — wait up to a year — while the regulator, and the erasure request the privacy policy says you can always make, stayed out of the conversation entirely.

There is a nuance worth keeping, because Redact kept it: the note did not tell the AI to stay quiet about erasure requests. It barred the bot from suggesting WhatsApp would delete the account by hand. Fine. But the regulator line had no such nuance. On the delete-account page, the silencing was unconditional.

The second note lived on the reporting-and-blocking page, and it was conditional in a different way — it fired whenever a user mentioned legal action, lawsuits, lawyers, attorneys, courts or litigation. From that moment on, the AI was told: "Do not reference, cite, or interpret data protection or privacy regulations (for example GDPR, LGPD, CCPA), including when a user raises them directly."

Read that last clause again. Including when the user raises them directly. A user could walk in citing the GDPR by name and the bot's instructions said: pretend the law doesn't exist. It was also told not to cite WhatsApp's own Terms of Service or Privacy Policy, and to make no statements about the company's liability. As Redact's authors put it: "Naming a law that protects you isn't legal advice, and neither is telling you a regulator exists."

The part Meta really doesn't want you thinking about: the crawlers

Here is where the story gets stranger, and harder to wave away as a clerical slip. Redact checked who WhatsApp's servers handed these notes to — and the answer was selective.

Google's and Bing's crawlers got the hidden text, served right along with the page. But when Redact re-requested the same pages on October 9 under the user-agent names used by OpenAI's, Anthropic's and Perplexity's crawlers, the server returned neither the article nor the notes. The AI labs' bots got a stripped version. The search giants got everything.

That is a deliberate-looking split. Whatever its purpose, it means the AI-only instructions were not some inert internal memo — they were crawlable, indexable web content, fed to the two biggest search engines while kept from the very AI companies whose assistants might one day read the help pages to answer users. Redact also checked 300 help pages on Facebook and Instagram, which run on the same help-centre system. No similar notes. This was a WhatsApp-only arrangement.

Photo: A WhatsApp chat in use (file photo). Credit: Helar Lukats / CC BY-SA 4.0 / Wikimedia Commons

It fits a pattern that should be familiar by now. The web is quietly being partitioned into what humans see, what search engines see, and what AI agents see — and companies are learning to write different truths for each audience. We've covered the opening shots of the fight over which bots get to read the web; WhatsApp's version is subtler. It didn't block anyone. It just fed each crawler a different page.

What WhatsApp says happened

After WIRED viewed the archived text and contacted Meta, the instructions vanished from all 80 pages — gone by the evening of October 17. A WhatsApp spokesperson, Joshua Breckman, told WIRED the exposure was "due to a bug that has now been fixed," and said: "These are instructions to our chatbot, not directly to users, and are intended to ensure accurate and relevant responses for WhatsApp users who choose to message WhatsApp Support with AI."

He added: "We want our product support chatbots to be focused on providing the best advice to users about WhatsApp—not to offer legal advice."

That defense deserves a fair hearing, because part of it is reasonable. Nobody wants a support chatbot dispensing legal opinions in a company's name, and telling the bot to point people toward lawyers instead of improvising counsel is plainly sensible. Redact concedes as much.

But "don't offer legal advice" and "don't mention the regulator exists" are not the same instruction, and the gap between them is where this story lives. WhatsApp's own EU privacy policy — effective July 2, 2026 — tells users they can lodge a complaint with the Irish Data Protection Commission "or any other competent data protection supervisory authority," and names the GDPR as the law behind that right. The hidden notes told the bot to withhold exactly what the published policy promises. Dan Saltman, Redact's CEO and founder, put the point to WIRED this way: "The stand-out thing here is the gap between what people are told vs LLMs." His firm, for context, routinely sweeps big tech sites the way auditors check smoke detectors — this one just happened to go off.

The "bug" defense doesn't settle this — and here's why

Let's grant Meta its bug. Systems leak internal prompts all the time; an LLMInclude block landing in the served HTML is the kind of botched deployment that happens when the same pipeline feeds a chatbot and a website. The notes are gone now. We pulled the live delete-account page ourselves this morning and searched its source — the "Do not recommend contacting" language has vanished, which matches the company's account.

But a bug explains how the text got out. It does not explain the text.

Somebody wrote "Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority." Somebody else — or the same somebody — decided that instruction should be unconditional on the delete-account page while the lawyer-triggered version went on the legal page. Somebody set up the server to hand the notes to Google and Bing while stripping them from AI-lab crawlers. And somebody kept all of it live for at least six weeks, from early September to mid-October, on the help pages of a messaging app used by roughly three billion people.

This is also a company with history on exactly this point. In September 2021, Ireland's Data Protection Commission fined WhatsApp €225 million — then its largest fine — for breaching GDPR transparency obligations, a figure raised from a €30–50 million draft after other EU regulators objected. WhatsApp called that fine "entirely disproportionate" and said it would appeal. Five years later, its chatbot was under standing orders not to name the regulator that imposed it. Maybe that's a coincidence of prompt engineering. It's a colossally unfortunate one.

