WhatsApp Hidden AI Instructions: Check Your Draft in 2026

WhatsApp's hidden AI instructions are a good reason to check the source of any AI-written support draft. Keep your own record of the problem separate from the company's help text. The archived deletion page from October 9 includes chatbot directions, including an instruction not to recommend a regulator. That historical copy confirms the instruction existed, but it doesn't prove the chatbot followed it.^[7]

On October 20, the investigation led to a long discussion on Hacker News. I read Redact's original report and WIRED's coverage, then checked the archived page source. If you're writing a support request, focus on a narrower question than the headline debate: whose instructions shaped the message you plan to send?^[1][2][3]

The archived instruction and the missing outcome

Redact reports 102 hidden blocks on 80 of the 545 help pages it reviewed. Its October 18 check found those blocks gone. The counts describe the investigator's findings, without establishing how many conversations were affected.^[1]

The archived deletion page gives us something concrete to check. Search its source for HelpContentGalaxyLLMInclude or the phrase about recommending a Data Protection Authority. That text appears in the supplied page data, but someone reading the regular help page would not see it as guidance. Use the dated archive for this check because the current page is a different artifact.^[7]

WIRED reviewed the exposed material independently and received a response from WhatsApp. The company said the directions belonged to its support chatbot and became public because of a bug. It said the bug had been fixed, and described keeping support focused on WhatsApp help rather than legal advice.^[2]

That explanation belongs next to the finding. Redact explicitly says it did not test whether WhatsApp's chatbot or outside assistants followed the notes. The evidence does not show that everyone's AI research was taken over. The historical copy should not be presented as the current page.^[1]

For a support message, I'd start by checking the source. The archived instruction is enough reason to check which parts of a source describe the product and which parts tell the model how to respond.

Write down the problem before importing an answer

Imagine a fictional project-management account whose exported task list is missing comments. Support asks you to export again. The second file has the same gap, and you want the comments recovered or an explanation of what the export includes.

Start with that account in a blank note. Record the export you ran, the date, and one task with missing comments. Add the earlier support response if you have it. Stick to what you know. You can't tell yet whether the export failed, the feature leaves out comments, or your permissions hid them.

A help page may explain a real limitation. If it does, record the limitation and its source in a separate note. If the page doesn't answer your case, leave the question open instead of filling the gap with a polished summary.

A sample dictation brief for this fictional request could read: "I exported the project twice. Both files include the task titles, but neither includes the comments for the sample task. Please confirm whether comments should be included and let me know what information you need to investigate."

You can compare that brief with the files. It doesn't accuse the vendor of losing data, invent a deadline, or accept an answer you haven't received. If a later AI rewrite adds, "I understand this is expected behavior," you'll know that concession came from the rewrite, not your notes.

This is how the WhatsApp story connects to everyday writing. A vendor's source can help you understand a problem, but it may also contain instructions for the vendor's assistant. Your draft still needs to state your question. A polite tone is fine, but review any sentence that decides the issue in the vendor's favor.

Give the assistant a small, labeled source packet

For a sensitive support exchange, I wouldn't paste a full extracted page alongside my instructions. I'd make a short packet with the relevant passage, its URL, and the date I checked it. I'd put my observations first and clearly label the passage as source material.

You can then ask the assistant to draft from your observations and use the passage only as evidence. Ask it to flag any questions the passage leaves unanswered. This can help with writing, but it doesn't guarantee security. A label can't reveal how every model or retrieval tool handles content from the page.

Read the proposed message with the source open beside it. First, check every claim about what you did. Did the draft say you tried another device when you only repeated the export? Did it turn a request for clarification into an agreement that nothing could be done? Remove those changes before polishing the wording.

Next, trace each policy claim to the passage it cites. An answer can link to a real page and still draw a conclusion the page doesn't support. The link gives you somewhere to start, but you still need to read the source.

The dates matter. The preserved October 9 page and the later removal of its hidden notes are separate parts of the record. A draft about the discovery should treat the archived instruction as historical. A draft asking for help with an account today should use current information about that account and service.^[1][7]

If the assistant can't explain where a sentence came from, leave it out or turn it into a question. You can still send a short, factual request without explaining everything the vendor might mean.

Dictate your account, then inspect the text

I build DictaFlow, so I have a commercial interest in voice input. Its role here is simple: hold a shortcut, speak, and insert the text into the active writing field. It can help you capture your own explanation before borrowing wording from a help page.^[6]

It doesn't tell you whether the vendor's answer is correct. Dictating a paragraph still requires a review. Before sending it, check the names, dates, negation, and any cleanup. In the fictional export request, "the comments aren't included" and "the comments are included" would send support in opposite directions.

Keep the first note nearby as you edit the final message. If you use AI to shorten it, compare the result with that note and the attached evidence. When a support reply arrives, save it with the request so you know which question it answered and which question you still need to ask.

Sources

[1] Redact's original investigation and dated updates

[2] WIRED's independent reporting and WhatsApp response

[3] Hacker News discussion, checked October 20, 2026

[6] DictaFlow official product workflow

[7] WhatsApp deletion help, archived October 9, 2026 at 05:48:48 UTC