Bridging the Knowledge Gap: Understanding User Expectations for Trustworthy LLM Standards

Michaela Benk, Léane Wettstein, Nadine Schlicker, Florian von Wangenheim, Nicolas ScharowskiAAAI 2025 · DOI

Abstract While experts and policymakers are busy drafting "top-down" ethical standards for Large Language Models (LLMs), a massive knowledge gap exists between these frameworks and the people actually using the tools. This study uses a mixed-methods approach to identify 68 specific criteria users look for to build trust, revealing that users prioritize privacy and safety over technical explainability and are deeply confused about who is actually accountable when an LLM fails.

Introduction

We are currently in an era of "top-down" AI governance. Organizations like the ISO and the EU are defining what "trustworthy AI" looks like, but these high-level principles—like "fairness" or "accountability"—often function as "empty vessels" that mean little to the average user. Trustworthiness is inherently subjective: it only exists if a user’s values align with the system’s implementation. If users don't understand the standards or can't find the "cues" that a standard has been met (e.g., a certification label or a specific interface feature), the standards fail to build actual trust. This paper seeks to bridge this gap by taking a "bottom-up" perspective—asking users what they actually need to see to trust an LLM.

Methods

The researchers employed a three-phase mixed-method approach:

  1. Literature Synthesis: Identification of 13 frequently cited high-level standards (e.g., Privacy, Robustness, Sustainability) from existing AI ethics frameworks.
  2. Qualitative Interviews (N=12): Deep-dive interviews with lay users to uncover a taxonomy of 68 user-relevant criteria—the specific, granular signals users use to evaluate high-level principles.
  3. Quantitative Survey (N=379): A larger online study where participants ranked the importance of the high-level standards and performed pairwise comparisons of the 68 individual criteria to determine which specific "trust signals" mattered most.

Key Results

The quantitative analysis revealed a clear hierarchy of user concerns. Privacy emerged as the "North Star" of user trust, followed closely by Safety and Security. Conversely, Sustainability was ranked as the least important standard, suggesting a significant disconnect between the high energy costs of LLMs and user concerns.

Figure 1: Radar chart illustrating the relative importance of 13 high-level trustworthiness standards as perceived by users.

Interestingly, users often view Safety and Security through the lens of data protection rather than technical robustness. Despite the massive research focus on Explainable AI (XAI), users in this study prioritized transparent information about data handling and system governance over technical explanations of how the model reached a specific conclusion. In pairwise comparisons for Privacy, "Data Protection" (e.g., secure login) was favored significantly more than "Transparent Information Provision," suggesting users want active protection over long privacy policies.

Figure 2: Pairwise comparison heatmap for Privacy criteria, showing preferences between granular trust signals.

Discussion: The SEC Taxonomy for LLMs

The authors mapped these findings onto the SEC (Search, Experience, Credence) framework from information economics to categorize how users verify if a standard is being met:

  • Search-based: Standards users can verify before use (e.g., checking an anonymity policy).
  • Experience-based: Standards verified during use (e.g., seeing a "ChatGPT can make mistakes" disclaimer).
  • Credence-based: Standards that are nearly impossible for users to verify even after use (e.g., checking if training data was unbiased). These require third-party "facework" like certification labels.

A major finding was the "Accountability Void": users are profoundly uncertain about who is responsible for LLM outputs. While they value accountability, they lack the "mental models" to understand how to contest an output or who to blame (the developer, the provider, or themselves).

Conclusion

Bridging the gap between top-down standards and bottom-up user expectations is critical for the adoption of trustworthy AI. Future work should explore how different communication channels (like interface cues vs. technical documentation) affect trust across different demographics and specific LLM use cases. The authors call for more research into "credence-based" standards—how to prove a model is fair when the user has no way of seeing the underlying data.

Summary updated

BibTeX
@inproceedings{benk2025bridging,
  author = {Michaela Benk and Léane Wettstein and Nadine Schlicker and Florian von Wangenheim and Nicolas Scharowski},
  title = {{Bridging the Knowledge Gap: Understanding User Expectations for Trustworthy LLM Standards}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i26.34928},
}
Figure