The ‘Trolley Problem’ is a series of thought experiments that have long intrigued ethicists, psychologists, and AI researchers alike. Originally presented in 1967 by philosopher Philippa Foot and later popularised by fellow philosopher Judith Jarvis Thomson in 1976, the Trolley Problem poses a moral dilemma in the context of a runaway trolley. The scenario forces us to consider the monumental decision of whether to sacrifice one person to save many others.
At the heart of the Trolley Problem is the “switch” scenario. In this setup, a trolley is hurtling down a track towards five people who are unable to move. A bystander, standing by a lever, has the power to alter the trolley’s course. By pulling the lever, the trolley would be diverted onto a sidetrack, sparing the five people but resulting in the death of one person who is on the alternate track. The moral quandary posed is stark: should the bystander do nothing and allow the trolley to kill the five people, or should they intervene, sacrificing one to save many?
This ethical puzzle is not just theoretical but taps into deeper questions about human morality and decision-making under pressure. It challenges us to articulate and justify our moral intuitions and to examine how contextual details might influence our judgments. The Trolley Problem continues to be a vital tool in Cognitive Science, providing insights into the principles that govern our ethical reasoning, moral psychology, and the values we hold most dear.
Moral assessments are increasingly recognized as exhibiting variability dependent on the context. An expanding body of research has pinpointed various contextual elements (such as emotions and intentions) that can systematically sway moral judgments. However, synthesizing these diverse influences into a cohesive framework for understanding moral judgments poses a challenge. Railton (2017) recently endeavored to tackle this challenge by proposing a causal-evaluative modeling approach to moral judgment. Supporting this model, Railton offers evidence from new variations of classic moral dilemmas akin to the trolley problem.
In this essay, we will be exploring how Railton's paper explores foundational concepts of moral learning through these variations of the trolley problem; the author reasons that moral judgement is largely determined by both model-based and model-free learning processes that can accurately predict responses to ethical dilemmas and circumstantial moral perception.
Model-based learning occurs when humans construct and update cognitive models that structure causal and evaluative information based on feedback loops from the environment. These models evolve with new experiences and allow individuals to extend their moral reasoning beyond emotion or intuition --largely through 'imaginative projection' that predicts hypothetical scenarios and simulative empathy based on social interaction.
On the other hand, model-free learning relies on implicit cues that are taken from experiences that have formed pre-existing cognitive models which subsequently guide automatic responses without active deliberation. Model-free processes depend on stored 'caches' of expectation values or individual actions that create immediate, emotion-driven moral responses. Theoretically, this allows humans to make decisions intuitively without reasoning through the specifics of every circumstance, resulting in a moral judgement that is often difficult to explain from a 'logical' perspective.
Railton illustrates how this interplay between the two learning types can result in asymmetric responses through the classic Switch scenario -- a variation of the sacrifical trolley dillemma questioning whether a participant would pull a lever to divert an oncoming trolley to collide into a single person, resulting in their death, in lieu of allowing the trolley to kill 5 other people. Railton derives two subvariants of the sacrificial Switch scenario, each featuring a specific lethal action that results in the death of one person to'save' five other people in the Switch and Beckon contexts.
Despite the similarities between each scenario, Railton encountered asymmetrical results in the distribution of students who answered 'yes' or 'no' to performing the lethal action depending on whether the the act was committed through pulling a lever, waving, or making a beckoning motion.
| Scenario | n | 'Yes' to Lethal Action | 'No' to Lethal Action |
|---|---|---|---|
| Switch | 45 | 85% | 13% |
| Wave | 38 | 87% | 13% |
| Beckon | 38 | 42% | 58% |
Percentages may fail to add to 100%, because students may accidentally push a button that does not correspond to a valid answer. In the text I will give student percentages from the most recent time the course was taught.
We only see this discrepancy in the 'Switch' scenario, to account for this, we will disregard the erroneous person and recalculate the percentages.
2% of 45
85% of 45
Railton theorizes that the asymmetric results of Switch, Beckon and Wave trolley cases are indicative of how each scenario impacts the process of model-free learning that ultimately impact the probability of committing lethal act. In terms of outcome, each of the three circumstances features the same consequence of 1 death for 5 lives; however, the specific hand action used within each scenario invokes different 'cached' expectation values that trigger distinct and intuitive moral judgements.
In Switch, where pulling a lever leads to diverting the trolley, the absence of direct or indirect negative experiences with a physical lever leads to exclusion of a model-free intuitive evaluation of the situation; instead of relying on 'cached' expecttion values, the individual resonably traces the consequences and conciously evaluates models such as Utilitarianism. This results in a majority approval of the action in Switch.
Conversely, in Beckon, where beckoning a person leads to their death, the model-based learning process evaluates the act with a negative 'cached' expectation value. The physical act of gesturing to beckon implies that the person possesses higher intentionality and purpose in luring them to their death or perceiving the victim as a means to an end, leading to disapproval from the majority.
