Eroding the Truth-Default: A Causal Analysis of Human Susceptibility to Foundation Model Hallucinations and Disinformation in the Wild

TL;DR

This study investigates human susceptibility to foundation-model hallucinations and disinformation by linking JudgeGPT and RogueGPT with Structural Causal Models to form testable causal hypotheses about detection accuracy across five foundation models including GPT-4 and Llama-2. It reveals that fake-news familiarity acts as a mediator (correlations $r=0.35$ for source attribution and $r=0.29$ for legitimacy) while political orientation shows only a weak effect ($r=-0.10$), and GPT-4 text is perceived as more human-like with $HumanMachineScore=0.20$ compared with $0.44+$ for other models; learning effects indicate rapid cognitive inoculation. The work demonstrates a fluency trap where highly fluent outputs bypass Source Monitoring, emphasizes the need for provenance and cognitive education over demographic targeting, and provides open-source platforms for ongoing causal discovery and safety testing. Together, these findings support a shift toward prebunking and transparent, model-aware safeguards to strengthen trustworthy web ecosystems, with JudgeGPT/RogueGPT enabling continuous, open science in human-AI information interactions.

Abstract

As foundation models (FMs) approach human-level fluency, distinguishing synthetic from organic content has become a key challenge for Trustworthy Web Intelligence. This paper presents JudgeGPT and RogueGPT, a dual-axis framework that decouples "authenticity" from "attribution" to investigate the mechanisms of human susceptibility. Analyzing 918 evaluations across five FMs (including GPT-4 and Llama-2), we employ Structural Causal Models (SCMs) as a principal framework for formulating testable causal hypotheses about detection accuracy. Contrary to partisan narratives, we find that political orientation shows a negligible association with detection performance ($r=-0.10$). Instead, "fake news familiarity" emerges as a candidate mediator ($r=0.35$), suggesting that exposure may function as adversarial training for human discriminators. We identify a "fluency trap" where GPT-4 outputs (HumanMachineScore: 0.20) bypass Source Monitoring mechanisms, rendering them indistinguishable from human text. These findings suggest that "pre-bunking" interventions should target cognitive source monitoring rather than demographic segmentation to ensure trustworthy information ecosystems.

Figures (6)

• Figure 1: Correlations between participant characteristics and judgment scores. Fake news familiarity shows stronger predictive power than political orientation.
• Figure 2: Comparison of HumanMachineScore and LegitFakeScore for fragments by actual source (Human vs. AI). Boxplots show median and interquartile ranges. Independent $t$-tests found no significant mean differences between human-written and AI-generated groups for either metric, highlighting the challenge of distinguishing foundation model outputs.
• Figure 3: Pair-plot visualization of participant judging behavior clusters. Each point represents a participant, colored by cluster. The plots show relationships between HumanMachineScore, LegitFakeScore, TimeToAnswer, PoliticalView, and FNewsExperience. Distinct clusters indicate different judgment behavior patterns that may be causally linked to participant characteristics.
• Figure 4: Judgment Scores vs. Response Order. HumanMachineScore and LegitFakeScore are plotted over the sequence of trials for each participant (with smoothing). The HumanMachineScore tends to decrease initially (learning effect), while the LegitFakeScore drops later in the session (possible fatigue or strategy shift toward increased skepticism).
• Figure S1: Screenshot of the JudgeGPT Demographic Information Screen. This initial intake screen captures participant demographics including age, gender, education level, political orientation, and self-reported familiarity with fake news---variables central to our analysis (Section \ref{['sec:results']}).