💻 computer science

Community Driving-Safety Deterioration as a Push Factor for Public Endorsement of AI Driving Capability

This study reveals that while community driving-safety concerns directly increase public endorsement of AI driving capabilities as a risk-mitigation strategy, this "push" effect is counterbalanced by a simultaneous suppression of general AI enthusiasm, resulting in a near-zero net impact on overall AI driving acceptance.

Original authors: Amir Rafe, Subasish Das

Published 2026-04-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Amir Rafe, Subasish Das

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Why We Might Want Robot Cars (But Also Fear Them)

Imagine you are standing at a crossroads. On one side, there is a Human Driver (the current way we travel). On the other side, there is an AI Driver (a self-driving car).

For years, researchers have asked: "What makes people want the AI driver?" They usually looked at the AI side of the road, asking things like: "Is the AI useful? Do I trust it? Is it cool?"

This study asks a different question: "Is the Human Driver side of the road so bad that it's pushing us toward the AI, even if we aren't sure about the AI yet?"

The researchers found a fascinating, two-sided story. It's like a tug-of-war happening inside people's minds.

The Two Forces at Play

1. The "Push" Factor (The Bad Road)

The Metaphor: Imagine your neighborhood street is a disaster. People are speeding, tailgating, texting while driving, and running red lights. It feels dangerous.
The Result: When people feel their local roads are unsafe, they think, "Okay, maybe a robot car would be safer than these crazy humans."
The Study's Finding: This is the Direct Effect. If you think your community driving is terrible, you are slightly more likely to say, "Yes, AI drivers are better than humans." This is the "Push" factor—dissatisfaction with the current system pushing you toward a new one.

2. The "Spillover" Factor (The General Mood)

The Metaphor: Now, imagine you are so stressed out by the dangerous driving on your street that you start feeling generally anxious about all technology. You start thinking, "If humans can't even drive safely, how can we trust a computer to do anything right? Maybe AI is dangerous too."
The Result: This is the Indirect Effect. Feeling unsafe in your community makes you generally less excited about Artificial Intelligence in any form (healthcare, jobs, etc.).
The Study's Finding: This is the "Risk Spillover." When people are worried about driving safety, their general enthusiasm for AI drops. And because they are less enthusiastic about AI in general, they are less likely to trust an AI driver specifically.

The Great Tug-of-War (The "Inconsistent Mediation")

Here is the twist: These two forces are fighting each other.

Force A (The Push): "My roads are bad, so I want AI drivers!" (Positive effect)
Force B (The Spillover): "My roads are bad, so I'm scared of all technology, including AI drivers." (Negative effect)

The Outcome: These two forces cancel each other out almost perfectly.

The "Push" makes people slightly more open to AI.
The "Spillover" makes people slightly less open to AI.
Net Result: The total effect is near zero.

The Analogy: Imagine you are trying to lose weight.

Force A: You start running every day (burning calories).
Force B: You start eating a massive pizza every night (gaining calories).
Result: Your weight stays exactly the same. You did a lot of work, but the two actions canceled each other out.

What About How Often You Drive?

The researchers wondered: "Does this change if you drive every day versus rarely?"

The Expectation: Maybe people who drive a lot are more annoyed by bad drivers, so the "Push" would be stronger.
The Reality: It didn't matter much. Whether you drive daily or rarely, the tug-of-war (the canceling out effect) happened the same way. The "Spillover" fear of technology was strong enough to neutralize the desire for a robot driver, regardless of how often you were behind the wheel.

The "Urban vs. Rural" Twist

There was one interesting difference based on where people lived:

Suburban and Rural Residents: The "Spillover" effect was strong. If they felt their roads were unsafe, they became generally wary of AI.
City Dwellers: The connection was weaker. City people might be used to seeing robots, delivery drones, and complex traffic systems, so a bad driving day didn't make them as scared of AI in general.

Why Does This Matter? (The Takeaway)

This study changes how we think about getting people to accept self-driving cars.

