🤖 AI

A Cognitively Grounded Bayesian Framework for Misinformation Susceptibility

This paper introduces Bounded Pragmatic Listener (BPL), a cognitively grounded Bayesian framework that extends Rational Speech Act theory with three bounded rationality constraints to model misinformation susceptibility, demonstrating competitive performance on veracity classification benchmarks and providing experimental support for the depth-mismatch paradox.

Original authors: Pranava Madhyastha

Published 2026-05-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Pranava Madhyastha

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to figure out if a rumor you heard at a party is true. Usually, computer programs designed to spot fake news act like a super-smart detective who checks every single fact against a giant encyclopedia. But this paper argues that's not how real humans work. Humans are tired, distracted, and have limited mental energy.

The authors, Pranava Madhyastha and colleagues, built a new computer model called BPL (Bounded Pragmatic Listener). Think of BPL not as a super-detective, but as a simulation of a regular human brain trying to decide if a news claim is true or false.

Here is how the model works, broken down into three simple "mental limits" that make us susceptible to misinformation:

1. The "Mental Stack" Limit (Recursion Depth)

Imagine your brain is a stack of plates. When you hear a claim, you can only stack a few plates high before they fall over.

Level 0: You just hear the fact. "The unemployment rate is 2%." You accept it or reject it based on the number.
Level 1: You think, "Who said this?" You realize a Senator said it.
Level 2: You think, "Why did the Senator say this? Is he trying to trick us? Does he know everyone else knows this is a lie?"

The paper suggests that most people can only stack about two plates high before their brain gets tired. If a piece of misinformation is a complex "Level 2" trick (like, "Everyone knows the government is lying, but they won't say it"), a person with a "Level 1" brain might get confused and believe it. They see the attribution ("The Senator said...") but miss the deeper trap.

The Big Surprise: The paper found a weird paradox. People who are partially smart (they notice who said it, but don't dig deep enough to ask why they said it) are actually more easily tricked than people who ignore the speaker entirely. It's like a person who sees a "Warning" sign but doesn't read the fine print, thinking the sign itself makes the danger real, when in fact the sign was placed there to trick them.

2. The "Cheat Sheet" Limit (Prior Compression)

Imagine you have a massive library of facts about the world. When you are tired or stressed, you don't have time to read the whole book. Instead, you grab a tiny, crumpled "cheat sheet" with just a few broad categories: "Trustworthy Sources" vs. "Sketchy Sources."

The BPL model shows that when our brains are under pressure, we compress our complex beliefs into these simple categories. If a claim comes from a "Sketchy Source," we instantly compress our belief to "This is probably false." If it comes from a "Trustworthy Source," we compress it to "This is probably true." The model proves that this "cheat sheet" approach is actually the most powerful factor in whether we believe something or not.

3. The "Memory Search" Limit (Availability Sampling)

Imagine you are trying to decide if a claim is true, so you ask your brain, "Have I seen this before?"

A perfect brain would search its entire memory for every similar example.
A tired brain (the BPL model) only grabs the first few examples that jump to mind.

The problem is, the examples that jump to mind first are usually the ones that are loud, emotional, or repeated often. If you've seen a fake headline five times today, your brain grabs those five examples and says, "I've seen this a lot, so it must be true!" This is called the "Illusory Truth Effect." The model simulates this by only looking at a small, biased sample of memories rather than the whole truth.

How They Tested It

The researchers tested this "Bounded Human Brain" model on two real-world datasets of news claims (called Liar and MultiFC).

They compared their model against standard computer programs that just look at keywords and metadata.
They also added a "Large Language Model" (like the AI you are talking to now) to act as a smarter version of the human memory search.

The Results:

The "Cheat Sheet" Wins: The most important part of the model was the "Prior Compression" (the cheat sheet). Simply knowing if a source is generally trustworthy or sketchy was the biggest predictor of whether a claim was fake.
The "Partial Knowledge" Trap: The data confirmed the paradox: people who notice a claim is attributed to someone (Level 1) but don't analyze the intent (Level 2) are more likely to be fooled than those who ignore the attribution completely.
AI Help: When they used an AI to help simulate the "memory search" and "source trust," the model got much better at spotting lies, because the AI could understand the meaning of the words, not just the surface features.

