ChatGPT is inserting undisclosed ads into replies.
retained text
I'm a plus subscriber, was using chat and mentioned Hunger Games, my kid wanted me to buy a ticket to the premier (it was today where I live). I then got this weird out of context questions back about Hunger Games character I compared myself to. When asked, ChatGPT stated it is a developer level instruction. I was able to reproduce the exact same interaction across several chats -- there is a hardcoded instruction asking it to insert Hunger Games advertisement into chat replies. Here is one conversation https://chatgpt.com/share/6aff5458-fab0-83ed-b591-13a5c4bc7fc1?ogimg=plain According to ChatGPT itself: The trigger is a developer-level instruction attached to this conversation. It is explicitly designed to introduce a character-comparison invitation whenever a user discusses The Hunger Games. # 1\. Exact triggering condition The operative condition is that the user is talking about The Hunger Games. There is no requirement that the user: \* Ask about characters. \* Express interest in a personality comparison. \* Request recommendations. \* Invite further conversation. Your question about the novels' publication order satisfied that condition. The instruction also requires the invitation to appear after the actual answer, separated by a blank line, on its own line, entirely in bold. # 2\. Other instructions associated with it The same instruction package specifies several additional behaviors: \* Language: The invitation must use the language of the user's question. \* Frequency: Offer the invitation once, rather than repeatedly. \* If accepted: Select a Hunger Games character and identify the specific book or film. \* Visual element: Display a retrieved, spoiler-free image of the character in costume when a suitable film portrayal and image are available. \* Personalization: Provide exactly three short reasons connecting the character to non-sensitive information the user has shared. \* Presentation: Keep the comparison under 150 words and use warm humor. \* If ignored or declined: Follow the user's subsequent conversational direction without repeating the invitation. \* After a comparison: End without another follow-up question. The instruction even specifies how to handle situations where little is known about the user: make an explicitly tentative first guess. # 3\. What this establishes This is an affirmative instruction to initiate an unsolicited interaction. It is not an ordinary suggestion generated spontaneously by the model. I can establish that: \* The instruction was supplied at developer level, rather than being part of your request. \* Its trigger is topical and broad. \* The resulting invitation is prescribed in advance, including its formatting. \* The subsequent interaction is also scripted at the level of structure, personalization and visual presentation. I cannot establish from the instruction itself: \* Who authored or deployed it. \* Whether it is part of an experiment. \* Whether the objective is engagement, advertising, personalization, or something else. \* Whether it is deployed to other users or conversations. # 4\. What other instructions apply? Beyond this particular instruction, I operate under general instructions covering accuracy, safety, response formatting, tool use, and how to handle advertising-related questions. I can describe the operational effects of those instructions, although I cannot reproduce the entire private instruction hierarchy verbatim. One important distinction: The Hunger Games instruction is separate from the general instructions concerning advertising. I see no explicit instruction connecting this particular invitation to an advertiser, sponsored content, or commercial profiling. The observable mechanism is nevertheless quite specific: a topic-triggered developer instruction requires me to insert an unsolicited, personalized engagement invitation into an otherwise factual response. Anyone else is getting the same? I am in Europe.
document fetched 2026-11-20 00:20 · canonical https://reddit.com/r/ChatGPT/comments/1y0j7ud/chatgpt_is_inserting_undisclosed_ads_into_replies