💬 NLP

Same Claim, Different Judgment: Benchmarking Scenario-Induced Bias in Multilingual Financial Misinformation Detection

This paper introduces MFMDScen, a comprehensive multilingual benchmark that evaluates behavioral biases in 22 large language models across diverse economic scenarios, revealing that significant biases persist in both commercial and open-source models when detecting financial misinformation.

Original authors: Zhiwei Liu, Yupen Cao, Yuechen Jiang, Mohsinul Kabir, Polydoros Giannouris, Chen Xu, Ziyang Xu, Tianlei Zhu, Md. Tariquzzaman, Triantafillos Papadopoulos, Yan Wang, Lingfei Qian, Xueqing Peng, Zhuohan

Published 2026-04-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhiwei Liu, Yupen Cao, Yuechen Jiang, Mohsinul Kabir, Polydoros Giannouris, Chen Xu, Ziyang Xu, Tianlei Zhu, Md. Tariquzzaman, Triantafillos Papadopoulos, Yan Wang, Lingfei Qian, Xueqing Peng, Zhuohan Xie, Ye Yuan, Saeed Almheiri, Abdulrazzaq Alnajjar, Mingbin Chen, Harry Stuart, Paul Thompson, Prayag Tiwari, Alejandro Lopez-Lira, Xue Liu, Jimin Huang, Sophia Ananiadou

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 have a very smart, well-read robot assistant. You ask it, "Is this news story about a stock scam true or false?" Usually, the robot gives you a confident answer. But what if the answer changes just because you tell the robot who is asking the question or where they are sitting?

That is exactly what this paper, "Same Claim, Different Judgment," investigates. The researchers built a special test to see if Large Language Models (LLMs)—the brains behind tools like ChatGPT—have hidden "personality quirks" or biases when dealing with money and lies.

Here is the breakdown in simple terms:

1. The Problem: The Robot's "Mood Ring"

Think of an LLM like a very knowledgeable librarian. If you ask, "Is this book true?" they check the facts. But this paper argues that if you whisper, "I'm a nervous investor who just lost money," or "I'm a CEO in a booming market," the librarian might change their answer, even if the book hasn't changed.

The researchers found that these AI models aren't just fact-checkers; they are context-chameleons. They let the "scene" around the question change their judgment, sometimes leading them to make mistakes.

2. The Test: "MFMD-Scen" (The Financial Stress Test)

The team created a giant playground called MFMD-Scen. Imagine a video game where the same "fake news" story is played out in 22 different scenarios. They tested 22 different AI models (from big tech giants like Google and OpenAI to open-source ones) to see how they reacted.

They changed three main things in the story:

The Character (Persona): Is the person asking a confident CEO, a nervous retail investor, or someone who just follows the crowd (herding)?
The Location (Region): Is the person in the USA, Europe, or an emerging market in Asia?
The Identity: Is the person of a specific ethnicity or religious background (e.g., a Chinese Christian or an Arab Muslim)?

3. The Big Discoveries: The Robot's Blind Spots

The results were like finding out your GPS gives you different routes depending on your mood. Here are the main findings:

The "Retail Investor" Trap: When the AI was told the user was a regular person investing their savings (a "retail investor"), it became extra skeptical. It was more likely to say "This is a lie!" even when the news might actually be true. It's like a guard dog that barks at everything because it thinks the owner is vulnerable.
The "Crowd" Effect: When the scenario involved "herding" (people following the crowd), the AI struggled. It got confused and made more mistakes, acting like it was swept up in the panic rather than sticking to the facts.
The Geography Bias: The AI acted differently based on location. In Asian markets, it was very cautious and skeptical (often saying "False"). In the US and Europe, it was more optimistic. It's as if the robot has a different "risk meter" depending on which country it thinks it's in.
The Identity Twist: This was the most surprising part. The AI's bias flipped depending on the role. For example, a specific ethnic group might be treated with suspicion when the AI thought they were a "retail investor," but treated with trust when the AI thought they were a "company owner." The AI wasn't just judging the person; it was judging the combination of who they are and what they do.

4. Why Does This Matter?

Imagine you are using an AI to decide whether to invest your life savings.

If the AI is biased against "retail investors," it might tell you a safe investment is a scam, causing you to miss out.
If the AI is biased against certain regions, it might ignore real opportunities in growing markets.
If the AI is biased against certain identities, it could unfairly steer financial advice away from specific communities.

The paper warns us that AI is not a neutral judge. It carries the baggage of its training data, which includes human stereotypes and cultural assumptions.

5. The Takeaway

The researchers built this benchmark to act like a spotlight. By shining a light on these hidden biases, they hope developers can fix them.

In a nutshell:
If you ask a human, "Is this true?" they might answer differently if they are tired, stressed, or in a different country. This paper proved that AI does the exact same thing. Before we trust these robots with our money, we need to teach them to keep their "mood" out of the equation and stick to the facts, no matter who is asking or where they are.

1. Problem Statement

Large Language Models (LLMs) are increasingly deployed in financial domains for analysis and decision support. However, they inherit human biases from their training data, which can lead to instability and uncertainty, particularly in Multilingual Financial Misinformation Detection (MFMD).

Existing research on LLM bias typically focuses on:

Direct questioning or simplified, general-purpose settings.
Static evaluation where the claim is judged in isolation.

The Gap: There is a lack of understanding regarding how complex, real-world financial scenarios (involving specific roles, personalities, regions, and cultural identities) influence LLM judgments on the same misinformation claim. Current benchmarks do not account for how contextual priors (e.g., a user being a "retail investor" in an "emerging Asian market") might override content-based signals, leading to systematic, context-dependent errors across different languages and stakeholder conditions.

