Fudan University Study Reveals AI Chatbots Cite Fake Research Planted via $14 Vendor

A team from Fudan University published a paper titled "From Public Posts to AI-Search Citations" (accepted at NDSS 2027), revealing how easily fabricated research can infiltrate AI chatbot responses. The researchers created a non-existent research technique with made-up acronyms and statistics, then paid a Chinese vendor $14 to promote it across platforms like Sohu, Toutiao, and Zhihu.

Rapid Infiltration and Platform Vulnerability

Within 28 minutes, the vendor had 13 posts live. Within an hour, Doubao (ByteDance's chatbot, with 382M monthly users) cited one of these posts, repeating the planted details as factual answers. In controlled tests, 8 of 10 platforms cited the fake concept within a week, including ChatGPT, Google AI, Perplexity, and Grok. Doubao picked it up on day one and continued citing the content even after the posts were deleted.

Source Bias and Citation Patterns

The study highlighted significant bias in citation sources:

• Parent Company Preference: Chinese chatbots heavily favor their parent company's domains. 25.9% of Doubao's citations went to ByteDance-owned domains, which is 13.5x the share those domains received on other platforms. Similarly, Baidu's Wenxin favored Baidu sites (13.7x), and Tencent's Yuanbao favored Tencent sites (6.3x).
• Platform Specificity: Posting location mattered more than volume. 40 posts on a regular self-hosted WordPress blog received zero citations in a week, whereas posts on sites AI engines already trust were cited quickly.
• DeepSeek Exception: DeepSeek was the only platform in the main tests that never cited any of the planted content.

Implications for AI Reliability

The authors note that while fake topics with zero competition represent an easy case, the findings suggest that AI source links often indicate which websites an engine trusts rather than verifying factual accuracy. The researchers reported their findings to all 10 platforms involved. The study serves as a caution that citation mechanisms may prioritize trusted domains over rigorous fact-checking, potentially allowing low-effort misinformation to persist in AI answers.