Transparency Paradox explores how capability disclosures in LLM-based assistants can improve user understanding, trust, and ethical awareness, moving beyond superficial transparency to foster more responsible AI engagement.

Affiliation

New York University

Sep 2024 - Dec 2024

Contribution

Research

Transparency Paradox

A research study to explore how capability disclosures in LLM-based assistants can improve user understanding, trust, and ethical awareness.

With language model (LLM)-based assistants increasingly woven into daily tasks, users need more than basic feature lists—they need clear, proactive disclosures about capabilities and constraints. Our research addressed a core challenge: current assistants often rely on reactive transparency, only revealing their limits after users encounter failures or confusion. This practice can erode trust, hinder effective use, and mask underlying ethical and cultural complexities, resulting in what we termed the “transparency paradox.”

I developed a qualitative, framework-driven evaluation to systematically assess how three well-known LLM-based assistants—ChatGPT, Claude, and Gemini—communicated their capabilities and limitations. This included:

Framework Creation

Defining key categories (e.g., capability disclosure, file-processing transparency) to guide analysis.

Scenario-Based Evaluation

Testing each assistant across predefined scenarios to observe when and how they revealed constraints.

Thematic Analysis

Identifying patterns in observation that clarified or obscured user understanding and trust formation.

Framework Creation

Developing key categories to guide analysis

To systematically investigate how LLM-based assistants communicate their capabilities and constraints, I began by establishing a clear evaluation framework. Drawing on existing AI transparency literature and socio-technical research, I defined key categories such as “conversational capabilities disclosure,” “file processing transparency,” “knowledge timeline context,” and “onboarding experience.” These categories provided a structured lens through which I could assess each assistant’s approach to capability disclosure—measuring not only whether they provided information, but also when, how, and how thoroughly. By ensuring that the evaluation criteria were theory-driven and comprehensive, I set the stage for a consistent and meaningful analysis.

Scenario-Based Evaluation

Testing and observing across predefined scenarios

With the framework in hand, I crafted a series of predefined scenarios designed to elicit specific responses from the three chosen assistants: ChatGPT, Claude, and Gemini. These scenarios ranged from straightforward queries, like asking directly about capabilities, to more complex tasks, such as attempting unsupported file uploads or referencing events beyond the assistant’s known timeframe. Observing each system’s behavior in these diverse contexts allowed us to capture how transparency and limitation-disclosure naturally unfolded, mirroring real-world user encounters. Through careful scenario-based testing, we documented each assistant’s proactive or reactive responses, creating a rich dataset of user-system interactions.

Thematic Analysis

Identifying patterns in obvservations

After collecting responses from each scenario, I applied a thematic analysis process to identify patterns and insights within the data. The coding focused on the timing, clarity, and user impact of each disclosure—especially noting where proactive guidance improved mental models or where reactive revelation led to confusion. I grouped these codes into emergent themes that highlighted both effective and ineffective transparency strategies. This qualitative analysis clarified the underlying tensions of what I call the “transparency paradox,” revealing that even when assistants disclose their capabilities and constraints, they often fail to provoke critical reflection or address deeper ethical implications. These themes ultimately informed the proposed design recommendations and provided a nuanced understanding of how LLM-based assistants can foster more informed, equitable, and trustworthy engagement.

Findings

Our research shows that embedding proactive, context-aware transparency strategies into LLM-based assistants can improve user understanding and trust. By presenting capability constraints upfront, adapting explanations based on user behavior, and incorporating reflective prompts that highlight ethical considerations, designers can move beyond superficial disclosures. These user-centered design strategies help ensure that transparency serves as a meaningful guide for more informed, responsible, and equitable AI interactions.

Proactive Capability Disclosure

Providing upfront information about system abilities and limitations reduces guesswork and frustration.

Adaptive Contextual Cues

Adjusting disclosures based on users’ ongoing queries and actions helps clarify uncertainties before they lead to confusion.

Reflective Prompts & Ethical Considerations

Encouraging users to question outputs and consider underlying biases ensures transparency fosters critical thinking rather than false assurance.

Scenario-Based Examples

Demonstrations of how to leverage the system’s strengths and navigate constraints enable users to form clearer mental models.

Equitable and Inclusive Engagement

Involving diverse perspectives and ensuring that transparency strategies resonate with a wide range of cultural and social contexts supports more fair and responsible AI use.