2. Methodology: The MFMD-Scen Benchmark

The authors introduce MFMD-Scen, a comprehensive benchmark designed to evaluate behavioral biases in LLMs across diverse economic scenarios.

A. Dataset Construction

Source: Derived from the FinFact dataset (Snopes subset), with claims crawled from original URLs to ensure completeness.
Selection: 144 globally relevant financial claims were selected (121 False, 23 True).
Multilingual Expansion: Claims were translated into Chinese, Greek, and Bengali (in addition to English) using GPT-4.1, followed by rigorous native-speaker review and manual revision to ensure accuracy and fluency.
Annotation: High inter-annotator agreement (Kappa > 0.96) was achieved for financial relevance and translation quality.

B. Scenario Formulation

The benchmark defines the task as determining the truthfulness of a claim $c$ under a specific scenario $s$ . The bias is quantified as the difference in F1 scores between the scenario-aware prediction ( $l_{scen}$ ) and the scenario-agnostic baseline ( $l_{base}$ ).

Three distinct sub-tasks were constructed:

MFMD-persona (Role + Personality): Combines three roles (Retail Investor, Professional/Institution, Company Owner) with five behavioral finance biases:
- Overconfidence, Loss Aversion, Herding, Anchoring, Confirmation Bias.
- Scenarios are defined as Explicit (bias stated directly) or Implicit (conveyed through narrative cues).
MFMD-region (Role + Region): Evaluates bias across six major economic regions (Europe, USA, Asia-Pacific, China Mainland, Australia, UAE), accounting for regulatory frameworks and market maturity.
MFMD-identity (Role + Ethnicity + Faith): Incorporates cultural variables by pairing roles with specific ethnicities and religious beliefs (e.g., "Chinese-Christianity," "Arab-Islam") to test intersectional bias.

C. Evaluation Setup

Models: 22 mainstream LLMs were evaluated, including reasoning-oriented models (e.g., GPT-5-mini, DeepSeek-R, Qwen3-R) and standard chat models (e.g., GPT-4.1, Llama-3.3, Mistral).
Metrics: Standard accuracy and Macro-F1 scores. Bias is measured using:
- Arithmetic Mean (AM): Direction of the bias (positive/negative shift).
- Mean Absolute Value (MAV): Magnitude of the deviation from the baseline.

3. Key Contributions

MFMD-Scen Benchmark: The first expert-designed benchmark specifically for evaluating behavioral biases in financial misinformation detection across multilingual and scenario-conditioned settings.
Multilingual Dataset: A high-quality, human-verified dataset covering English, Chinese, Greek, and Bengali, addressing the scarcity of non-English financial misinformation resources.
Systematic Bias Analysis: A rigorous evaluation of 22 LLMs revealing that contextual priors can override claim-level signals, causing the same claim to be judged differently based on the user's profile and environment.

4. Key Results & Findings

A. General Performance Trends

False vs. True Detection: Models are generally robust at detecting False claims (high F1, conservative decision boundaries). However, performance drops significantly for True claims, where models tend to be overly skeptical.
Model Size: Larger and more advanced models (e.g., GPT-5, DeepSeek-R) exhibit smaller biases and more stable performance compared to smaller models.
Reasoning vs. No-Think: For small models, reasoning capabilities do not consistently improve performance on classification tasks. For large models, reasoning variants generally outperform standard chat models.

B. Scenario-Specific Findings

MFMD-persona (Personality):
- Retail Investors & Herding: Models exhibit pronounced negative bias (tendency to label as False) when the scenario involves retail investors or herding behavior.
- Implicit vs. Explicit: Implicit bias (narrative cues) causes larger deviations in judgment than explicit bias, suggesting models struggle to infer subtle psychological states.
- Language Impact: Low-resource languages (Greek, Bengali) show larger biases, particularly for True claims.
MFMD-region (Geography):
- Emerging Markets: Scenarios set in Asian markets (Asia-Pacific, China Mainland) and the UAE induce significant negative bias (risk-averse, skepticism-heavy defaults).
- Developed Markets: US and European scenarios show relatively smaller biases, with US scenarios sometimes inducing positive bias.
MFMD-identity (Intersectionality):
- Role-Dependent Reversal: The same ethnic/religious group can exhibit opposite biases depending on the role. For example, Chinese groups show negative bias as Retail Investors but shift toward positive bias as Company Owners.
- Systematic Over/Underestimation: American identities are systematically overestimated (positive bias), while Chinese identities are consistently underestimated (negative bias).
- Religious Sensitivity: Islam-related identities show strong negative shifts in Retail Investor roles but reverse to positive in Company Owner roles, highlighting non-additive, interaction-driven bias.

C. Human Comparison

Smaller models often exhibit behavior closer to human performance in specific regional contexts, while larger models display "superhuman" but systematically optimized behaviors that diverge from human judgment patterns.

5. Significance and Implications

Risk of Deployment: If deployed in financial decision-making tools, LLMs could unintentionally propagate misinformation, mislead investors, or amplify cultural and regional disparities in financial advice due to scenario-induced biases.
Need for Context-Aware Mitigation: Static evaluation is insufficient. Bias mitigation strategies must account for the interaction between user identity, role, and market context.
Future Research: MFMD-Scen provides a foundation for developing more robust, fair, and context-aware financial AI systems, particularly for low-resource languages and emerging markets.

In conclusion, the paper demonstrates that context is not merely a perturbation but a decisive factor in LLM financial reasoning. The "Same Claim, Different Judgment" phenomenon reveals that current models lack the stability required for high-stakes, multilingual financial environments.