TLDR: A new research paper explores how different methods of presenting factuality in AI-generated text impact user trust. The study found that strategies which hide or make less factual content ambiguous (Opaque and Ambiguity) lead to higher user trust and perceived answer correctness, challenging the idea that full transparency about AI errors is always the best approach. This suggests that carefully managing the disclosure of uncertain information can improve user confidence in AI systems.
Large Language Models (LLMs) have revolutionized many aspects of our digital lives, but they come with a significant challenge: the tendency to generate information that sounds convincing but is factually incorrect, a phenomenon often called “hallucinations.” This issue can lead to serious consequences, from financial losses for companies to misinformed decisions by individuals.
For years, researchers have explored various methods to communicate the factuality of AI-generated content to users, aiming to prevent blind trust and erroneous decisions. A common approach has been to highlight parts of the AI’s output that are estimated to be less factual, or conversely, to highlight the factual parts. However, a recent study by Hyo Jin Do and Werner Geyer from IBM Research delves into a less explored but crucial question: What if, instead of revealing content estimated to be factually incorrect, we hide it altogether?
Rethinking Factuality Communication
The research, titled “Hide or Highlight: Understanding the Impact of Factuality Expression on User Trust”, challenges the conventional wisdom that more transparency is always better. In human-human communication, we often avoid stating things we believe to be false to maintain trust. This paper explores whether a similar principle applies to human-AI interaction.
The study tested five distinct strategies for presenting AI-generated answers with factuality assessments in question-answering scenarios:
- Baseline: The original AI response is shown without any factuality information.
- Transparent: Content estimated to be less factual is highlighted (e.g., in orange).
- Attention: Content estimated to be highly factual is highlighted (e.g., in blue).
- Opaque: Less factual content is simply removed from the response, often indicated by a placeholder like “[..]”.
- Ambiguity: Less factual content is replaced with vague statements that are not factually incorrect, reducing precision rather than removing content entirely.
Key Findings: Hiding Builds Trust
The researchers conducted a human subjects study with 148 participants, who engaged in question-answering tasks using AI-generated biographies. Participants were asked to act as journalists and verify statements based on the AI’s output and a reference link (Wikipedia).
The most striking finding was that the Opaque and Ambiguity strategies consistently led to higher user trust. Participants in these conditions reported significantly higher “Trust Belief” (perceptions of the system’s trustworthiness, ability, benevolence, and integrity) compared to those in the Transparent, Attention, and Baseline conditions. Furthermore, these strategies resulted in higher “Appropriate Compliance,” meaning users were better at accepting correct statements and rejecting incorrect ones.
Interestingly, the study also found that participants perceived the correctness of the AI-generated answers significantly higher in the Opaque and Ambiguity conditions. This suggests that by removing or neutralizing incorrect information, the actual correctness of the presented answer increased, leading to a higher perception of accuracy by users.
Contrary to some expectations, the Opaque and Ambiguity strategies did not negatively impact other perceived qualities of the AI answer, such as relevance, completeness, conciseness, or coherence. This implies that even with content removed or made vague, the overall quality of the response was maintained in the users’ eyes.
Also Read:
- Unpacking User Engagement with AI Explanations: A Surprising Look at Trust and Decision-Making
- Assessing Language Model Reliability in Varied Contexts
Implications for AI Design
This research offers significant insights for AI practitioners and application developers. It suggests that focusing solely on transparency by highlighting potential errors might not always be the most effective way to build user trust. In fact, it can sometimes erode it. Instead, strategies that actively hide or make less factual content ambiguous can be more beneficial for fostering trust and appropriate reliance on AI systems.
The Ambiguity strategy, in particular, is a novel approach that allows AI to accommodate varying interpretations without committing to precise but incorrect assertions. While generating ambiguous text might seem counterintuitive to the goal of clarity, this study demonstrates its practical utility in managing AI hallucinations and enhancing user trust.
The findings encourage a shift in design philosophy: rather than always exposing AI’s uncertainties, strategically managing the disclosure of low-factuality content can lead to a more positive and trustworthy user experience. This is especially relevant for general-purpose LLMs where accurate information is highly valued, such as in summarization or learning tasks.
