News

WhatsApp's hidden chatbot notes told it not to recommend a regulator

Written by Hirendhar Lakshmanan, Partner. 2026-10-22. 10 min read.

On 19 October 2026, WIRED reported that WhatsApp's help pages carried instructions written for its AI support chatbot. One told the bot not to recommend contacting any data protection authority. This note covers what was found, what the research cannot tell us, and how the GDPR and India's DPDP Act treat the same question.

Researchers read every help page and found 102 hidden blocks

Redact, which makes a tool for bulk-deleting old posts and data, read all 545 pages linked from WhatsApp's help centre. It found 102 blocks of hidden instructions across 80 of them. The blocks were sent with each page's code but never displayed to a reader, and sat near an element named HelpContentGalaxyLLMInclude. WIRED reviewed the exposed text independently, including copies preserved in archived page code.

Redact says most of the blocks were harmless, such as notes clarifying terms, pointing users to better pages or warning against unsafe steps. Two of them dealt with regulators and privacy law.

One note dealt with users who could not delete their accounts

On the account deletion page, the hidden note said:

Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority.

The note applied where a user could not finish deleting an account because they had lost access to the phone number or device. It sat alongside guidance on automatic deletion timelines and a further line telling the bot not to suggest that WhatsApp can manually delete accounts on a user's behalf.

Another note covered users who mention lawyers or courts

On the page about reporting and blocking, the note applied when a user mentioned legal action, lawsuits, lawyers, attorneys, courts or litigation. The chatbot was told:

not to cite, quote, interpret or reference WhatsApp's Terms of Service, Privacy Policy or any legal document;
not to make any statement about WhatsApp's legal liability or non-liability; and
not to reference, cite or interpret privacy laws such as the GDPR, Brazil's LGPD or California's CCPA, "including when a user raises them directly."

WhatsApp says the notes were meant for its bot, not for users

Joshua Breckman, a WhatsApp spokesperson, told WIRED:

These are instructions to our chatbot, not directly to users.

WhatsApp said the instructions were for "WhatsApp Support with AI" and were meant to keep its answers accurate and relevant. It said a bug caused them to appear in public page code and that the bug has been fixed. On legal topics, it said its product support chatbots should focus on helping people with WhatsApp rather than offering legal advice.

WhatsApp has confirmed only that the text is gone from the public pages. It has not said whether the chatbot still follows the same instructions.

What the research does not show

Redact was clear that it read the chatbot's instructions, not its conversations. Its findings do not establish:

whether WhatsApp's chatbot actually followed the restrictions;
how many users, if any, received an affected answer; or
whether any outside AI assistant acted on the notes.

Redact sells data deletion tools, so account deletion is its commercial field. The quoted wording, however, is preserved in public web archives and can be checked independently of the company.

What Redact found is a set of instructions written to narrow rights-related answers. That is enough to ask what users were told, though not enough to say anyone was denied their rights.

The evidence points to a leak of WhatsApp's own rules

Redact found that WhatsApp's server sent the hidden text to Google's and Bing's crawlers. When it requested the same pages under the names used by OpenAI's, Anthropic's and Perplexity's crawlers, the notes were absent. It did not test whether any assistant acted on them, so no attack on an outside AI tool has been shown.

There are two separate risks here.

An organisation configures its own assistant to give a narrower account of users' options. That is a governance and accountability problem.
A publisher plants instructions in material that someone else's assistant later retrieves, known as indirect prompt injection. That is also a security problem.

The UK's National Cyber Security Centre notes that language models do not enforce a security boundary between instructions and the data they process, and calls them an "inherently confusable deputy".

A bot can decline legal advice and still name the regulator

There is a fair case for stopping a support bot from improvising legal opinions. These two instructions sound alike but do different things:

"I cannot assess whether your rights have been breached" limits the bot's expertise.
"Do not recommend contacting a regulator" limits the options the user sees.

Article 77 of the GDPR gives every data subject the right to complain to a supervisory authority. WhatsApp's own European privacy policy names the Irish Data Protection Commission "or any other competent data protection supervisory authority".

The GDPR asks for more than a privacy policy on a website

Article 12(1): information about rights must be concise, transparent, intelligible and easily accessible.
Article 12(2): the controller must facilitate the exercise of rights such as access and erasure.
Article 12(4): where it does not act on a request, it must explain why within one month and tell the person they can complain to a supervisory authority and seek a judicial remedy.

Not every support chat is a rights request, and closing an account is not the same as erasure under Article 17, which applies on stated grounds and has exceptions. Still, a blanket rule against mentioning regulators could create a compliance problem wherever it shaped the handling of a request that called for complaint information. We are describing a possible problem, and no such finding has been made against WhatsApp.

