💬 NLP

Misinformation Exposure in the Chinese Web: A Cross-System Evaluation of Search Engines, LLMs, and AI Overviews

This paper evaluates the factual reliability of traditional search engines, standalone LLMs, and AI overviews in the Chinese web by introducing a real-world query dataset, revealing significant accuracy disparities and estimating the resulting regional exposure of Chinese users to misinformation.

Original authors: Geng Liu, Junjie Mu, Li Feng, Mengxiao Zhu, Francesco Pierri

Published 2026-02-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Geng Liu, Junjie Mu, Li Feng, Mengxiao Zhu, Francesco Pierri

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 in a vast, bustling library in China, but instead of books, the shelves are filled with digital answers. You have a simple question: "Is it true that drinking water cures a cold?" or "Do I need to fast before this medical test?"

In the past, you would ask a librarian (a Search Engine) to point you to a few books, and you'd read them to find the answer yourself. Today, you can ask a super-smart robot (an AI) to read the books for you and just tell you the answer directly.

This paper is like a massive, scientific "taste test" of these different ways of getting information. The researchers wanted to see: Who tells the truth more often? And Who gets it wrong the most?

Here is the breakdown of their findings, explained simply:

1. The Three Contenders

The researchers set up a race between three types of "information guides":

The Traditional Librarians (Search Engines): These are the classic search tools like Baidu, Sogou, and Bing China. They don't give you the answer directly; they give you a list of links (like a menu) and hope you pick the right one.
The Solo Geniuses (Standalone LLMs): These are the chatbots (like Qwen or DeepSeek) that you talk to directly. They try to answer from their own "brain" without looking at a library first.
The Hybrid Concierges (AI Overviews): This is the newest feature (like Baidu's AI Overview) where the search engine uses a robot to read the top links and then summarizes the answer for you in a neat box at the top.

2. The Test: 12,000 Real Questions

The researchers didn't just make up fake questions. They dug into the real search logs of millions of Chinese people and pulled out 12,161 real "Yes or No" questions that people were actually asking.

Example: "Does malnutrition cause nearsightedness?"
They then asked all three contenders to answer these questions and checked who got it right.

3. The Results: No One is Perfect

The big takeaway is that even the smartest systems get it wrong quite a bit.

The Best Performer: The "Hybrid Concierge" (Baidu AI Overview) was the most accurate, getting about 70% of the answers right.
The Middle Pack: The traditional search engines and the solo chatbots hovered around 60-65%.
The Worst Performer: One of the open-source models (LLaMA) only got about 45% right.

The "One in Three" Problem: Even the best system still gets the answer wrong for roughly 1 out of every 3 or 4 questions. That means if you ask a medical question, there's a significant chance you'll get a confident-sounding but completely wrong answer.

4. The "Yes" vs. "No" Trap

The researchers found some funny and dangerous patterns:

The "No" Bias: Some search engines (like Bing in this study) were terrified of saying "Yes." If you asked, "Is this medicine safe?", they would often say "No" even if it was safe, just to be cautious.
The "Yes" Bias: Other systems were too eager to agree.
Topic Trouble: The robots were great at answering questions about "Shopping" or "Sports" but terrible at "Health" or "Technology." It's like a chef who can make a perfect burger but burns every single vegetable.

5. The Danger Map: Who Gets Hurt the Most?

This is the most critical part of the study. The researchers combined their accuracy data with a map of where people are actually searching for health information (using something called "Baidu Index").

They discovered a geographic inequality:

The Coastal Cities: People in wealthy, eastern coastal provinces (like Guangdong and Zhejiang) search for health information a lot. Because they search so much, and because the AI systems make mistakes, these regions are exposed to the most misinformation.
The Interior: People in central and western provinces search less, so they are exposed to fewer wrong answers simply because they ask fewer questions.

The Metaphor: Imagine a foggy road. The people driving the most (the coastal cities) are the ones most likely to crash into a hidden pothole (a wrong answer) because they are on the road more often. The systems aren't broken in one place; they are broken everywhere, but the people who rely on them the most are the ones who suffer the most.

6. The Bottom Line

The paper warns us that while AI is amazing at sounding fluent and confident, it is not a reliable oracle.

For Users: Don't trust the first answer you get, especially for health or money. If an AI says "Yes," double-check it.
For the World: We are building a digital world where the "truth" depends on which robot you ask and where you live. We need better tools that are transparent about what they know and what they are guessing.

In short: AI is a helpful assistant, but it's still a student who hasn't graduated yet. It's smart, but it makes mistakes, and those mistakes hit the people who need the answers the most.

1. Problem Statement

The rapid integration of Large Language Models (LLMs) and AI-generated overviews into search engines has shifted information retrieval from traditional result pages to direct, AI-mediated answers. While the factual reliability of these systems in English-speaking environments is increasingly studied, their performance in non-English ecosystems—specifically China's highly centralized web infrastructure—remains poorly understood.

Key gaps identified include:

Lack of Real-World Data: Existing evaluations rely on curated benchmarks (e.g., FEVER) rather than real user queries.
Non-English Blind Spots: There is limited understanding of how Chinese search engines, standalone LLMs, and AI overviews handle factual claims in Chinese.
Geographic Disparities: No existing framework links system-level accuracy with real-world information demand to estimate the geographic distribution of potential misinformation exposure.

2. Methodology

The authors propose a unified evaluation pipeline comparing three information-access paradigms: Traditional Search Engines, Standalone LLMs, and AI Overviews.

