Highlight All the Phrases: Enhancing LLM Transparency Through Visual Factuality Indicators

Hyo Jin, Rachel Ostrand, Philipp Geyer, Keerthiram Murugesan, Dennis Wei, Justin D. WeiszAAAI 2025 · DOI

Abstract

Large language models (LLMs) are susceptible to generating inaccurate or false information, often referred to as "hallucinations" or "confabulations." While several technical advancements have been made to detect hallucinated content by assessing the factuality of the model's responses, there is still limited research on how to effectively communicate this information to users. To address this gap, we conducted two scenario-based experiments with a total of 208 participants to systematically compare the effects of various design strategies for communicating factuality scores by assessing participants' ratings of trust, ease in validating response accuracy, and preference. Our findings reveal that participants preferred and trusted a design in which all phrases within a response were color-coded based on factuality scores. Participants also found it easier to validate accuracy of the response in this style compared to a baseline with no style applied. Our study offers practical design guidelines for LLM application developers and designers, aimed at calibrating user trust, aligning with user preferences, and enhancing users' ability to scrutinize LLM outputs.

Introduction: The Challenge of LLM Hallucinations

Large Language Models (LLMs) frequently generate "hallucinations"—content that is linguistically fluent but factually incorrect. While automated methods for detecting these errors are improving, a critical human-computer interaction (HCI) challenge remains: how to present these factuality assessments to users. Effective communication is essential for "trust calibration," helping users identify when to rely on a model and when to be skeptical. Without clear indicators, users may either over-trust inaccurate outputs or under-trust accurate ones.

Methodology

The researchers conducted two scenario-based experiments involving 208 participants to evaluate different visual design strategies. The study compared three visual styles across two levels of linguistic granularity:

Visual Styles

  1. Highlight-all: The entire response is color-coded (red-to-green) based on factuality scores.
  2. Highlight-threshold: Only sections with low factuality scores (below 0.5) are highlighted to signal inaccuracies.
  3. Score: Numerical factuality scores are displayed in small bubbles or underlines next to the text.

Granularity Levels

  • Term-level: Factuality is assessed word-by-word.
  • Phrase-level: Factuality is assessed for groups of words or short clauses.
Visual design strategies across term and phrase granularity levels.

Experiment 1 utilized manually curated responses regarding general knowledge (e.g., Beyoncé's filmography). Experiment 2 employed automated methods, using the spaCy library to calculate semantic similarity between LLM responses and source documents in specialized domains like HR and Medicine.

Key Insights

  • The "Phrase" Sweet Spot: Participants preferred phrase-level granularity over term-level. Highlighting every individual word often resulted in "visual noise," making the text difficult to read, whereas phrases provided context without being overwhelming.
  • Trust Calibration: Visual indicators did more than just increase trust; they calibrated it. Users who initially over-trusted incorrect responses became more cautious when seeing red highlights, while skepticism toward correct responses decreased when green highlights were present.
  • Preference for "Fuzziness": Users favored intuitive color highlights over precise numerical scores. Interpreting a number like "0.74" requires more cognitive effort than processing a color cue.
  • The Value of Positive Reinforcement: "Highlight-all" was preferred over "Highlight-threshold." Seeing green highlights for correct information served as a "safety signal," making the red highlights for errors more impactful.

Key Results

The study found that transparency significantly improves user experience and trust compared to a standard "black box" response.

Trust and Ease of Validation ratings across different design strategies.
  • The Optimal Design: The phrase-level highlight-all strategy emerged as the most preferred and trusted design. It was also rated as the easiest for validating accuracy.
  • Ease of Validation: In Experiment 1, participants found it significantly easier to validate accuracy using the phrase-level highlight-all design compared to the baseline ($p = .03$).
  • Trust Impact: All six design strategies significantly increased user trust compared to the baseline ($p < .001$).
  • Granularity Preference: In Experiment 2, phrase-level designs were rated as significantly easier to validate than term-level designs ($p < .001$).

Limitations and Future Directions

While promising, the study has several limitations:

  • Demographics: Participants were IBM employees, who may be more tech-savvy than the general public.
  • Task Scope: The research focused on Q&A tasks; factuality indicators might be distracting in creative contexts like storytelling.
  • Ground Truth: The study assumes source documents are 100% accurate, which may not hold in real-world Retrieval-Augmented Generation (RAG) systems.
  • Accessibility: The red-green color scale is not accessible for color-blind users. Future research should explore alternative encodings like patterns or icons.

Summary updated

Abstract

Large language models (LLMs) are susceptible to generating inaccurate or false information, often referred to as "hallucinations" or "confabulations." While several technical advancements have been made to detect hallucinated content by assessing the factuality of the model's responses, there is still limited research on how to effectively communicate this information to users. To address this gap, we conducted two scenario-based experiments with a total of 208 participants to systematically compare the effects of various design strategies for communicating factuality scores by assessing participants' ratings of trust, ease in validating response accuracy, and preference. Our findings reveal that participants preferred and trusted a design in which all phrases within a response were color-coded based on factuality scores. Participants also found it easier to validate accuracy of the response in this style compared to a baseline with no style applied. Our study offers practical design guidelines for LLM application developers and designers, aimed at calibrating user trust, aligning with user preferences, and enhancing users' ability to scrutinize LLM outputs.
BibTeX
@inproceedings{jin2025highlight,
  author = {Hyo Jin and Rachel Ostrand and Philipp Geyer and Keerthiram Murugesan and Dennis Wei and Justin D. Weisz},
  title = {{Highlight All the Phrases: Enhancing LLM Transparency Through Visual Factuality Indicators}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aies.v8i1.36588},
}
Figure