The DOI matches its Crossref record; 43 of the 60 references it cites were found in a registry or an index.
'AI is from the devil.' Behaviors and Concerns Toward Personal Data Sharing with LLM-based Conversational Agents
Open accessAbstract
With the increased performance of large language models (LLMs), conversational agents (CA), such as ChatGPT, are nowadays available to any individual requiring little technical knowledge and skills. Initial studies that have investigated related privacy risks primarily focused on either technical aspects and misuse of these tools, or captured overall perceptions of CA users in small-scale qualitative evaluations. Complementing and extending previous work, we used a quantitative user-centered approach to analyze and compare the behaviors and concerns of users and non-users. We conducted a survey study (N=422) with (1) service users, i.e., users of CA services, (2) local users, i.e., users of a local instance of CA (partially local users, or fully local users), and (3) non-users. We collected self-reported usage patterns and personal data-sharing behavior as well as privacy concerns related to different types of personal data (e.g., health data, demographics, or opinions). Furthermore, we analyze individuals' intention to use CA services in multiple scenarios. Our findings show that users of CA services generally have fewer privacy concerns than non-users. While users rarely share data related to personal identifiers and account credentials, they tend to often share data related to lifestyle, health, standard of living, and opinions. Surprisingly, partially local users tend to share more data with CA services as they also generally use CA services more often and for more diverse purposes. Also, while the majority of CA services users declared not being willing to prioritize CA services as an information source in the described scenarios such as seeking legal advice, between about one-quarter and one-third of partially local users would use CA services for all scenarios. Furthermore, half of the users were willing to stop using CA for privacy reasons (e.g., in case of data leaks), whereas a large majority of non-users reported not using CAs simply because they do not have the need or the opportunity. Our work highlights the high privacy risks for CA services users as CA services largely expand the amount of any type of personal information that can be collected by companies.
Overview
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
In the citing paper
“Similarly, users of conversational agents tend to exhibit fewer privacy concerns than non-users, sharing sensitive lifestyle and health data while withholding direct identifiers [51].”
From Understanding Users' Privacy Perceptions Towards LLM's RAG-based Memory (2025) · cites this paper as [51]
Citations
Cites
Statistical power analyses using G∗Power 3.1: tests for correlation and regression analyses
Faul, F., Erdfelder, E., Buchner, A., & Lang, A.-G. (2009) · Behavior Research Methods · doi:10.3758/brm.41.4.1149 · Cited by 37,615 · cited 1×
“99 probability of type II errors, G*Power [19] estimated significant results ( 𝑝 0 .”
The Coding Manual for Qualitative Researchers
Lungu, M. (2022) · American Journal of Qualitative Research · doi:10.29333/ajqr/12085 · Cited by 10,826 · cited 1×
“Given the well-defined research questions and simple, straight-forward responses in the open-ended questions, the first author coded the answers based on a codebook developed through open coding [44].”
Prolific.ac—A subject pool for online experiments
Palan, S., & Schitter, C. (2018) · Journal of Behavioral and Experimental Finance · doi:10.1016/j.jbef.2017.12.004 · Cited by 3,450 · cited 1×
“We recruited our survey respondents via Prolific , which is con-sideredareliablecrowdsourcing platformforscientificresearch[41].”
Qualitative Content Analysis: Theoretical Background and Procedures
Mayring, P. (2015) · Advances in Mathematics Education · doi:10.1007/978-94-017-9181-6_13 · Cited by 2,263 · cited 1×
“After that, they proceeded to a first inductive coding round by creating main categories to classify and merge related data types [35].”
Extracting Training Data from Large Language Models
Carlini, N., Tramer, F., Wallace, E., Jagielski, M., Herbert-Voss, A., Lee, K., Roberts, A., Brown, T., Song, D., Erlingsson, U., et al. (2020) · arXiv · doi:10.48550/arxiv.2012.07805 · Cited by 727 · cited 6×
“This is not the case for the other categories (see Figures 9,17, 18).”
InfluentialMulti-step Jailbreaking Privacy Attacks on ChatGPTPreprint
Li, H., Guo, D., Fan, W., Xu, M., Huang, J., & Song, Y.-Q. (2023) · Conference on Empirical Methods in Natural Language Processing · doi:10.48550/arxiv.2304.05197 · Cited by 514 · cited 1×
“…have recently investigated privacy risks in the context of LLM-based CAs primarily focused on either technical efforts aimed at measuring [9, 29], mitigating risks during training or related to inferences [15, 24] or captured overall perceptions of CA users in small-scale qualitative…”
Gerber, N., Gerber, P., & Volkamer, M. (2018) · Computers & Security · doi:10.1016/j.cose.2018.04.002 · Cited by 511 · cited 2×
“Also, it may be due to the well-known privacy paradox where individuals tend to claim that they care deeply about their privacy but continue to use tools that may be harmful to it [3, 22].”
