Does Explanation Presentation Format Affect How Users Assess and Understand AI?
A between-subjects study comparing a narrative Data Story and a Q&A interface, asking whether format shapes the quality and accuracy of user critiques of an AI system.
Does the format of a training data explanation change how accurately users can assess and critique an AI system?
Between-subjects experiment (n = 39) comparing a narrative Data Story against a structured Q&A format, with identical content across both conditions.
Format determined accuracy. Data Story produced 66% high-accuracy critiques; Q&A produced 42%. Same information, different packaging, measurably different outcomes.
Overview
Prior work showed users value training data transparency. But how do they actually use that information? And does the format it comes in change what they do with it? I designed two contrastive explanation formats and ran a study to find out.
Designed both explanation conditions from scratch: a scrollable Data Story with annotated visuals and infographics, and a Q&A interface replicating the CHI 2021 format. Both presented identical training data content.
Recruited and ran all 39 participants in-person and remotely, managing session logistics and participant communication across both conditions. Sessions averaged 90 minutes.
Performed initial iterative coding of all 352 SWOT comments for topic coverage, bias type, and accuracy alongside a second researcher. Ran ANCOVA with AI literacy as covariate and chi-squared tests.
Conducted and analyzed all semi-structured post-study interviews via thematic analysis to surface attitudinal patterns and interpretive strategies across both conditions.
The Problem
We know users value training data explanations. What we do not know is how they use and comprehend the information, and whether the format it comes in changes what they actually do with it. Interface designers make format decisions constantly; this study asks whether those decisions have measurable consequences for critique quality.
Design Question
If two users receive exactly the same training data information: one as a narrative and one as a structured Q&A, do they end up equally equipped to critique the AI system? This study says no.
Research Questions
What types of training dataset information do users draw on when critiquing an automated system, and does presentation style affect which information they use?
Does presentation style affect the accuracy of user critiques and their ability to identify data biases?
How do subjective impressions of the explanation differ between formats?
Methods
Between-Subjects Experiment
39 participants randomly assigned to Data Story (n=19) or Q&A (n=20). Same training data content, different format.
SWOT Analysis
Participants produced structured critiques identifying Strengths, Weaknesses, Opportunities, and Threats in the AI hiring system based on its training data explanation.
Content Coding
Two researchers coded all 352 SWOT comments for topic coverage, bias type, and accuracy level. ANCOVA with AI literacy as covariate.
Semi-Structured Interviews
Post-study interviews probed attitudes toward AI, perceptions of the explanation, and reasons behind critique choices. Analyzed via thematic analysis.
Research Process
The Gap
CHI 2021 showed users value training data transparency. But that work did not study how users actually use or comprehend the information. The format question was entirely open.
Designing the Data Story
I drew on data storytelling literature to build a scrollable, author-driven narrative interspersed with annotated visualizations and infographics. The Q&A interface replicated the structured format from prior work. Both presented identical content about the same fictitious hiring dataset.
Participants were placed in the role of an HR employee tasked with recommending whether to purchase an automated hiring system, based solely on its training data explanation. They used their assigned format to complete a SWOT analysis, then answered Likert-scale questionnaires on trust, fairness, and understanding, and finished with a semi-structured interview. Sessions averaged 90 minutes.
Format shaped critique accuracy. Data Story participants produced significantly more accurate critiques: 66% rated as high understanding vs. 42% in the Q&A.
Format directed attention differently. Data Story participants focused more on demographics and instance-level data. Q&A participants focused more on dataset overview information.
Both formats surfaced bias. Most participants identified at least one data bias. Data Story led to more Representation and Measurement biases; Q&A led to more Temporal biases.
Q&A felt more educational. A trend suggested participants felt they learned more about ML/AI from the Q&A, despite producing less accurate critiques.
The Two Conditions
The study only works if both formats contain identical information. The design challenge was ensuring neither felt more thorough than the other while making them genuinely distinct in interaction model.
Same training data content. Different interaction model. Different critique outcomes.
I would've gone with the system without any thinking if I didn't see the explanation. I would've bought the system. If I don't have any information on what it's doing, I wouldn't care, because it's making my life easier.Participant in the Q&A condition
Key Insights
Format is not neutral
Presentation style systematically shaped which information users drew on, what biases they identified, and how accurate their critiques were, even when the underlying content was identical.
Narrative drives accuracy; Q&A drives perceived learning
Data Story participants produced more accurate critiques. Q&A participants felt they learned more about AI/ML. Accuracy and sense of learning pointed in opposite directions.
Visuals direct attention more than structure does
The Data Story's annotated graphics led participants to focus on demographics and instance-level information. The Q&A's structure directed attention toward dataset overview. Format determines what users notice.
Neither format is universally better
Each format has genuine advantages. Data Story is preferable when accurate critique is the goal. Q&A is preferable when the goal is helping users feel they are learning about the system.
Design Implications
Core Insight
Format is a design decision with measurable consequences. Holding content constant and varying only presentation style produced significantly different critique outcomes. Explanation designers cannot treat format as neutral.
- Match format to the evaluation goal. If the goal is accurate critique of an AI system, a narrative format with integrated visuals is preferable. If the goal is giving users a sense of learning about ML, structured Q&A has advantages.
- Use visuals and repetition to direct attention. The Data Story's annotated graphics and infographic slideshows drew participants to demographic imbalances and processing decisions. Format choices determine what users notice.
- Prefer accuracy over the feeling of learning. The Q&A format created a stronger sense of learning while producing less accurate critiques. Subjective impressions of comprehension do not reliably track actual comprehension.
- Consider hybrid designs. The tradeoffs observed suggest an ideal format might combine narrative structure with on-demand depth, offering coherent framing while preserving user agency.
Impact
For product teams: when the goal is helping users evaluate an AI system accurately (procurement, auditing, onboarding), format matters as much as content. A narrative with integrated visuals outperformed a structured Q&A on critique accuracy by 24 percentage points, with identical underlying information. Choosing the wrong format for the task is not a neutral decision. Published at IEEE VL/HCC 2024.