Introduction
On-demand ridesharing services, such as Uber Pool or Lyft Shared, offer significant potential for reducing urban traffic congestion and lowering $CO_2$ emissions. Despite these societal benefits, adoption rates remain lower than expected. A primary barrier is user satisfaction: passengers often perceive the inconvenience of additional stops or longer routes more acutely than the cost savings or environmental impact. This paper investigates how AI-driven explanations can bridge this gap by selectively highlighting the benefits of shared rides to improve passenger sentiment.
The Explanation Selection Problem
When a shared ride concludes, an automated system has access to a wealth of data. It could inform the passenger about their savings relative to a private taxi, the time lost compared to a direct route, or their efficiency relative to public transit. However, providing all available information—a strategy suggested by classical game theory—can overwhelm users or inadvertently highlight negative aspects of the experience.
The authors define the "explanation selection problem" as the challenge of identifying which specific pieces of information maximize a human user's satisfaction. This requires moving beyond purely rational models to account for human psychological biases and preferences.
Theoretical vs. Empirical Approaches
The researchers evaluated two distinct types of agents to solve this problem:
1. The Rational Baseline (PBE Agent)
The interaction was first modeled as a signaling game between an AI agent and a passenger. Using the Perfect Bayesian Equilibrium (PBE) framework, the authors mathematically demonstrated that if both the agent and the passenger are perfectly rational "Econs," the agent must disclose all available information. In this theoretical model, full disclosure is the only way to maximize the passenger's estimate of the ride's value, as withholding information might lead the passenger to assume the worst about the undisclosed metrics.
2. The Human-Centric Agent (AXIS)
Recognizing that human passengers do not always behave like rational agents, the authors developed AXIS (Automatic eXplainer for Increasing Satisfaction). AXIS is a machine-learning agent designed to predict which explanations will resonate most with human psychology.
- Architecture: AXIS utilizes a neural network with two hidden layers trained on human-labeled data.
- Features: The model processes seven key inputs, including absolute and relative differences in cost and time, as well as environmental metrics like $CO_2$ savings.
- Persuasive Strategy: Unlike the PBE agent, AXIS employs "anchoring" and "framing" techniques. It selects specific benchmarks (anchors) and presents data in ways that influence the user's perception of the service's value.
Methodology and Data
To train and validate AXIS, the researchers combined real-world data with human feedback:
- Ride Simulation: Using the NYC Green Taxi Trip Data (2016), the authors simulated shared-ride scenarios. Since the original data only contained private trips, they employed a brute-force algorithm and a Traveling Salesman Problem (TSP) solver to generate realistic multi-passenger routes.
- Human Feedback: The team conducted four experiments via Amazon Mechanical Turk, collecting data from 953 participants. These participants rated their satisfaction with various ride scenarios and the suitability of different explanations, providing the "ground truth" for the AXIS training set.
Key Insights into Human Psychology
The study revealed several critical nuances in how humans process explanations:
- The Power of the "Private Taxi" Anchor: Comparisons to private taxis were significantly more effective at increasing satisfaction than comparisons to public transit. Psychologically, users view shared rides as a variation of a taxi service rather than an upgraded bus service, making the taxi the most relevant baseline.
- Framing Matters: The researchers found that users prefer different perspectives depending on the metric. For absolute differences (e.g., "You saved $5"), users preferred the shared-ride perspective. For relative differences (e.g., "A taxi would have cost 50% more"), they preferred the alternative perspective.
- Strategic Silence: AXIS learned that in some scenarios, the best explanation is no explanation. If all available comparisons are mediocre or negative, providing information can actually decrease satisfaction.
Results
In human trials, the AXIS agent significantly outperformed both the "rational" PBE agent and a random selection baseline ($p < 0.001$).
- Satisfaction Gains: On a 1–7 satisfaction scale, AXIS consistently achieved higher ratings by providing a curated subset of explanations (averaging 2.5 explanations per ride) rather than the full data dump required by the PBE model.
- Predictive Power: The neural network achieved 74.9% accuracy in identifying the most suitable explanations for a given ride scenario.
Conclusion and Future Directions
The success of AXIS demonstrates that in human-AI interaction, more information is not always better. By prioritizing psychological relevance over mathematical completeness, AI agents can more effectively influence user behavior and satisfaction.
Future work may involve moving beyond template-based language to Natural Language Generation (NLG) for more natural interactions. Additionally, personalizing AXIS to individual user profiles—identifying who cares more about time, money, or the environment—could further enhance the effectiveness of these persuasive explanations.
Summary updated