User Perceptions of Smart Home IoT Privacy
Zheng, S., Apthorpe, N., Chetty, M., & Feamster, N. (2018) · Proceedings of the ACM on Human-Computer Interaction · doi:10.1145/3274469 · Cited by 423 · cited 1×
“Many studies have investigated the users’ data-sharing behavior and privacy concerns in the context of different emerging technologies, e.g., overall investigations of smart cities [49], smart homes [55, 56], messaging apps [27], or social media [30].”
Unveiling Security, Privacy, and Ethical Concerns of ChatGPTPreprint
Wu, X., Duan, R., & Ni, J.-B. (2023) · Journal of Information and Intelligence · doi:10.48550/arxiv.2307.14192 · Cited by 365 · cited 1×
“[52] explored the potential risks of integrating ChatGPT into daily life, focusing on security, privacy, and ethical concerns.”
Shahsavar, Y., & Choudhury, A. (2023) · JMIR Human Factors · doi:10.2196/47564 · Cited by 341
Cited by
Understanding Users' Privacy Perceptions Towards LLM's RAG-based Memory
Zhang, S., Ma, R., Ma, Y., Li, S., Xu, Y., Yi, X., & Li, H. (2025) · Proceedings of the 2025 Workshop on Human-Centered AI Privacy and Security · doi:10.1145/3733816.3760750 · Cited by 6
Towards Usable, Privacy Respecting Long-Term Memory for LLM-based Conversational Agents
Malki, L. M. (2026) · CHI Conference Extended Abstracts on Human Factors in Computing Systems (CHI EA) · doi:10.1145/3772363.3799198 · Cited by 3
“Asitcurrently stands, user-centred privacy research addressing LLM-based CAs remains qualitative and exploratory [13, 14], or focuses on perceptual constructs such as privacy concerns and information disclosure [24, 31].”
Can LLMs Infer Conversational Agent Users' Personality Traits from Chat History?
Cögendez, D., Zimmermann, V., & Zufferey, N. (2026) · arXiv · doi:10.48550/arxiv.2604.19785 · Cited by 3
“For example, our data shows that at least half of the users shared data related to Job and Education , Lifestyle and habits , Family Life and Relationship , and Physical Health at least once, which corresponds to previous work [38].”
InfluentialSelf-citationPrivatar: A Privacy-Preserving Framework for LLM Agents
Jian, R., Zhu, Z., Shang, Y., Shen, Y., & Liu, K. (2025) · IEEE ... International Conference on Data Mining workshops · doi:10.1109/icdmw69685.2025.00011 · Cited by 1
“This raises pressing concerns about data privacy, regulatory compliance, and user trust [9].”
Designing Proactive Memory Privacy Visualization and Management in LLM-based ChatbotsPreprint
Zhang, S., Ye, L., Li, S., Xing, H., Jingyu, T., Shui, B., Nie, K., liu, P., Yi, X., & Li, H. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.7288197
Metadata
- Authors
- Zufferey, N., Abdelwahab Gaballah, S., Marky, K., & Zimmermann, V.
- Affiliations
- ETH Zurich (CH)
- Ruhr University of Bochum
- Ruhr University Bochum (DE)
- Venue
- Proceedings on Privacy Enhancing Technologies
- Published
- 2025-07
- Volume
- 2025
- Issue
- 3
- Pages
- 5-28
- ISSN
- 2299-0984 (ISSN-L)
- OpenAlex
- W4410475227
- Semantic Scholar
- e7273188d36c6effd531361a665e31a2b58d547b
- Cited by
- 6
- References
- 60
- Citations by year
- 2026: 4 · 2025: 2
- FWCI
- 10.45
- Type
- journal article
- Language
- en
- Open access
- hybrid
- doi.org
- Locations
- Proceedings on Privacy Enhancing Technologies (publishedVersion, cc-by)
- Topics
- Social Sciences › Sociology and Political Science › Privacy, Security, and Data Protection
- Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
- Physical Sciences › Computer Science › Artificial Intelligence › Privacy-Preserving Technologies in Data
- Keywords
- large language models, chatbots, personal data sharing, privacy concerns, user behavior, local LLM deployment, health data, user trust in AI, data leakage
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: 'AI is from the devil.' Behaviors and Concerns Toward Personal Data Sharing with LLM-based Conversational Agents. https://app.cochat.ai/papers/10.56553/popets-2025-0086
- The paper
- Zufferey, N., Abdelwahab Gaballah, S., Marky, K., & Zimmermann, V. (2025). 'AI is from the devil.' Behaviors and Concerns Toward Personal Data Sharing with LLM-based Conversational Agents. Proceedings on Privacy Enhancing Technologies. https://doi.org/10.56553/popets-2025-0086
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.