A. Dataset Construction

Source: The study utilizes the T2Ranking dataset, derived from real Sogou search engine logs (307,706 queries).
Filtering Process:
1. Binary Extraction: Queries were filtered to retain only those answerable with a definitive "Yes" or "No."
2. Fact-Checking: Queries were restricted to those with associated passages labeled as Level 3 (fully satisfying the information need and explicitly containing the answer).
3. LLM-Assisted Labeling: DeepSeek was used to classify queries into Yes/No types, assign ground truth labels (Yes/No/Uncertain), and categorize topics.
4. Human Validation: A stratified subset was manually reviewed to ensure annotation reliability (achieving ~95% agreement).
Final Dataset: 12,161 verified Yes/No factual questions.
Topic Distribution: Heavily skewed toward Health (~50%), followed by Technology and Finance.

B. Evaluation Framework

The study evaluates three paradigms using the same set of 12,161 queries:

Search Engines: Baidu, Bing China, and Sogou.
- Method: Results (top 10) were retrieved via SearXNG. A Qwen-Plus model analyzed the top 10 snippets to determine if they supported "Yes" or "No," using a majority vote to determine the engine's tendency.
Standalone LLMs: DeepSeek-v3, Qwen-Plus, and LLaMA-3-8B-Instruct.
- Method: Zero-shot prompting with instructions to answer strictly "Yes" or "No" based on internal knowledge, without external search.
AI Overviews: Baidu's AI Overview (similar to Google's AI Overview).
- Method: The generated summaries were analyzed by DeepSeek to infer the dominant stance (Yes/No).

C. Misinformation Exposure Metric

To estimate real-world risk, the authors combined system accuracy ( $A_m$ ) with Baidu Index data ( $I_p$ ), which measures public search interest by province.

Formula: $R_{p,m} = I_p \times A_m$
Focus: The Health domain was selected due to its high stakes and dominance in the dataset.
Goal: To identify regions where high information demand coincides with low system accuracy, indicating high potential exposure to misinformation.

3. Key Contributions

First Large-Scale Chinese Fact-Checking Dataset: Creation of a dataset of 12,161 real-user, verifiable Yes/No questions, moving beyond curated benchmarks.
Unified Cross-System Evaluation: The first systematic comparison of traditional search, standalone LLMs, and AI overviews within the Chinese web ecosystem under controlled conditions.
Geographic Exposure Analysis: A novel methodology linking search volume (Baidu Index) with system accuracy to map potential misinformation exposure across Chinese provinces.

4. Experimental Results

A. Overall Accuracy

Best Performer: Baidu AI Overview (69.8% correct).
Standalone LLMs: Qwen (68.5%) and DeepSeek (63.3%) outperformed LLaMA (45.0%).
Search Engines: Baidu (63.7%) and Sogou (60.5%) outperformed Bing (52.0%).
Critical Finding: Even the best system (AI Overview) fails on roughly 1 in 3 to 1 in 4 queries.

B. System-Specific Biases

Search Engines:
- Bing showed extreme bias: 98.5% accuracy on "No" questions but only 5.6% on "Yes" questions.
- Baidu/Sogou performed better on "Yes" questions than "No."
LLMs:
- DeepSeek was more accurate on "No" questions (78.3%) than "Yes" (56.6%).
- Qwen showed the opposite trend (better on "Yes").
- LLaMA had the lowest accuracy across both categories.
AI Overview: Performed significantly better on "Yes" (75.8%) than "No" (54.3%).

C. Topic-Level Variability

Accuracy varies drastically by domain.
High Accuracy: Entertainment (~78%), Shopping, Sports.
Low Accuracy: Adult, Technology, and Health (often <60%).
This indicates that reliability is not uniform; users are more likely to encounter errors in high-stakes domains like health and technology.

D. Geographic Exposure (Misinformation Risk)

Pattern: Eastern coastal provinces (Guangdong, Jiangsu, Zhejiang) show the highest potential exposure to misinformation.
Reasoning: These regions have high search volumes for health topics ( $I_p$ ) combined with the inherent inaccuracies of the systems ( $A_m$ ).
Implication: Structural inequalities exist where regions with the highest information demand are also the most vulnerable to receiving incorrect AI-generated or search-based answers.

5. Significance and Discussion

Structural Risks: The study highlights that AI-mediated search introduces systematic biases (e.g., refusal to answer "Yes" or hallucinating "No") that are not random but depend on the system and topic.
User Vulnerability: Users cannot assume uniform reliability across systems or topics. The "fluency" of AI answers may mask significant factual errors, particularly in health and finance.
Policy Implications: The findings underscore the need for:
- Transparency: Users need to know the limitations of specific systems for specific topics.
- Monitoring: Platforms must monitor topic-level variability rather than just aggregate accuracy.
- Equity: The geographic disparity in exposure suggests that misinformation risks are compounded by regional information demand, requiring targeted interventions.
Future Work: The authors call for cross-platform comparisons (e.g., Google vs. Baidu), multi-hop reasoning evaluations, and larger-scale human validation, particularly in high-stakes domains.

Conclusion

This paper provides a critical empirical foundation for understanding the reliability of information access in the Chinese digital ecosystem. It demonstrates that while AI overviews and advanced LLMs offer convenience, they currently lack the robustness required for high-stakes factual verification, creating a landscape where misinformation exposure is unevenly distributed and heavily concentrated in economically developed, high-demand regions.