TL;DR
This paper introduces a computational model for social causality and responsibility judgment in multi-agent systems, drawing heavily on psychological attribution theory. The model aims to enable AI agents to understand and assign blame or credit for actions in social interactions, mirroring human reasoning, and has shown promising empirical validation against human judgments.
What Problem Does This Paper Address?
Traditional AI often focuses on individual problem-solving and rational behavior, which is insufficient for understanding and participating in human social interactions. This paper tackles the challenge of social causality – how an agent infers the reasons behind social behavior, including mental states and coercive circumstances – and responsibility judgment – how agents assign credit or blame based on these inferences. This is crucial for developing more human-like intelligent agents and systems that can effectively interact with people. The authors highlight that social causality goes beyond simple physical cause-and-effect, incorporating factors like intention and freedom of choice.
Methods
The authors propose a domain-independent computational framework that models social causality and responsibility judgment. Their approach is grounded in psychological attribution theory, specifically drawing from Shaver's and Weiner's models. Key aspects of their methodology include:
- Hierarchical Plan Representation: Causal knowledge is encoded using a hierarchical plan representation, similar to STRIPS operators, which allows for modeling actions, their preconditions, effects, and decomposition into sub-actions. This representation also captures performer and authority relationships, modeling power dynamics.
- Dialogue Inference: Communicative events (speech acts) are used to infer beliefs about agents' epistemic states (knowledge, desires, intentions).
- Causal Inference: This process uses the plan representation and observed actions to infer agency, intention, and coercion.
- Attribution Algorithm: A multi-step algorithm combines dialogue and causal inferences to determine primary and secondary responsible agents and assign a degree of responsibility (high, medium, low).
- Belief-Desire-Intention (BDI) Model: Concepts from BDI are used to map psychological terms to AI concepts.
- Commonsense Reasoning: The inference engine relies on commonsense rules to derive attributions.
Datasets
The paper describes two main experimental studies:
- Company Program Scenarios: Three variants of a "company program" scenario were used, along with questionnaires to assess human judgments of beliefs and evidence used for inference.
- Firing Squad Scenarios: The original "firing squad" scenario and its variants were used to compare the proposed model's predictions against human judgments and other computational models.
Key Insights and Ideas
- Social Causality is More Than Physical Causality: The model explicitly accounts for mental states (intention, foreknowledge) and social factors (coercion, authority) in determining responsibility, moving beyond simple physical cause-and-effect.
- Dialogue is a Rich Source of Social Information: The model leverages speech acts to infer crucial attributional variables, demonstrating the importance of communication in social reasoning.
- Human-like Judgment Requires Multiple Inference Types: Combining dialogue inference, causal inference, and a structured attribution algorithm allows the model to mimic human reasoning patterns in complex social situations.
- Coercion is Nuanced: The model differentiates between different types and degrees of coercion, and how it impacts responsibility assignment, reflecting psychological findings.
- Empirical Validation is Crucial: The authors emphasize the importance of validating their computational model against human data to ensure its psychological plausibility and practical utility.
Key Results
- Good Agreement with Human Judgments: The model's predictions for inferred beliefs and the evidence used for inference showed substantial agreement with human subjects (average Kappa agreement of 0.732).
- Accurate Inference Rule Application: The model's inference rules achieved an average accuracy of 0.85 in predicting how humans use evidence.
- Superior Performance Compared to Alternatives: In comparative studies with simple cause, simple authority, and the Chockler & Halpern (C&H) model, the proposed M&G model generally performed better in matching human judgments across various scenarios, particularly in capturing nuanced social interactions.
- Distinguishing Blame and Responsibility: The model successfully differentiates between physical cause, responsibility, and blame, and can assign different degrees of responsibility.
Limitations and Future Work
- Asymmetry of Credit and Blame: While the model currently uses a uniform approach for credit and blame, the authors acknowledge that psychological findings suggest an asymmetry, with credit assignment criteria being stricter. Future work could incorporate this asymmetry.
- Subjective Biases and Individual Differences: The model does not currently account for perceiver biases, emotional states, or dispositional differences that influence responsibility judgment. Related work from the authors' lab has explored these aspects.
- Delegation and Shared Responsibility: The paper notes that subjects often assign shared blame, and some work exists on modeling delegation. The model's architecture has the potential to incorporate these extensions.
- Negligence and Other Factors: The authors mention that extensions to include models of negligence have been explored.
- Implementation Gaps: Not all causal inference rules (e.g., C22-C25) were implemented in the Java engine used for experiments.
- Computational Complexity: While not explicitly detailed as a limitation, the complexity of modeling communicative acts and their flexible content in approaches like C&H is noted as a challenge.
Summary updated