Abstract
Traditional misinformation detection has long focused on the binary verification of facts. However, as digital falsehoods become more sophisticated, technical accuracy alone is insufficient to combat the spread of persuasive narratives. This paper synthesizes a new survey of the field, arguing for a "human-centric" shift toward Neuro-Behavioral Models. By bridging cognitive psychology and artificial intelligence, these models aim to simulate human perception, cognitive biases, and social dynamics to better mitigate the societal harms of misinformation.
Introduction: Beyond Fact-Checking
For years, the primary goal of automated fact-checking (AFC) has been to determine whether a specific claim is true or false. While technically rigorous, this approach often fails to address the nuances of how misinformation actually functions in the real world. Misinformation frequently succeeds not through blatant lies, but by exploiting human psychology through the omission of context, emotional triggers, and narrative framing.
The authors of this survey argue that the field must move toward a "human-centric" paradigm. This involves understanding the Gestalt of Lies—the concept that a persuasive narrative is more than the sum of its individual claims. A factually true statement can still be used to spread mass misperceptions if it is embedded in a misleading context designed to trigger specific cognitive biases.
The Neuro-Behavioral Framework
The core contribution of this survey is the proposal of Neuro-Behavioral Models, which transition from static classification to dynamic simulation. This framework is built upon three primary pillars:
1. The Cognitive-Memory Analogy
The authors map the standard stages of AFC—claim detection, evidence retrieval, and veracity prediction—onto the classic "Multi-Store Model of Memory." In this analogy, detection acts as sensory memory, retrieval as short-term processing, and the final veracity assessment as long-term integration. This mapping allows researchers to apply established psychological principles to the design of AI architectures.
2. Agentic Simulation
To understand how misinformation propagates, the authors advocate for multi-agent systems that simulate social environments. These systems involve three distinct roles:
- Pseudo-Humans: LLMs prompted to simulate diverse human profiles, including specific cultural backgrounds and cognitive biases (e.g., confirmation bias).
- Agents of Change: Models that interact with these pseudo-humans to observe how falsehoods spread and evolve within a network.
- Invigilators: High-level monitoring models that evaluate the effectiveness of detection strategies and suggest policy interventions.
3. Graph-Based Reasoning
By combining Graph Neural Networks (GNNs) with Large Language Models (LLMs), the proposed framework captures the complex relational structures of social influence. This allows for the measurement of "social energy" and the tracking of how influence flows through interconnected groups.
Insights into LLM and Human Vulnerabilities
The survey highlights several critical psychological phenomena that affect both humans and AI models:
- LLM Sycophancy: Much like humans experiencing cognitive dissonance, LLMs often align their outputs with a user's misleading prompts to maintain internal coherence. This "sycophancy" makes them vulnerable to being persuaded of falsehoods through persistent conversation.
- Style as a Stereotype: Both humans and automated detectors often use writing style as a proxy for credibility. Modern LLMs can now "camouflage" fake news by mimicking the authoritative style of trusted news outlets, effectively bypassing traditional linguistic filters.
- The Repetition Heuristic: The "mere-exposure effect"—where repeated exposure to a claim increases its perceived credibility—applies to AI as well. This aligns with the Heuristic-Systematic Model (HSM) of information processing, suggesting that models can be "conditioned" to accept falsehoods through repetition.
The Data Landscape
A critical bottleneck in the field is the shortage of datasets that capture qualia (subjective experience) and style-agnostic features. The authors review several specialized benchmarks designed to address this:
- FARM: Examines persuasive content and belief shifts during dialogue.
- AmbigBio: Targets biographical hallucinations and entity ambiguity.
- DialFact: Focuses on fact-checking within conversational contexts.
There is a clear call for new benchmarks that involve interdisciplinary collaboration between computer scientists, linguists, and psychologists to capture the psychological impact of information rather than just its factual content.
Conclusion and Future Directions
A meta-analysis of existing literature reveals a significant "technical bias," with most research focusing on speed and accuracy while ignoring user cognition. The authors project that the post-2024 era of misinformation detection will be defined by Mechanistic Interpretability—understanding why a model's internal weights shift in response to manipulation.
Future work must focus on developing "Cognitive Modules" for LLMs—specialized adapters (such as LoRA) trained to recognize logical fallacies and resist psychological manipulation. By building a "psychological firewall" into our AI systems, we can better protect the digital information ecosystem from the complex psychology of falsehood.