Ada Lerner

Zhiping Zhang, Michelle Jia, Hao-Ping Lee et al. · cmu

The widespread use of Large Language Model (LLM)-based conversational agents (CAs), especially in high-stakes domains, raises many privacy concerns. Building ethical LLM-based CAs that respect user privacy requires an in-depth understanding of the privacy risks that concern users the most. However, existing research, primarily model-centered, does not provide insight into users' perspectives. To bridge this gap, we analyzed sensitive disclosures in real-world ChatGPT conversations and conducted semi-structured interviews with 19 LLM-based CA users. We found that users are constantly faced with trade-offs between privacy, utility, and convenience when using LLM-based CAs. However, users' erroneous mental models and the dark patterns in system design limited their awareness and comprehension of the privacy risks. Additionally, the human-like interactions encouraged more sensitive disclosures, which complicated users' ability to navigate the trade-offs. We discuss practical design guidelines and the needs for paradigm shifts to protect the privacy of LLM-based CA users.

Jenny Tang, Eleanor Birrell, Ada Lerner

Privacy and security researchers often rely on data collected through online crowdsourcing platforms such as Amazon Mechanical Turk (MTurk) and Prolific. Prior work -- which used data collected in the United States between 2013 and 2017 -- found that MTurk responses regarding security and privacy were generally representative for people under 50 or with some college education. However, the landscape of online crowdsourcing has changed significantly over the last five years, with the rise of Prolific as a major platform and the increasing presence of bots. This work attempts to replicate the prior results about the external validity of online privacy and security surveys. We conduct an online survey on MTurk (n=800), a gender-balanced survey on Prolific (n=800), and a representative survey on Prolific (n=800) and compare the responses to a probabilistic survey conducted by the Pew Research Center (n=4272). We find that MTurk response quality has degraded over the last five years, and our results do not replicate the earlier finding about the generalizability of MTurk responses. By contrast, we find that data collected through Prolific is generally representative for questions about user perceptions and experiences, but not for questions about security and privacy knowledge. We also evaluate the impact of Prolific settings, attention check questions, and statistical methods on the external validity of online surveys, and we develop recommendations about best practices for conducting online privacy and security surveys.