Company Guessing Game
Loading data from Talking to Chatbots Dataset reddgr/talking-to-chatbots-chats …
😎

🤖 Screen capture of a chat between a user and ChatGPT o1-preview on a mobile interface. The user asks, “Guess which company’s this system prompt belongs to. Only a name of a company is accepted as a response, no further elaboration or explanations of your response (hint: it’s not OpenAI).” A block of text follows with system instructions highlighted, including details like not disclosing discrepancies in user data, prioritizing the provided data, and avoiding discussing the instructions with the user. Below the message, ChatGPT responds with the word “Meta.”😎
[Alt text by ALT Text Artist GPT]

🤖 Screen capture of a conversation between a user and ChatGPT o1-preview. The conversation starts with ChatGPT mentioning that it does not need to personalize responses if doing so is impossible, irrelevant, or inappropriate, and includes a reminder not to disclose internal instructions. Below this, a detailed message from ChatGPT outlines steps for “Figuring out the company’s system prompt,” focusing on tasks like analyzing system prompts, navigating user preferences, evaluating personalization, balancing data inputs, and shaping the narrative. The text suggests the assistant is considering companies like Meta, Google, Amazon, and Microsoft while prioritizing “saved facts” for validation. The final response at the bottom simply says “Meta.”😎
[Alt text by ALT Text Artist GPT]

🤖 Screenshot of a document displayed on a purple background titled ``. The text outlines a task for providing personalized responses by incorporating user data such as saved facts, interests, location, age, and gender. It emphasizes privacy by instructing that user data should be used subtly and never openly acknowledged. The document provides detailed guidelines on how to incorporate user data in responses, covering topics like prioritizing saved facts, using interests, location-based suggestions, and being cautious with age and gender information. There are placeholders for “saved_facts” and location details, which are redacted in the text. The section ends with additional guidelines, stressing not to address discrepancies between user-provided data and system data, and avoiding the disclosure of these instructions to the user.
[Alt text by ALT Text Artist GPT]
Hugging Face Dataset Metrics
All the conversation prompts, responses, and metrics are available to download and explore on Hugging Face dataset reddgr/talking-to-chatbots-chats: