In formal logic, a contradiction is a failure. If an agent believes both $A$ and $\neg A$, traditional systems view this as a state that must be resolved or "repaired" to maintain integrity. However, human reasoning rarely adheres to this strict requirement for consistency. We learn, forget, and shift our logic based on social context or local boundaries.
Research in behavioral argumentation modeling seeks to bridge this "rationality-psychology gap." By moving from prescriptive models (how agents should reason) to descriptive models (how humans actually reason), researchers aim to build AI systems that can effectively tutor, persuade, and collaborate with people.
The Prescriptive-Descriptive Gap
Most formal argumentation frameworks are normative. As Gabbay et al. observe, these frameworks define how an idealized rational agent ought to evaluate conflicting information. While mathematically elegant, these models lack mechanisms to account for the cognitive distortions, logical fallacies, and messy natural language inherent in human discourse.
This disconnect is particularly visible in educational and training settings. Ikeda et al. argue that maintaining a perfectly consistent representation of a student’s mental state is nearly impossible because human learning is nonmonotonic—students learn new facts that contradict old ones, forget rules, or hold fragmented beliefs. Similarly, Murakami et al. note that human behavior in virtual training environments is often driven by conflicting motivations that standard computational models struggle to reconcile.
The Social Utility of Inconsistency
Why are humans so "illogical"? Empirical evidence suggests that what looks like a logical failure is often a social feature. Rosenfeld et al. conducted studies showing that human debaters frequently introduce contradictory arguments within the same discussion. They do this not because they are confused, but to project a sense of neutrality or balance. This behavior directly violates the "conflict-freedom" principle—a cornerstone of formal argumentation theory—suggesting that social signaling often overrides strict logical consistency.
Furthermore, human reasoning is often idiosyncratic and context-dependent. Fan et al. demonstrate that in analogy tasks, humans may shift their underlying logic mid-task, applying different rules at boundaries or specific data points. Traditional models that enforce a single global rule fail to capture these "abnormal" but characteristically human shifts.
Validating Belief Revision
To address these gaps, researchers are beginning to test formal logical postulates against empirical human data. Baker investigates the psychological realism of the AGM framework—the standard theory for how agents should update their knowledge—by comparing its predictions to actual human belief-change patterns. The goal is to determine if these formal axioms are perceived as natural or useful by humans when they encounter conflicting information.
Even when models attempt to incorporate uncertainty, they still require validation. De Bona et al. have expanded epistemic graphs to include conditional probabilities and labeling constraints, bridging the gap between logic and probabilistic reasoning. However, the authors emphasize that it remains unknown whether the logical constraints derived from these graph topologies actually align with how individuals update their degrees of belief during complex debates.
Toward Psychologically Realistic Agents
If an AI assistant provides advice based on a logic the user doesn't share, the interaction fails. Rosenfeld et al. argue that AI debate assistants often provide counter-intuitive suggestions because they prioritize mathematical justification over psychological relevance.
To build better systems, Rosenfeld proposes "implicit interaction" models that predict human actions based on observed behavior rather than assuming logical consistency. Similarly, Hunter explores how automated systems can "read the room" in asymmetric dialogues, strategically selecting arguments that resonate with a user's messy, non-standard belief updates. As Habernal suggests, the future of computational argumentation lies in moving beyond pure semantics and logic to embrace the multifaceted reality of human communication, including common sense and social context.
Key Foundational Work
Modeling Human Behavior for Virtual Training Systems (2005) — contributed 1 concept elements to this direction.
Nonmonotonic Model Inference-A Formalization of Student Modeling (1993) — contributed 1 concept elements to this direction.