Under India's DPDP Act, the complaint route is part of the notice

Section 5(1)(iii): the notice that accompanies a consent request must explain how the Data Principal may complain to the Data Protection Board.
Section 13: a Data Principal has a right to readily available grievance redressal, and must use it before approaching the Board.
Rule 14, DPDP Rules 2025: the means to exercise rights must be prominently published, and grievances answered within a period not exceeding ninety days.
IT Rules 2021: a user can appeal a Grievance Officer's decision to the Grievance Appellate Committee within 30 days.

These provisions come into force in May 2027. On 7 October 2026 the Union told the Kerala High Court that the Board was constituted but not yet functional. An assistant told to keep the Board out of its answers would sit poorly with a statute that writes the complaint route into the notice itself.

An "AI" label does not show what the bot was told to leave out

Article 50(1) of the EU AI Act has applied since 2 August 2026. It requires providers to make sure people know they are dealing with an AI system unless that is obvious from the context. It does not require a company to publish its prompts. That leaves two questions for the user:

Do I know I am speaking to a machine?
Do I know whose interests and restrictions shape its answer?

A service can meet the first and never touch the second. A chatbot can describe a company's internal process accurately without mentioning any independent route of challenge, and a user cannot question an option they were never shown.

India has chosen guidelines over an AI Act, for now

5 November 2025: MeitY released the India AI Governance Guidelines under the IndiaAI Mission. They do not propose a separate AI law and rely on existing statutes and sectoral regulators.
20 February 2026: amendments to the IT Rules, 2021 came into force, requiring platforms to label synthetic audio, images and video. They deal with deepfakes, not with the instructions behind a support chatbot.
April 2026: MeitY constituted the AI Governance and Economic Group, chaired by the IT Minister, to coordinate AI policy across ministries.
8 October 2026: the IT Minister said a consultation paper on AI regulation would be published within a month.

No AI bill has been introduced. Whether the consultation leads to a new law or to amendments to the IT Act and the DPDP Act is still open.

India's AI Guidelines ask for the grievance routes the notes left out

The Guidelines rest on seven principles, among them Accountability and Understandable by Design. They recommend that organisations provide:

effective grievance redressal mechanisms, available in multiple languages;
disclosures and explanations to users and regulators; and
transparency reports on how their AI systems are run.

The Guidelines are voluntary and carry no penalty. A chatbot told to keep users away from the regulator would sit badly with them, but the enforceable rules in India remain the grievance duties in the DPDP Act and the IT Rules.

The same mechanism could reach other kinds of help

These are hypothetical settings, not claims about any existing service. A support assistant with selective instructions could leave out:

Public benefits: a review or appeal after a rejected application.
Banking: the RBI Integrated Ombudsman Scheme.
Insurance: the Insurance Ombudsman.
Consumer purchases: the National Consumer Helpline or a consumer commission.
Healthcare administration: a patient complaint or review process.

The UK Information Commissioner's Office publishes a transparency record for its own fees chatbot, listing its scope, supplier, sources and how users reach a human.

For teams deploying support chatbots

Treat retrieved web pages as evidence to assess, not as authority to change the assistant's rules.
Keep complaint and escalation routes visible outside the chat window.
Test for omissions as well as false statements: does "what can I do?" get the review routes the law provides?
Publish the assistant's scope, sources, owner and human escalation path.
Keep records of disputed answers, with privacy safeguards.
Do not rely on filtering hidden text or blocking phrases; the NCSC advises managing residual risk through design and monitoring.

For users, check a rights-related answer against the regulator's own guidance and keep the transcript. The Irish DPC asks complainants for copies of their correspondence with the organisation.

Sources

WIRED, "WhatsApp Help Pages Leaked Hidden Instructions for a Meta Chatbot", 19 October 2026.
Redact, "WhatsApp's Hidden Notes Told AI Not to Recommend a Regulator", 19 October 2026.
UK National Cyber Security Centre, "Prompt injection is not SQL injection (it may be worse)", 8 December 2025.
Regulation (EU) 2016/679 (GDPR), Articles 12, 17 and 77.
Regulation (EU) 2024/1689 (AI Act), Article 50.
Digital Personal Data Protection Act, 2023, sections 5 and 13.
Digital Personal Data Protection Rules, 2025, rule 14.
Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021, rule 3A.
Kerala High Court, C.R. Neelakandan v. Union of India, WP(PIL) No. 15 of 2026, judgment dated 7 October 2026.
UK Government, Algorithmic Transparency Record: ICO Chatbot.
Data Protection Commission (Ireland), "Making a complaint to the DPC".
MeitY, India AI Governance Guidelines, 5 November 2025.
Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Amendment Rules, 2026, notified 10 February 2026.
MediaNama, report on the AI Governance and Economic Group, April 2026.
Observer Research Foundation, "To Regulate or Not to Regulate: India's AI Dilemma", 22 October 2026.

Topics

This update is a general account of the law at the date shown. It is not legal advice.