TL;DR
This systematic review analyzes 155 research papers to map how users perceive privacy, security, and trust in Conversational AI (CAI). It reveals a fragmented research landscape where these three pillars are rarely studied together, highlighting an urgent need for standardized, CAI-specific measurement scales to replace outdated models borrowed from the 1990s.
What Problem Does This Paper Address?
Conversational AI (CAI)—ranging from voice assistants like Alexa and Siri to text-based chatbots—has become ubiquitous. However, as these systems integrate deeper into our private lives (smart homes, healthcare, banking), they trigger significant user anxiety regarding "constant listening," data misuse, and transparency.
The authors identify a critical gap: while developers want to build "trustworthy" AI, the research community lacks a unified understanding of what that actually means for a user. Existing studies often use "privacy," "security," and "trust" interchangeably or rely on measurement scales designed for 1990s e-commerce websites. This paper seeks to organize these fragmented concepts, evaluate the reliability of current research methods, and provide a roadmap for developing more robust, user-centric CAI.
Methods
The authors conducted a Systematic Literature Review (SLR) following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines.
- Search Strategy: They queried five major databases (ACM, IEEE, Scopus, Web of Science, and SpringerLink) using two distinct keyword sets: one for privacy/security perceptions and one for trust perceptions.
- Selection: Out of thousands of initial results, they narrowed the field to 100 papers on privacy/security and 58 on trust (with only 3 papers overlapping both categories initially).
- Analysis Framework: The papers were coded based on application fields (e.g., healthcare, retail), research methods (qualitative vs. quantitative), and the specific "subconstructs" measured (e.g., perceived risk, benevolence, or competence).
Datasets
As a systematic review, the "dataset" is the corpus of existing literature.
- Primary Corpus: 155 unique peer-reviewed articles and conference papers published between 2015 and 2023.
- Participant Demographics in Reviewed Papers: The authors noted that most studies relied on crowdsourced workers (Amazon Mechanical Turk, Prolific) or students, which may limit the generalizability of the findings to broader or more vulnerable populations (like the elderly or children).
Key Insights and Ideas
- The "Silo" Problem: Privacy, security, and trust are the "three pillars" of adoption, yet they are almost always studied in isolation. Only a tiny fraction of research looks at how a security breach specifically erodes trust in a conversational context.
- Outdated Foundations: Many "trust" studies still rely on the Mayer Model (1995). While foundational, this model was designed for human-to-human organizational trust. Applying it to a disembodied AI voice requires significant adaptation that many researchers skip.
- Construct Overlap: The review found that subconstructs like "Perceived Risk" are used to measure both privacy concerns and trust. This suggests that users don't necessarily distinguish between these concepts the way researchers do.
- Voice vs. Text Bias: Privacy and security research is heavily skewed toward voice assistants (smart speakers), likely due to the "always-on" microphone concern. Conversely, trust research is more evenly split between voice and text-based chatbots.
- The Rise of "Privacy Cynicism": A few recent papers highlight a shift where users feel they have no control over their data, leading to a "cynical" acceptance of the technology despite high privacy concerns.
Key Results
- Methodological Dominance: Quantitative methods rule the field, used in 60% of privacy/security papers and 81% of trust papers.
- Reliability Gap: Privacy and security scales demonstrated much higher statistical rigor. 76% of P&S papers sufficiently assessed internal consistency, compared to only 45% of trust papers.
- Validity Neglect: Only about 27% of trust-related papers reported on discriminant or convergent validity, suggesting that many "trust" findings in CAI may be built on shaky psychometric ground.
- Geographic Concentration: Research is highly centralized, with the USA (27%), UK (13%), and Germany (9-10%) producing the lion's share of the literature.
Limitations and Future Work
- Disembodied Focus: The review explicitly excluded "embodied" AI (physical robots), which might have different social trust dynamics.
- The LLM Gap: Because the search ended in mid-2023, the explosive impact of Large Language Models (LLMs) like GPT-4 on user perception is only briefly touched upon.
- Future Directions: The authors call for longitudinal studies (most current research is a one-time "snapshot") and the development of modular, CAI-specific scales that can be adapted for different contexts (e.g., a banking bot vs. a music-playing bot). They also emphasize the need to study "non-users"—those who refuse to adopt CAI—to understand the ultimate barriers to trust.
Summary updated