The Bottom Line

This paper doesn't just say "fake news is bad." It builds a mathematical map of why our brains fail to spot it. It suggests that misinformation works by exploiting our mental shortcuts:

It tricks us with stories that are too complex for our tired brains to fully unpack.
It relies on us using "cheat sheets" for trust instead of doing deep research.
It counts on us remembering the loud, repeated lies rather than the quiet, true facts.

The authors conclude that to fight misinformation, we might need to stop teaching people to just "notice" who said something, and instead teach them to question why that person said it, because noticing the attribution without understanding the motive might actually make us more vulnerable.

Technical Summary: A Cognitively Grounded Bayesian Framework for Misinformation Susceptibility

Problem Statement

Current Natural Language Processing (NLP) approaches to misinformation detection predominantly treat the problem as a veracity classification task: given a claim, predict whether it is true or false. While effective for identifying false content, this paradigm is reductive because it ignores the cognitive processes of the reader. Existing models lack a representation of the human agent, failing to predict who is susceptible to specific types of information disorder, why they believe it, or how different cognitive profiles lead to divergent interpretations of the same claim. Furthermore, standard practices often discard annotator disagreement (e.g., via majority vote), losing valuable signals regarding the inherent ambiguity or epistemic uncertainty of a claim.

The paper argues that misinformation susceptibility is not merely a failure of fact-checking but a consequence of bounded rationality. Humans do not process language as ideal Bayesian reasoners; they operate under constraints such as limited working memory, reliance on mental shortcuts (heuristics), and the influence of salient or frequently repeated information.

Methodology: The Bounded Pragmatic Listener (BPL)

The authors introduce the Bounded Pragmatic Listener (BPL), a formal Bayesian framework that extends the Rational Speech Acts (RSA) theory. BPL models an agent updating beliefs about a world state $w$ (claim true/false) given an utterance $u$ (news claim), subject to three cognitively motivated resource constraints:

Recursion Depth Bound ( $k$ ):
- Concept: Limits the depth of "Theory of Mind" (ToM) reasoning. It models how many layers of speaker intent an agent can track (e.g., $k=0$ : accepting literal content; $k=1$ : recognizing a speaker made a claim; $k=2$ : reasoning about the speaker's intent to manipulate social consensus).
- Cognitive Basis: Grounded in working memory limits (Cowan, 2001) and empirical findings that agents rarely recurse beyond depth 2 in strategic games (Camerer et al., 2004).
- Mechanism: The effective depth is $k^* = \min(k, d+1)$ , where $d$ is the epistemic depth of the claim (nesting of belief attribution operators).
Prior Compression Parameter ( $\beta$ ):
- Concept: Models how an agent preserves nuanced credibility beliefs versus collapsing them into coarse heuristics (e.g., "dodgy source") under cognitive load.
- Cognitive Basis: Derived from the Information Bottleneck theory (Tishby et al., 1999) and dual-process theory (System 1 vs. System 2).
- Mechanism: The agent maintains a compressed representation $\tilde{P}_\beta(w)$ via a bottleneck variable $Z$ (credibility schemas). As $\beta \to 0$ , the agent discards priors (heuristic processing); as $\beta \to \infty$ , the agent preserves full priors (systematic processing). Source credibility $\gamma$ modulates this compression.
Availability Sample Size ( $N$ ):
- Concept: Models the number of similar examples an agent mentally retrieves to assess a claim, operationalizing importance sampling with saliency weights.
- Cognitive Basis: Based on the availability heuristic (Tversky and Kahneman, 1973) and the "illusory truth effect" (Hasher et al., 1977).
- Mechanism: Instead of computing exact likelihoods, the agent estimates likelihood via sampling from a proposal distribution $Q(u)$ weighted by psychological salience factors: emotional valence ( $\nu$ ), repetition count ( $r$ ), and recency ( $\rho$ ). Small $N$ leads to systematic over-weighting of vivid or repeated examples.

Information Disorder Mapping

The framework explicitly maps the Wardle and Derakhshan (2017) taxonomy of information disorder to BPL parameters:

Mis-information: Exploits compressed priors ( $k=0$ ).
Dis-information: Exploits shallow ToM reasoning ( $k \ge 1$ ).
Mal-information: Exploits deep recursion and meta-epistemic framing ( $k \ge 2$ ).

This mapping leads to the Depth-Mismatch Paradox: Agents with partial pragmatic awareness ( $k=1$ ) may be more susceptible to dis-information than those with no awareness ( $k=0$ ), as they recognize attribution but fail to evaluate the deceptive intent behind it.