Fixing the Roads Might Backfire: If we fix traffic and make human driving super safe, people might actually become less interested in AI cars because they no longer feel the "Push" to switch.
Fixing the "Mood" is Key: The study suggests that the most powerful way to get people to accept AI drivers isn't just to fix the roads; it's to improve people's general attitude toward AI. If we can make people feel excited and safe about AI in general (through education and good experiences), that positive feeling will spill over and make them trust AI drivers, even if the roads are still a bit messy.

Summary in One Sentence

Feeling unsafe on the road makes you want a robot driver, but it also makes you scared of robots in general, so the two feelings cancel each other out—meaning we need to fix our general trust in technology, not just our traffic, to get people to accept self-driving cars.

1. Problem Statement and Research Gap

Road traffic crashes cause approximately 1.19 million deaths annually, primarily due to human error. While Autonomous Vehicles (AVs) offer a technological solution, existing literature on AV acceptance is dominated by technology-centered "pull" factors (e.g., perceived usefulness, ease of use, and trust in the AI system itself).

The study identifies three critical gaps in current research:

Neglect of "Push" Factors: Literature rarely examines whether dissatisfaction with the incumbent system (human driving) acts as a "push" factor driving acceptance of alternatives.
Asymmetric Safety Framing: Safety research focuses on the fear of AI risks rather than the perceived deterioration of human driving safety in one's community.
Mediating Mechanism: General attitudes toward AI (Generalized AI Orientation) are typically treated as background covariates or direct predictors, not as a mediating mechanism linking environmental stressors (community safety concerns) to specific technology judgments.

Research Questions:

Does Perceived Community Driving-Safety Concern (PCSC) predict evaluations of AI vs. human driving?
Is this relationship mediated by Generalized AI Orientation?
Does personal driving frequency moderate this relationship?

2. Methodology

The study employs a rigorous quantitative approach using a nationally representative probability sample.

Data Source: Pew Research Center's American Trends Panel (ATP), Wave 152 (August 2024).
- Sample Size: $N = 5,410$ (Effective sample size $n_{eff} = 3,993$ after weighting).
- Sampling: Address-based sampling with post-stratification weights for demographics and political affiliation.
Variables:
- Predictor (PCSC): Latent construct measured by 6 ordinal items (speeding, aggressive driving, impaired driving, cellphone distraction, cyclist/pedestrian risk).
- Mediator (Generalized AI Orientation): Latent construct measured by 7 items reflecting excitement vs. concern, trust, and perceived impact of AI.
- Outcome: Ordinal item comparing AI vs. human driving capability.
- Moderator: Personal driving frequency (ordinal).
- Controls: 12 covariates including demographics, prior AI exposure, and political ideology.
Analytic Strategy:
- Model: Weighted Structural Equation Modeling (SEM) with a Moderated Mediation design.
- Estimation: Weighted Least Squares Mean and Variance Adjusted (WLSMV) to handle ordinal indicators via polychoric correlation matrices.
- Inference: Bias-corrected bootstrap (10,000 resamples) for indirect effects.
- Robustness Checks: Seven pre-registered checks including Imai sensitivity analysis, E-value confounding thresholds, Propensity Score Matching (PSM), reverse mediation, and multigroup analysis by urbanicity.

3. Key Results

The analysis reveals a complex inconsistent mediation (suppression) pattern, characterized by a "risk-spillover" mechanism.

A. Measurement Model

Both PCSC and Generalized AI Orientation demonstrated excellent fit (CFI > 0.97, RMSEA < 0.08) and strong reliability ( $\omega > 0.84$ ).
The two constructs were empirically distinct (low inter-construct correlation $r = -0.07$ ).

B. Structural Paths

Path $a$ (PCSC $\to$ AI Orientation): Significant negative effect ( $\beta = -0.073, p < 0.001$ $β = - 0.073, p < 0.001$ ).
- Interpretation: Higher concern about community driving safety leads to lower general enthusiasm for AI. This suggests a "risk-spillover" where fear of human driving errors generalizes to a broader skepticism of AI technology.
Path $b$ (AI Orientation $\to$ AI Driving Eval): Significant positive effect ( $\beta = 0.293, p < 0.001$ $β = 0.293, p < 0.001$ ).
- Interpretation: Individuals with a generally positive orientation toward AI are significantly more likely to endorse AI driving capabilities.
Path $c'$ (Direct Effect: PCSC $\to$ AI Driving Eval): Significant positive effect ( $\beta = 0.031, p = 0.020$ $β = 0.031, p = 0.020$ ).
- Interpretation: Holding general AI attitudes constant, higher community safety concerns directly increase the preference for AI drivers (a domain-specific "push" effect).

C. Indirect and Total Effects

Indirect Effect: The mediated path (PCSC $\to$ $\to$ AI Orientation $\to$ $\to$ AI Driving) is negative and statistically significant across all levels of driving frequency (Low, Mean, High).
- Mechanism: Community safety concerns suppress general AI enthusiasm, which in turn reduces AI driving endorsement.
Total Effect: The net effect is near zero ( $0.037 + (-0.025) \approx 0.012$ $0.037 + (- 0.025) \approx 0.012$ ).
- Conclusion: The positive "push" from dissatisfaction with human driving is almost entirely canceled out by the negative "pull" of suppressed general AI enthusiasm.
Moderation: Driving frequency did not significantly moderate the indirect effect (Index of Moderated Mediation was non-significant). The mechanism operates consistently regardless of how often a person drives.

D. Robustness

Sensitivity: The indirect effect is sensitive to small unmeasured confounders ( $\rho^* = -0.020$ ), suggesting cautious interpretation is needed.
Urbanicity: The negative path from PCSC to AI Orientation was significant in suburban and rural areas but non-significant in urban areas, suggesting urban residents may have more diverse channels for AI attitudes that dilute this spillover.
Cross-Task Specificity: The direct "push" effect was specific to driving and news writing tasks but not medical or hiring tasks, indicating the mechanism is partially domain-specific.

4. Key Contributions

Theoretical: Introduces a Risk-Spillover Mechanism. It challenges the linear "push-pull" assumption by showing that dissatisfaction with the incumbent system (human driving) can simultaneously push users toward an alternative (AI) and pull them away by suppressing general trust in the technology class.
Methodological: Sets a new standard for AV acceptance research by utilizing a probability-based national sample with WLSMV-estimated SEM on ordinal data, moving beyond the convenience samples and simple regressions common in the field.
Empirical: Demonstrates that Generalized AI Orientation is an active mediator, not just a covariate. It reveals that the "safety" argument for AVs is psychologically complex: fearing human drivers makes people want AI, but fearing the environment makes them distrust AI generally.
Policy: Suggests that improving traffic safety (reducing the "push") might paradoxically slow AV adoption by removing the comparative advantage of AI. Conversely, improving AI literacy (boosting the mediator) is a more potent lever for increasing AV acceptance than fixing road conditions.

5. Significance and Implications

The study fundamentally reframes how public acceptance of AVs should be understood. It argues that acceptance is not merely a function of how "good" the AI is, but a tension between:

Domain-Specific Pragmatism: "Human drivers are dangerous; I need AI."
Domain-General Risk Aversion: "The world is dangerous and unpredictable; I don't trust AI either."

Policy Implications:

Dual-Track Interventions: Policymakers cannot rely solely on traffic safety campaigns to drive AV adoption. They must simultaneously address general AI literacy to prevent safety concerns from spilling over into a generalized distrust of AI.
Communication Strategy: Messaging should decouple general AI anxiety from specific driving contexts. Highlighting the specific failure of human drivers (the push) must be balanced with efforts to build general trust in AI systems to counteract the negative indirect effect.

Limitations:

Cross-sectional design limits causal inference (though reverse mediation was tested).
Sensitivity to unmeasured confounding (e.g., general risk aversion).
US-specific context may not generalize to cultures with different driving norms or AI regulatory environments.