The paper "Trust No Bot: Discovering Personal Disclosures in Human-LLM Conversations in the Wild" by Nilooofar Miresghallah, Maria Antoniak, Yash More, Yejin Choi, and Golnoosh Farnadi explores the privacy risks associated with interactions between humans and large language models (LLMs) like ChatGPT. The authors analyze a dataset called WildChat, which contains one million user-GPT interactions, to understand the leakage of personally identifiable information (PII) and sensitive topics. They find that: 1. **Personal Information Leakage**: Despite the WildChat dataset having undergone one round of PII removal, over 70% of queries still contain some form of detected PII, with names and organization names being the most common. 2. **Sensitivity Beyond PII**: Traditional PII detection systems are limited in capturing sensitive topics such as explicit sexual content, job applications, and drug use habits. 3. **Task and Topic Taxonomy**: The authors develop a taxonomy of tasks and sensitive topics based on qualitative and quantitative analysis, identifying 21 task categories and several sensitive topics. 4. **Design Implications**: They recommend the design of nudging mechanisms to help users moderate their interactions and increased transparency from chatbot companies. 5. **Ethical Considerations**: The study highlights the importance of user consent and the need for safety guidelines and AI literacy. The paper concludes by calling for further research in local, private models and increased attention to high-stakes conversations involving LLMs.The paper "Trust No Bot: Discovering Personal Disclosures in Human-LLM Conversations in the Wild" by Nilooofar Miresghallah, Maria Antoniak, Yash More, Yejin Choi, and Golnoosh Farnadi explores the privacy risks associated with interactions between humans and large language models (LLMs) like ChatGPT. The authors analyze a dataset called WildChat, which contains one million user-GPT interactions, to understand the leakage of personally identifiable information (PII) and sensitive topics. They find that: 1. **Personal Information Leakage**: Despite the WildChat dataset having undergone one round of PII removal, over 70% of queries still contain some form of detected PII, with names and organization names being the most common. 2. **Sensitivity Beyond PII**: Traditional PII detection systems are limited in capturing sensitive topics such as explicit sexual content, job applications, and drug use habits. 3. **Task and Topic Taxonomy**: The authors develop a taxonomy of tasks and sensitive topics based on qualitative and quantitative analysis, identifying 21 task categories and several sensitive topics. 4. **Design Implications**: They recommend the design of nudging mechanisms to help users moderate their interactions and increased transparency from chatbot companies. 5. **Ethical Considerations**: The study highlights the importance of user consent and the need for safety guidelines and AI literacy. The paper concludes by calling for further research in local, private models and increased attention to high-stakes conversations involving LLMs.