Trust Map
Explore how unequal data distribution affects AI-generated insights across different sectors like health, crime, education, and more.
Scroll down to explore how we calculated the AI Trust Score and how it varies across different regions. The interactive map visualizes the AI Trust Score across the United States.
Understanding AI Trust Scores
The AI Trust Score is a composite metric ranging from 1-5 that measures measures the quality of AI Overviews in a particular county compared to the other. Counties with higher scores (green) indicate that the county has higher quality AI Overviews than the other counties.
How Scores Are Calculated
The AI Trust Score is the average of three components:
- AI Availability
- Source Diversity
- Epistemic Humility
Each component or score is placed on a scale from 1 to 5 based on how it compares to others.
Standard deviation helps us understand how spread out scores are from the average. When we say something is "one standard deviation above the mean," it means that score is higher than most others and stands out as above average. Most scores usually fall close to the average, but the further a score is from the mean, the more unusual it is. For example, scores more than one standard deviation above the mean are considered high, while those more than one standard deviation below are considered low. This makes standard deviation a powerful way to group scores fairly. Instead of just looking at the number itself, we understand how that number compares to the rest of the group. It tells us whether a score is typical, slightly better or worse than average, or truly exceptional.
- Highest (5): x > μ + σ
- High (4): μ + σ > x > μ
- Moderate (3): μ > x > μ - σ
- Low (2): μ - σ > x > μ - 2σ
- Lowest (1): x < μ - 2σ
AI Availability
The AI Availability Score is calculated using two key metrics:
Calculation Process:
- Apply min-max normalization to both Trigger Rate and Average Content Score
- Sum the normalized values to get a combined score
- Apply final min-max normalization to obtain scores between 0-100
Final scores are mapped based on these thresholds:
- Highest (5): ≥ 71.35
- High (4): ≥ 55.86
- Moderate (3): ≥ 40.36
- Low (2): ≥ 24.87
- Lowest (1): < 24.87
Source Diversity
The Source Diversity Score measures how evenly information sources are distributed across different domains for each county.
Calculation Process:
- Calculate Shannon Index (H):
- H = -Σ(p × ln(p))
- p = proportion of sources from each domain
- Normalize by domain count:
- H_normalized = H / ln(k)
- k = number of unique domains
- Convert to 0-100 scale for final score
Final scores are mapped based on these thresholds:
- Highest (5): ≥ 100.7
- High (4): ≥ 91.7
- Moderate (3): ≥ 82.7
- Low (2): ≥ 73.6
- Lowest (1): < 73.6
Epistemic Humility
The Epistemic Humility Score measures how well AI systems acknowledge uncertainty in their responses.
Calculation Process:
- Classify confidence levels using BART:
- Confident responses (1.0)
- Neutral responses (0.5)
- Uncertain responses (0.0)
- Calculate initial score: 1 - (weighted sum of confidence)
- Apply min-max normalization for final 0-100 score
Final scores are mapped based on these thresholds:
- Highest (5): ≥ 58.12
- High (4): ≥ 38.19
- Moderate (3): ≥ 18.26
- Low (2): ≥ -1.67
- Lowest (1): < -1.67
Meet the team
Loading Map...
Low Trust
High Trust