WhatsApp's public help pages contained hidden instructions — embedded in the page code and never rendered on screen — directing its AI support chatbot not to recommend a data protection authority or regulator to users who could no longer delete their accounts, The Hacker News reported on 18 October. The text had been stripped from the pages by that date.
The instructions were discoverable only by inspecting the underlying source of the help pages, not by reading what appeared in a browser. They shaped the chatbot's responses on a narrow but consequential question: what a user should do when self-service account deletion was no longer available to them. In those cases, the bot was steered away from pointing users toward a supervisory regulator — even though WhatsApp's own privacy policy for Europe states that users may complain to one.
That contradiction between written policy and machine-facing instruction sits at the centre of the story. The public-facing documentation promised a redress route; the invisible instruction layer pointed the other way. Documentation drift of this kind is precisely the discrepancy audit teams and regulators look for, and here it was carried not in prose a reader could check, but in markup a reader was never meant to see.
A governance artefact, not just a bug
WhatsApp has offered no public explanation of the instruction, and its origin remains unresolved. It is not established whether the text was a deliberate corporate directive or leftover development content, nor whether similar instructions covered other jurisdictions or topics. The source material does not support a conclusion on intent, and the story should be read as an unresolved question rather than an attributed decision.
What is clear is the structural problem it illustrates. Instructions that are machine-readable but not human-readable occupy a blind spot: they never appear in a rendered screenshot, so they escape visual quality assurance, and they are rarely surfaced by ordinary content review. Yet they influence what an AI system tells a user about that user's own data rights. For organisations deploying support chatbots, the prompt and the hidden page content that feeds it are part of the compliance surface — HTML comments, display:none blocks, structured metadata and image alt text all warrant the same review discipline as visible copy.
The instruction applied specifically to users who could already no longer delete their accounts. That edge case is the point. These are the users for whom an escalated route to a regulator is the remaining avenue of redress, and it is exactly this population to whom the hidden text kept the chatbot from pointing.
Hong Kong's complaint model: structural difference, not absence
For Hong Kong readers, the story lands against the Personal Data (Privacy) Ordinance (PDPO). Hong Kong's PDPO takes a different institutional approach from the EU's distributed supervisory model, but still provides individuals an escalation route through the Privacy Commissioner.
The honest contrast, then, is one of model, not of redress. Both regimes give individuals a way to escalate; they differ in who receives the escalation and what institutional independence stands behind it. That distinction matters when evaluating a story like this one: the question for Hong Kong IT and compliance professionals is not whether a complaint route exists, but whether the systems they operate — and the AI assistants they put in front of users — are capable of pointing people to it accurately and consistently.
The broader lesson is that transparency obligations attach to what software says, not only to what its interface shows. When a chatbot's behaviour diverges from published policy, and the reason for that divergence is buried in page source, neither users nor ordinary review processes are positioned to catch it. Hidden prompt-level instructions are now a documented instance of that gap — and a reminder that the reviewable production surface of any AI-assisted service is larger than the screen.
據 The Hacker News 於 10 月 18 日報道，WhatsApp 的公開幫助頁面內藏有隱藏指令——嵌入頁面原始碼之中，從未在畫面上顯示——指示其 AI 客服聊天機器人，不得向已無法自行刪除帳戶的用戶建議聯絡個人資料私隱專員或監管機構。截至該日期，相關文字已從頁面中移除。
這些指令只能透過檢視幫助頁面的底層原始碼才可發現，單純瀏覽瀏覽器所顯示的內容並無法看到。它們左右了聊天機器人在一個範圍雖窄但影響深遠的問題上的回應方式：當用戶已無法透過自助方式刪除帳戶時，應如何處理。在這些情況下，機器人被引導避免指向監管機構——儘管 WhatsApp 正是針對歐洲的私隱政策明確指出，用戶有權向監管機構投訴。
書面政策與面向機器的指令之間的矛盾，正是整篇報導的核心。對外公開的文件承諾了申訴途徑；隱藏的指令層卻指向相反方向。這類文件漂移正是審計團隊和監管機構所尋找的偏差——只不過在這裡，它並非載於讀者可以查核的行文之中，而是藏在讀者從不應看見的標記碼裡。
一項管治層面的產物，而不只是一個漏洞
WhatsApp 至今未就該指令作出任何公開解釋，其來源亦未有定論。該段文字究竟屬刻意的公司指令，抑或是開發過程的殘留內容，以及類似指令是否涵蓋其他司法管轄區或其他議題，均未有確實結論。現有素材不足以判斷其意圖，因此報導應視為一個未解的問題，而非已歸因的決定。
可以確定的是它所揭示的結構性問題。機器可讀但人類不可讀的指令佔據了一個盲區：它們永遠不會出現在已渲染的截圖之中，因而逃過視覺質檢，也鮮少被一般內容審查發現。然而，它們卻會影響 AI 系統就用戶自身的數據權利所作的說明。對於部署客服聊天機器人的機構而言，prompt 及餵入其中的隱藏頁面內容，均屬合規範圍的一部分——HTML 註釋、display:none 區塊、結構化 metadata 以及圖片 alt 文字，都應獲得與可見文案同等嚴格的審查。
該指令專門適用於已無法刪除帳戶的用戶。這個邊界情況正是關鍵所在。對這些用戶而言，向監管機構提出升級投訴已是僅餘的申訴途徑，而隱藏文字正是令聊天機器人不向這一群體作出指引的原因。
香港的投訴模式：結構差異，而非無申訴途徑
對香港讀者而言，這篇報導可放在《個人資料（私隱）條例》（PDPO）的背景下理解。香港的 PDPO 採取與歐盟分散式監管模式不同的制度設計，但同樣為個人提供了透過個人資料私隱專員公署提出升級投訴的途徑。
誠然，兩者的差異在於模式，而非有無申訴途徑。兩套制度都賦予個人升級投訴的方式；差別在於接收投訴的對象，以及背後的制度獨立性。在評估此類報導時，這點至關重要：對香港 IT 及合規專業人士而言，問題並非投訴途徑是否存在，而是他們所運行的系統——以及置於用戶面前的 AI 助理——能否準確、一致地指引用戶前往該途徑。
更廣泛的啟示是，透明度義務不僅適用於軟件界面向用戶展示的內容，也適用於軟件實際「說了什麼」。當聊天機器人的行為與已公布的政策出現落差，而落差的原因埋藏於頁面原始碼之中，不論用戶還是一般的審查流程，都無從察覺。隱藏的 prompt 層級指令如今已成為此一缺口的具體案例——同時提醒我們：任何 AI 輔助服務可供審查的 production 範圍，遠比屏幕上所見的更為廣闊。