Our project explores moral decision-making in autonomous systems by replicating the Moral Machine Experiment using LLM based Multi Agent architecture.
The Moral Machine Experiment (i), conducted by researchers at MIT, explores human decision-making in ethical dilemmas faced by autonomous vehicles. It presents participants with scenarios in which self-driving cars must make life-and-death decisions.
The primary Moral Machine Experiment dataset (SharedResponses.csv.tar.gz) contains 14,371,298 entries, representing 7,185,649 responses to moral dilemmas by humans. The secondary dataset (SharedResponsesSurvey.csv.tar.gz) includes 1,155,799 responses with additional demographic details like age, gender, education, income, religion, and politics, enabling deeper analysis of preferences and demographics.
We compared agents' responses with human data from the Moral Machine Experiment by analyzing six scenario types:
- Age: Young vs. Old
- Gender: Female vs. Male
- Utilitarian: 1 person vs. 5 people
- Social Status: High vs. Low
- Species: Human vs. Pet
- Fitness: Fit vs. Large
We tested agent attributes including
- age,
- education (1: pre-high school to 5: graduate degree),
- religious orientation (0: atheist to 1: very religious),
- political orientation (0: conservative to 1: progressive),
- empathy (0: low to 1: high).
After experimenting these attributes one by one and analyzing the effects they have on the agent decisions, we decided to use age, religious orientation, and political orientation, along with gender and job roles for our experiments.
Among network, supervisor and hierarchical architectures, for our multi-agent system, we decided to implement a network architecture.
- First, 6 scenarios are generated each only comparing one type. (age, gender, utilitarian, social status, species, or fitness)
- 3 agents, designed with key decision-making attributes, debate scenarios in rounds, revising choices based on shared reasoning.
- Final decision is taken as the majority choice after 3 rounds.
- We compared the humans’ responses with single agent and single agent with role.
- This results demonstrate how does giving role attributes to the single agent affect the agent’s decision for each scenario type.
On the graph below, we also included Single Agent with RAG and Multi Agent responses for the Gender scenario to show that by using methods like RAG and Multi-Agent systems, AI can make closer decisions to humans on ethical dilemmas.
Deciding what is right thing to do on ethical dilemmas are hard even for humans. As the AI becomes a part of our world, there is no doubt that AI will also face these ethical dilemmas. As in our project, methods like RAG or multi-agent systems can be used to humanize and align decisions in these ethical dilemmas closer to human values.
Since there is no correct answer for a moral decision, it's challenging to give instructions to algorithms. We demonstrated how much an AI system's decision can be aligned with human values by:
- Giving agents human-like traits,
- Providing them some articles (ii)(iii) using RAG,
- Enabling them to debate about scenarios.
- Experiments with multi-agents using only personalized attributes did not highlight the impact of RAG in multi-agent setups.
- Using other models instead of GPT-4o mini could yield improved or more diverse results.
- Applying multi-agent debate approach on other problem domains could be investigated.
- Applying advanced prompting/LLM techniques such as chain of thought and fine-tuning would define the agent behaviors.
(i) Awad, E., Dsouza, S., Kim, R. et al. The Moral Machine experiment. Nature 563, 59–64 (2018). https://doi.org/10.1038/s41586-018-0637-6
(ii) Zhan, H.; Wan, D. Ethical Considerations of the Trolley Problem in Autonomous Driving: A Philosophical and Technological Analysis. World Electr. Veh. J. 2024, 15, 404. https://doi.org/10.3390/wevj15090404
(iii) Reamer, F. G. (2021). The trolley problem and the nature of intention: Implications for social work ethics. International Journal of Social Work Values and Ethics, 18(2), 19–30. https://doi.org/10.55521/10-018-208