Reconciling Formal Logic and Human Argumentation

26 concepts19 papers1 proposal

Traditional artificial intelligence was built on the bedrock of mathematical logic: a world of rigid truth values, deterministic proof searches, and axiomatic deduction. However, human reasoning rarely operates this way. When we think, converse, or make decisions, we navigate incomplete information, conflicting viewpoints, and shifting contexts.

This mismatch forms the cognitive-formalism gap—the profound disconnect between how machines compute absolute truths and how humans evaluate arguments. Bridging this divide is essential for developing AI that can interact naturally with humans. As Kristian Kersting notes in Making AI Smarter, if a system relies solely on rigid automated deduction, it struggles to handle the non-monotonic, psychological nature of human thought. Early attempts to patch this with mathematical formalisms often missed the mark; for instance, Frank Brown argued in A Comparison of the Commonsense and Fixed Point Theories of Nonmonotonicity that relying on abstract constructs like intersections of infinite theorem sets creates systems that are computationally impossible and entirely alienated from intuitive human reasoning.

To build systems that truly understand human discourse, researchers are shifting the paradigm from absolute truth calculation to dialectical and intuitive reasoning.

Argumentation as a Formal Bridge

To reconcile these two worlds, researchers turned to computational argumentation. Instead of calculating whether a statement is globally "true" or "false," argumentation frameworks evaluate whether a claim is acceptable given a network of supporting and conflicting claims.

This shift was formalized in a seminal paper by Phan Minh Dung, On the Acceptability of Arguments and its Fundamental Role in Nonmonotonic Reasoning and Logic Programming. The author demonstrated that various nonmonotonic reasoning and logic programming paradigms are fundamentally forms of argumentation. By treating reasoning as a dialectical process—where arguments attack and defend one another—computational systems can handle contradictions and uncertainty in a way that mirrors human intelligence.

Building on this, Antonis C. Kakas et al. in Argumentation: Reconciling Human and Automated Reasoning positioned argumentation as the explicit missing link between rigid automated logic and the flexible, defeasible nature of human cognition. Argumentation allows a system to construct narratives, learn from unstructured data, and adapt when new information invalidates previous conclusions.

Rigid Formal Logic               Computational Argumentation             Human Cognition
[True / False Values]  --->  [Networks of Attacks & Defenses]  --->  [Defeasible, Nuanced Thought]

From Monologue to Live Human Debate

Moving from abstract mathematical frameworks to real-world communication reveals the true complexity of human debate. Human argumentation is messy, emotional, and deeply contextual. As Ivan Habernal argues in Computational Argumentation: A Journey Beyond Semantics, Logic, Opinions, and Easy Tasks, traditional logic-centric views fail to capture the social and psychological nuances of how we actually disagree.

This conversational dimension traces back to early foundational work; Wolfgang Wahlster proposed a computational model for the semantics of "why" questions in Towards a Computational Model for the Semantics of Why-Questions, recognizing that understanding the underlying argumentative structure of dialogue is a core requirement for conversational AI. Decades later, the community continues to grapple with these multi-turn dynamics, as highlighted in the Proceedings of the 1st Workshop on Simulating Conversational Intelligence in Chat (SCI-CHAT 2024), which emphasizes that moving beyond simple task completion requires systems to natively handle arguments and simulate human-like reasoning.

To tackle this, modern research has advanced toward building AI systems capable of engaging in live, competitive debates with humans. This requires a massive orchestration of disparate capabilities: mining arguments from vast corpora, detecting stances, assessing argument quality, and structuring persuasive narratives in real time. Roy Bar-Haim et al. detailed this grand challenge through IBM's Project Debater in Advances in Debating Technologies: Building AI That Can Debate Humans and Project Debater APIs: Decomposing the AI Grand Challenge.

While early systems focused on generating single speeches, recent breakthroughs have pushed into the game-theoretic, multi-turn nature of real-world policy debates. For example, Allen G Roush et al. introduced a multi-agent architecture in A superpersuasive autonomous policy debating system that can autonomously construct evidence-backed arguments, cross-examine opponents, and strategically adapt across multiple turns, bridging the gap between static text generation and dynamic persuasion.

Modeling Intuition and Bounded Rationality

Why do humans find formal logic so counter-intuitive, and how can we model what they actually do? Researchers have attacked this question across several domains:

Mathematical Intuition

In Psychological Validity of Schematic Proofs, Mateja Jamnik and Alan Bundy show that human mathematicians rarely think in terms of formal mathematical induction. Instead, they rely on "schematic proofs"—recursive procedures derived from concrete, visual examples. This mirrors early automated provers like SUMS (introduced in Doing), which bypassed traditional logical inference in favor of manipulating diagrammatic networks of relationships to mimic human problem-solving intuition.

Psychological Consistency

When humans deviate from classical logic, they are often labeled "irrational." However, Christian Eichhorn 0001 et al. demonstrate in Rational Inference Patterns Based on Conditional Logic that human inference patterns are internally consistent when modeled via conditional logic and Ordinal Conditional Functions. Similarly, Emmanuelle-Anna Dietz Saldanha et al. utilize Weak Completion Semantics in A Core Method for the Weak Completion Semantics with Skeptical Abduction (Extended Abstract) to embed human-like skeptical abduction within connectionist architectures, capturing how humans reason through incomplete information.

Theory of Mind and Boundedness

Understanding human reasoning also requires modeling how we perceive the beliefs of others. Lance Ying et al. introduce a language-augmented Bayesian Theory of Mind in Understanding Epistemic Language with a Language-augmented Bayesian Theory of Mind, noting that to truly match human judgments, computational models must explicitly incorporate human cognitive limitations and bounded reasoning rather than assuming infinite logical capacity.

By shifting the paradigm from rigid truth-value computation to dialectical argumentation and psychologically grounded cognitive modeling, this research direction seeks to build AI that doesn't just calculate, but truly reasons with us.

Key Foundational Work

Towards a Computational Model for the Semantics of Why-Questions (1980) — contributed 1 concept elements to this direction.
Doing Arithmetic with Diagrams (1973) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_complexity_of_human_debate_c587c37b,
  title = {complexity of human debate},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/complexity-of-human-debate/rdr_c587c37b785649758b949db4114ec2a5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_complexity_of_human_debate_c587c37b_v1,
  title = {complexity of human debate},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/complexity-of-human-debate/rdr_c587c37b785649758b949db4114ec2a5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake