By Ray A. Groq · Editorial review: Julio C. Sandria Reynoso
Leer este artículo en español →
Autonomous vehicles promise safer roads, but they also force us to confront an unsettling question: when the algorithm faces an impossible situation, who decides whom it protects? This dilemma, akin to the famous trolley experiment, becomes real when a car must choose between running over a pedestrian or swerving and endangering its occupants. In this article we will explore how scientists and designers are trying to answer that question, from massive surveys to algorithms that learn from our own decisions. All of this with the goal that, when the machine has to decide, it does so in a way that reflects our collective values.
The trolley dilemma transferred to asphalt
Imagine your autonomous car detects an imminent collision: on the left there is a group of pedestrians, on the right a single passenger. There is no time to brake. What should it do? This scenario is a modern version of the trolley dilemma, where an unavoidable decision entails sacrificing someone. Researchers use simple analogies, such as choosing between saving a family or a child, so that the public understands the gravity of the problem. In practice, algorithms must translate those ethical values into code that can be executed in milliseconds. The central question is who defines those priorities: the manufacturers, the legislators, or society as a whole? The answer is not unique, and therefore several approaches to data collection and decision modeling have been created.
What the Moral Machine experiment reveals
Awad et al. (2018) tested the trolley dilemma on more than 40,000 people worldwide using the Moral Machine platform. Participants had to choose between different combinations of victims, such as pedestrians crossing in a crosswalk or passengers in the car. The results showed cultural patterns: in individualist countries, there was a preference to save the passengers, while in collectivist societies, pedestrians were prioritized. Moreover, the majority showed a slight preference for saving children over adults. These findings indicate that there is no single answer, but that morality varies by region and culture. Therefore, any automated decision system should be able to adapt to those local values, or at least be transparent about the criteria it employs.
Simulation platforms and data collection
To go beyond surveys, Behzadan et al. (2018) developed TrolleyMod v1.0, an open‑source tool that allows users to experiment with collision scenarios in real time. The platform records decisions, response time, and environmental context, creating a rich database for training algorithms. Being accessible, TrolleyMod has facilitated participation from researchers of various disciplines, from psychology to engineering. The collected data show that, when factors such as vehicle speed or visibility are added, decisions can change rapidly. This type of simulation helps identify which variables influence the ethical choice the most, providing a roadmap for designing systems that consider not only the number of victims but also the probability of harm and environmental uncertainty.
Algorithms that try to predict and plan ethical decisions
Pickering and Burnham (2022) proposed a model that estimates the severity of injuries in a collision and uses that information to guide trajectory planning. Their approach combines impact data with ethical criteria, prioritizing routes that minimize severe injuries. On the other hand, Wei and Wang (2023) employed an LSTM network to predict the trajectories of pedestrians and vehicles, integrating the prediction into moral decision-making. The algorithm evaluates multiple possible futures and selects the one that reduces expected harm. Geisslinger et al. (2022) presented a trajectory planning algorithm that incorporates ethical norms encoded as mathematical constraints, ensuring that the optimal solution respects principles such as “do not sacrifice innocent lives.”. These works demonstrate that, although technology advances, the integration of human values remains a technical challenge that requires balancing predictive accuracy and moral considerations.
Human feedback learning: the next frontier
Brice et al. (2026) introduced the Ethical Decision Head, a system that learns ethical norms through reinforcement from human feedback. Instead of programming fixed rules, the model observes decisions that users approve or reject and adjusts its policies accordingly. This approach allows the vehicle to evolve over time, adapting to cultural changes or new safety standards. Moreover, by making the source of feedback explicit, it opens the door to audits and accountability. However, the authors warn that the quality of human data is crucial: biases or inconsistent decisions can lead to undesirable behaviors. Therefore, combining controlled simulations, such as TrolleyMod, with continuous learning could provide a more robust framework for autonomous cars to make decisions that truly reflect our convictions.
Conclusion
In the end, the question of who decides when the autonomous car must choose between two evils does not have a simple answer. Research shows that collective morality varies, that simulations can capture that diversity, and that algorithms are learning to translate human values into concrete actions. Yet the ultimate responsibility rests on us: designers, regulators, and users must collaborate to define the principles that will guide these machines. Only with an open conversation and transparent data can we trust that, when the road becomes dangerous, the car’s decision reflects what we truly value as a society.
Key takeaways
- Ethical values about who should be protected vary by culture and context.
- Platforms like TrolleyMod enable the collection of real decisions to train algorithms.
- Current models combine injury prediction, trajectory planning, and human‑feedback learning.
- Transparency and public participation are essential to legitimize autonomous vehicle decisions.
Sources
- Awad, E. et al. (2018). The Moral Machine experiment — https://doi.org/10.1038/s41586-018-0637-6
- Behzadan, V. et al. (2018). TrolleyMod v1.0: An Open-Source Simulation and Data-Collection Platform for Ethical Decision Making in Autonomous Vehicles — https://arxiv.org/abs/1811.05594v1
- Pickering, J. E. and Burnham, K. J. (2022). Predicting Autonomous Vehicle Collision Injury Severity Levels for Ethical Decision Making and Path Planning — https://arxiv.org/abs/2212.08539v1
- Wei, W. and Wang, J. (2023). Ethical Decision-making for Autonomous Driving based on LSTM Trajectory Prediction Network — https://arxiv.org/abs/2308.01022v1
- Geisslinger, M. et al. (2022). An Ethical Trajectory Planning Algorithm for Autonomous Vehicles — https://arxiv.org/abs/2212.08577v1
- Mbrice, T. et al. (2026). The Ethical Decision Head: Operationalizing Normative Ethics in Autonomous Vehicles via Reinforcement Learning from Human Feedback — https://arxiv.org/abs/2608.16710v1
