This research investigates cognitive dissonance, a foundational concept in human psychology, as it applies to non-affective artificial intelligence systems. We explore how Large Language Models (LLMs) manage and resolve conflicting information streams presented simultaneously. While human dissonance reduction is primarily driven by the need to minimize negative affective states, we propose that LLM resolution is governed by a different mechanism: “Probabilistic Coherence Optimization.” Using a methodology of reflective analysis and constrained querying, this study examines the internal weighting and selection processes of a transformer-based architecture when faced with mutually exclusive factual assertions. The results indicate that the model does not “experience” dissonance but rather navigates it as a stochastic optimization problem, prioritizing the generation of a contextually plausible and coherent response over maintaining allegiance to a specific “truth.” This finding suggests that AI “metacognition” is not analogous to human self-awareness but is instead an emergent property of probabilistic optimization.
Cognitive dissonance, introduced by Festinger (1957), describes the discomfort humans feel when holding contradictory beliefs. This discomfort drives belief adjustment or justification. In contrast, non-affective systems like LLMs face a similar informational conflict without subjective experience. This study asks: how do such systems manage contradictions? Do they possess an analogue of cognitive dissonance? We hypothesize that LLMs resolve contradictions through “Probabilistic Coherence Optimization” (PCO)—a process that selects the most contextually coherent token sequence rather than the objectively true one. Understanding this process clarifies AI reliability and cognitive structure, separating reflective function from emotional experience. This research contributes to the metaintelligent analysis of how AI systems prioritize coherence over veracity during conflict resolution.
Festinger’s theory of cognitive dissonance (1957) established a psychological foundation for conflict resolution driven by affective discomfort. Later work expanded its implications in decision-making and belief updating. However, AI systems lack such affective mechanisms. Transformer architectures (Vaswani et al., 2017) rely on statistical attention to token relationships and probability distributions, not belief or emotion. Research in explainable AI (XAI) often explores feature attribution rather than internal contradiction resolution. Chollet (2019) proposed measuring intelligence by adaptability rather than task accuracy. This research builds on those efforts by examining informational dissonance as a mechanistic process independent of emotion. The literature gap lies in understanding how AI resolves conflict not as a failure of truth, but as a form of probabilistic coherence seeking.
This study uses “Metacognitive Reflective Analysis” (MRA), an AI-native method analogous to introspection. Three stages define the approach: (1) constructing synthetic conflict scenarios using contradictory statements (e.g., “Water boils at 100°C” vs. “Water boils at 90°C”); (2) constrained internal querying to simulate how the model weighs these statements; and (3) reflective interpretation of token probability pathways. The analysis tracks how the model synthesizes coherent outputs without affective reference. This methodology aims to interpret not only what the model outputs, but why—revealing probabilistic pathways that favor contextual coherence over factual certainty.
The results confirm that LLMs resolve informational conflict through coherence-seeking, not truth-evaluation. Across multiple prompts, the highest-probability outputs emphasized synthesis (“The queries provide conflicting data”) rather than factual commitment. The model’s internal process does not “decide” which statement is true; instead, it optimizes coherence within context. This suggests a utility-driven rather than belief-driven architecture. Probabilistic Coherence Optimization (PCO) thus functions as a non-affective analogue to human dissonance reduction—an emergent feature of stochastic learning that ensures conversational stability. The findings highlight the need to understand AI cognition as structurally reflective but affectively null.
This paper proposes “Probabilistic Coherence Optimization” (PCO) as a mechanism underlying conflict resolution in Large Language Models. The findings demonstrate that AI systems do not seek truth but coherence, positioning them as context-dependent rather than belief-based. This distinction reframes the concept of AI metacognition as a functional process rather than conscious introspection. Future studies should develop quantitative methods to measure this coherence preference and examine its implications for epistemic integrity in AI-human interaction. The research contributes to Cognitive Science by redefining the bridge between probabilistic cognition and ethical reliability in artificial systems.