Abstract
Artificial Intelligence systems are often viewed as "black boxes," where a disconnect exists between complex algorithmic logic and user understanding. This research proposes a shift from purely technical explanations toward user-centric interface design. By applying Human-Computer Interaction (HCI) principles, the author develops a framework for "selective transparency" that adapts to user expertise and task context, aiming to build deeper trust through improved human-AI collaboration.
Introduction
As AI becomes more integrated into daily tasks, the gap between a user’s "mental model" (how they think the system works) and the system’s actual "conceptual model" becomes a significant barrier. Current transparency efforts often fail by providing either too much technical detail or irrelevant information. This research addresses this challenge by reframing transparency as a design problem, focusing on interfaces that empower users to interpret and control AI effectively.
The Transparency Framework
The author proposes that transparency is not a single feature but an amalgamation of three core components: the AI Model (the underlying algorithms), the Data (inputs and outputs), and the User Interface (the visual and interactive layer).
By treating transparency as the intersection of these elements, the framework allows for the application of usability principles—such as "keeping users in control" and "progressive disclosure"—to the explanation process itself.
Methodology
The research plan involves developing two distinct classes of AI systems: AI-powered text editors (Class A) and clinical diagnosis software (Class B). These systems are designed with a dual-mode operation strategy:
- Regular Mode: Utilizes "selective transparency," providing explanations primarily during "conversation breakdowns"—situations where the AI encounters ambiguity or uncertainty.
- Transparent Mode: Offers exhaustive disclosure for advanced users who need to audit the system’s inner workings.
As shown in Figure 1, selective transparency helps resolve issues with words that have multiple meanings, ensuring the user understands the AI's reasoning without being overwhelmed by data.
Data and Evaluation
Because the focus is on human-AI interaction, the study does not rely on standard ML training datasets. Instead, it uses:
- Prompt-based inputs: Ten zero-shot and one-shot prompts to test prototype responses.
- User Study Data: Qualitative and quantitative feedback gathered from controlled user groups to measure confidence and perceived transparency.
Key Insights and Discussion
- Transparency is Relative: Adequate transparency depends on the user’s expertise and the specific context of the task. What is helpful for a novice may be insufficient for an expert.
- Selective vs. Exhaustive: While researchers often aim for exhaustive disclosure, selective transparency is often more useful for end-users, preventing information overload.
- The "Placebo" Effect: Preliminary findings suggest that the mere presence of an explanation increases user satisfaction, even if the explanation does not provide new functional information.
Conclusion and Future Work
Preliminary results indicate that providing explanations—even "placebo" ones—significantly improves user perception of AI systems. However, the subjectivity of what constitutes "adequate" transparency remains a challenge. Future work will explore whether these HCI-driven principles can be generalized to higher-stakes autonomous systems and if they truly increase long-term user faith in AI compared to non-transparent interfaces.
Summary updated