Transparency Doesn't Fix AI—Control Does

Users trust biased ML models more when explainability visualizations are harder to understand. The easier it is to see the bias, the less they trust the system.

Method: Researchers tested five visualization types (bar charts, confusion matrices, decision trees, feature importance plots, and SHAP plots) across 180 participants. The inverse relationship held across all formats: participants who scored higher on comprehension tests showed significantly lower trust scores. SHAP plots, despite being the most detailed, produced the highest trust among users who couldn't actually interpret them correctly.

Caveats: Study focused on classification tasks with demographic bias. Doesn't address regression models or other bias types like sampling bias.

Reflections: Does the inverse relationship persist when users have domain expertise in ML fairness? · Can hybrid visualizations balance comprehension and trust without sacrificing either? · How does repeated exposure to biased model outputs affect the comprehension-trust relationship over time?

People overestimate support for extreme policies among political opponents by 20-30 percentage points. This misperception fuels partisan violence and anti-democratic sentiment.

Method: Tested four visualization formats (bar charts, icon arrays, distribution curves, and text-only) to correct misperceptions about outparty policy support. Icon arrays—where each icon represents a person—reduced misperceptions by 15% more than text alone and increased willingness to engage with opponents by 12%. The effect persisted one week later. Distribution curves performed worst, likely because they require statistical literacy.

Caveats: Effects measured immediately and at one week. Long-term persistence unknown. Tested only U.S. partisan contexts.

Reflections: Do icon arrays work for correcting misperceptions in non-political domains like healthcare or workplace diversity? · What's the threshold where icon arrays become too cluttered to process effectively? · Can adversarial actors weaponize these formats to amplify misperceptions instead?

Instructors point, circle, and sketch on slides without verbal descriptions. Low-vision learners miss these cues entirely or waste cognitive load hunting for visual indicators.

Method: VeasyGuide uses motion detection to identify instructor actions (pointing, marking, sketching) and generates personalized audio descriptions based on user-specified preferences: verbosity level, description timing, and action priority. In a study with 12 low-vision participants, users who set high verbosity caught 89% of instructor actions versus 34% without the tool. Users could toggle between real-time narration and on-demand summaries.

Caveats: Co-designed with only three low-vision participants. Motion detection fails when instructors move off-camera or use subtle gestures.

Reflections: Can the system distinguish between meaningful gestures and fidgeting or pacing? · How do sighted learners respond to the same personalized audio—does it reduce their cognitive load too? · What's the optimal delay between action detection and audio narration to avoid disrupting flow?