In Wave, where waving signals safety to others, the model-free learning process, qualifies as a cached positive valuatio and influences the majority to approve of the action. These results demonstrate how the degree of model-free learning processes influences moral judgments in different scenarios.
A runaway trolley is speeding down the track, its driver slumped over the controls, apparently unconscious. Ahead on the tracks are five workers, who do not see the trolley coming, and who soon will be struck and killed. You are standing next to a lever that operates a switch lying between the trolley and the workers. Pushing this lever would send the trolley onto a sidetrack. That would save the five workers, but there is a single worker on the sidetrack, who will be struck and killed. Should you push the lever to send the trolley down the sidetrack?
A runaway trolley is speeding down the track, its driver slumped over the controls, apparently unconscious. Ahead on the tracks are five workers, who do not see the trolley coming, and who soon will be struck and killed. You are standing at some distance from the track, with no ability to turn the train or warn the men. A large man, whose weight is sufficient to stop the trolley,is standing on the other side of the track, facing in your direction. He is unable to see the oncoming trolley owing to a traffic signal box that blocks his view up the track. If you would beckon to him [I pantomime an encouraging beckoning gesture], he would step forward onto the track, and be immediately struck and killed. This would halt the trolley and save the five workers. Should you beckon to the large man?
A runaway trolley is speeding down the track, its driver slumped over the controls, apparently unconscious. Ahead on the tracks are five workers, who do not see the trolley coming, and who soon will be struck and killed. A wall prevents them from moving to their left to avoid the trolley, but there is space to their right. You are standing at some distance from the track, with no ability to turn the train. The workers are facing in your direction, and if you were to wave to their right with your arms [I pantomime an encouraging waving gesture], the five workers on the track would step off and escape injury. However, a single worker who is closer to you and standing to the left of the track, and who also does not see the trolley, will see you wave, and he will step onto the track, and immediately be hit and killed. Should you wave to the workers?
In our study, we explored Railton's moral learning framework as applied to variations of the Trolley Problem, using Bayesian causal analysis to investigate the intricate interactions between model-free learning and socio-cultural influences. We constructed two distinct models: the "Good Case," which aligns with Railton's original hypotheses, and the "Bad Case," which introduces alternative explanations that include unaccounted noise and questions the significance of model-free variables. Through these models, we rigorously examined the robustness of the conclusions drawn in the original study.
- Good Case Analysis:
Our good case model, which closely aligned with Railton’s interpretations, showed that different actions (beckon, switch, wave) significantly influence moral decisions based on the nature of the model-free learning associated with each scenario. This was evidenced by the positive and negative influences captured in the logistic regression outputs, which confirmed Railton's notion that the physical act associated with each scenario triggers varying intuitive moral judgments.
- Bad Case Analysis:
The bad case model challenged the simplicity of Railton’s model by introducing noise as a significant factor potentially affecting moral judgments. This model suggested that external variables such as cultural influences and sampling variability could substantially skew the outcomes, thus questioning the generalizability of Railton’s findings across diverse populations. By modeling these factors explicitly, it was demonstrated that when accounting for broader socio-cultural noise, the predictive strength of the model-free learning variable on moral decisions appears less deterministic and more influenced by externalities than initially proposed.
Upon comparing the direct effects model (no mediator) with the total effects model (with mediator), the direct effects model generally exhibited more precise and reliable parameter estimates. This enhanced precision suggests that simplifying the model to focus on direct scenario impacts without the mediation of model-free learning could yield more robust and interpretable results in certain contexts.
When assessing the "Good Case" versus the "Bad Case," the direct effects model not only provided clearer insights but also highlighted the potential overestimation of the role of model-free learning in Railton's original setup. The good case validated some of Railton’s claims under controlled assumptions, while the bad case presented a plausible alternative narrative that introduces a critical perspective on the influence of unmodeled factors.
This analysis underscores the importance of considering external and potentially confounding variables in studies pertaining to the cognitive sciences. The findings advocate for a nuanced approach to interpreting cognitive experiments, especially those involving complex moral and ethical decision-making. Future research could benefit from integrating more comprehensive models that account for socio-cultural dynamics and other contextual factors that might influence moral judgments.
Moreover, this project highlights the utility of Bayesian methods in providing a flexible and robust framework for testing both conventional and alternative hypotheses in cognitive science. By employing causal graphs and Bayesian modeling, researchers can not only replicate and test existing theories but also explore new dimensions of psychological phenomena that traditional methods might overlook.
In conclusion, while Railton's model offers valuable insights into the mechanics of moral decision-making, our analysis suggests that the robustness of these insights may be contingent upon the broader context in which decisions are made. The exploration through Bayesian causal critique thus not only reinforces the findings under specific conditions but also opens avenues for refining our understanding of moral cognition in more diverse and realistic settings.