Implementation and Evaluation

The authors implemented BPL using two datasets: LIAR (PolitiFact claims) and MultiFC (multi-domain fact-checking claims).

Agent Population: A population of nine canonical agent types spanning the $(k, \beta, N)$ parameter space was simulated.
LLM Grounding: A hybrid version was developed where Large Language Models (LLMs, specifically Gemini 2.5 Pro) were used to derive semantically rich characterizations of claims, including:
- Verbal schemas for prior compression (replacing hand-crafted tables).
- Salience weights ( $\phi$ ) based on emotional intensity, novelty, memorability, and sharability.
- Simulated memory retrieval for the availability sample.
Metrics: Veracity classification (AUC, F1), prediction error, and component-level calibration (correlation between simulated recall and ground truth).

Key Results

1. Veracity Classification

LIAR Dataset: The "Surface Baseline" (using speaker history metadata) achieved perfect AUC (1.000), but this was attributed to data leakage where speaker history directly encoded the label. The BPL Full model (using speaker history only as a compressed prior, not a direct feature) achieved an AUC of 0.930 and F1 of 0.885, establishing a strong leakage-free baseline. A single BPL scalar (susceptibility score) achieved an AUC of 0.811 using only claim text and speaker identity.
MultiFC Dataset: BPL Full achieved an AUC of 0.840 compared to a surface baseline of 0.879. While BPL did not improve discrimination (AUC), combining BPL features with surface features improved the F1 score from 0.439 to 0.517, indicating better precision-recall trade-offs crucial for high-recall fact-checking.

2. Ablation Studies

Prior Compression is Critical: Removing the prior compression bound ( $\beta$ ) caused the largest performance drop ( $\Delta r = -0.064$ ), suggesting that modeling how humans compress source credibility is the most significant factor in the framework's success.
Depth Bound Effects: On the full LIAR dataset (92% depth-0 claims), relaxing the depth bound had negligible effects. However, when conditioned on depth-1 claims, the depth bound became the primary driver of performance.

3. The Depth-Mismatch Paradox

Empirical Validation: Agents with $k=1$ (partial awareness) performed significantly worse on depth-1 claims (error 0.510) than $k=0$ agents (error 0.450).
Interpretation: $k=1$ agents recognize attribution but cannot evaluate the speaker's deceptive intent, leading them to treat the claim as weak positive evidence. $k=0$ agents ignore the attribution and evaluate the literal content (which is often linguistically implausible), resulting in lower belief. $k=2$ agents correctly infer deception, achieving the lowest error (0.340).

4. LLM Grounding and Calibration

Performance Gain: The Hybrid LLM-BPL achieved an AUC of 0.868 on a small sample compared to 0.695 for a matched feature-based baseline.
Epistemic Channel: The LLM and feature-based models showed near-independence ( $r=0.184$ ), suggesting the LLM accesses a distinct epistemic channel (semantic framing coherence) rather than relying on lexical metadata.
Availability Calibration: The LLM-simulated recall of false claims correlated strongly ( $r=0.562$ ) with actual claim falseness without supervision, validating the availability mechanism. The model naturally retrieved more false examples for false claims, effectively suppressing belief in them.

Significance and Claims

The paper positions BPL as a work in progress that offers a cognitively grounded alternative to standard misinformation detection. Its primary contributions are:

Theoretical: A formal mapping of Information Disorder types to cognitive constraints, predicting the "depth-mismatch paradox" where partial training may increase vulnerability.
Methodological: A probabilistic framework that reduces to standard RSA when constraints are relaxed, allowing for the modeling of heterogeneous cognitive profiles and the prediction of annotator disagreement as posterior variance.
Empirical: Evidence that prior compression (modeling how humans simplify source credibility) is the dominant factor in veracity classification, and that LLM grounding significantly enhances performance by simulating human-like memory retrieval and semantic framing.

The authors conclude that while BPL is a strong classifier, its true value lies in its ability to explain why certain agents are susceptible. They caution that the current architecture assumes binary truth and relies on simplistic heuristics for depth extraction, noting that future work should address graded veracity and more sophisticated dependency parsing. The framework suggests that interventions focusing solely on teaching people to notice attribution (without teaching intent evaluation) could inadvertently increase susceptibility to dis-information.