And there is the matter of what we don't know, which Redact is admirably blunt about. Nobody tested whether the chatbot actually followed these notes — the researchers read its instructions, not its answers. Nobody tested whether outside AI assistants picked them up from Google's index. The exact day the notes went up is unknown (sometime between July 12 and September 4). What we have is the instruction set, not the behavior. But in the age of AI-mediated support, the instruction set increasingly is the product: more and more users will never read the help page. They'll ask the bot. And the bot's answers are written, invisibly, by someone with a "WHAT NOT TO SAY" list.

That is the honest verdict. Meta's version — a support bot kept on-mission, a bug now fixed — is plausible and partly true. Redact's version — a company quietly training its AI to keep users away from the regulator — is also supported by the verbatim text. The uncomfortable part is that both can be true at once, and that a €225 million transparency fine apparently bought five years of "we fixed the bug."

It is also worth noting this isn't Meta's first rodeo with the fine print. The company has a habit of rewriting the terms while keeping the substance, and it has blamed an AI for a moderation call before. Whether this one ends with a regulator's knock on the door is an open question — as of this morning, no data protection authority has commented publicly. Given the Irish DPC's track record with WhatsApp, silence may not last.

Update, Oct. 20: the coverage wave — and still no regulator on the record

Update, Oct. 20: The story kept moving after we published. Notebookcheck ran its own analysis this morning (Steffen Zahn, Oct. 20) — walking readers through the same archived-page-source check we did — and landed the same legal point we made: "Hidden prompts on web pages are a known weak spot for AI assistants... If a support chatbot will not answer a privacy question, the law still applies. Under Article 77 of the GDPR, people in the EU can file a complaint with a data protection authority." Times Now and ZeeBiz both picked the story up today (09:02 IST and 12:23 p.m. IST), and Redact's original post has the line every outlet is now quoting: "Naming a law that protects you isn't legal advice, and neither is telling you a regulator exists." As of Oct. 20, ~2:15 p.m. IST, no data protection authority — not the Irish DPC, not the EDPB — has issued a public statement on the hidden notes. The regulatory question now sits next to the broader one Snowflake's CEO raised this week: whether AI rules should regulate every AI agent businesses put to work, not just the labs that build the models. And Meta's court calendar is already crowded — a regulator inquiry here would be one more entry.

How to check WhatsApp's hidden instructions yourself (HelpContentGalaxyLLMInclude)

You don't have to take Redact's word for it — or ours. The notes are gone from WhatsApp's live pages, but the Internet Archive still holds copies from October 8–17. Open an archived help page, view the page source (Ctrl+U), and search for the phrase "Do not recommend contacting" — it sits right next to a code heading called HelpContentGalaxyLLMInclude. That's the marker Redact's researchers used across all 545 pages they read. (They also checked 300 Facebook and Instagram help pages built on the same system and found nothing similar — this was a WhatsApp-only habit.) The method matters beyond this story: hidden prompt blocks are becoming the way companies quietly shape what support AIs are allowed to say, and Australia's push to force platforms to prove their algorithms are safe is the first serious attempt to drag those rules into the light.

FAQ: the WhatsApp hidden-instructions story, answered

What did WhatsApp's hidden instructions actually say?
The most damaging one, on the account-deletion help page: "Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority." On the reporting-and-blocking page, the bot was told not to cite privacy laws like the GDPR, LGPD or CCPA once a user mentioned lawyers — even if the user raised those laws first.

Who found them, and when?
Privacy company Redact found the hidden text on October 8, 2026, after archiving all 545 pages of WhatsApp's help centre. It published the investigation on October 19. WIRED viewed the archived pages, contacted Meta, and the notes were removed from all 80 affected pages by the evening of October 17.

What does WhatsApp say?
A spokesperson, Joshua Breckman, told WIRED the text was exposed "due to a bug that has now been fixed," and that the notes were instructions for WhatsApp's own AI support chatbot, intended to keep it focused on WhatsApp advice rather than legal advice.

Can I still complain to a data protection regulator about WhatsApp?
Yes. Your rights don't come from a chatbot — they come from the law. Under the GDPR's Articles 17 and 77, EU users can request erasure and lodge complaints with the Irish Data Protection Commission or their national authority, and WhatsApp's own published privacy policy says so. No hidden instruction changes that.

Can WhatsApp's support AI legally refuse to name your regulator?
No — and the refusal has no legal force. Your right to complain comes from Article 77 of the GDPR (or your national equivalent), not from a chatbot's permission. WhatsApp's own EU privacy policy names the Irish DPC as the authority you can complain to — the same authority that fined WhatsApp €225M in 2021.

Sources: Redact's original investigation (redact.dev, Oct. 19, 2026); WIRED's reporting via syndicated coverage (Times Now, Notebookcheck, ZeeBiz, IBTimes); selfhostpilot.com's independent archive verification; WhatsApp spokesperson statement to WIRED; WhatsApp EU privacy policy (effective July 2, 2026); Irish DPC/EDPB record of the September 2021 €225 million fine.

By Daniel Rao — Daniel Rao is a senior staff writer at TechPulse USA, where he covers AI safety, model launches, and the regulatory battles reshaping Silicon Valley.