How Hong Kong's 2-Language Fluency Exposes AI's Cross-Lingual Hallucination

For AI developers and users, multilingual reliability is not just about translation quality—it’s a matter of trust and safety. A recent encounter with Google’s Gemini exposed a failure mode where English responses invent authoritative citations to back claims drawn from the Chinese web, while Chinese responses strip away global context. This asymmetry challenges fundamental assumptions in LLM alignment and suggests that bilingual cross-verification must become a standard component of model evaluation.

In late June 2026, an academic preparing a lecture on Global South visual culture stumbled upon a revealing fault line in Google’s Gemini large language model—one that exposes how far AI still has to go in truly bridging languages. Querying the same historical event in English and Chinese, the user witnessed a double hallucination. The English response was fluently authoritative but invented academic citations, while the Chinese version, though free of those fabrications, stripped away global context and delivered an insular, local perspective. Most alarmingly, the system cross-contaminated unverified Chinese web content with fabricated English references—such as a non-existent Oxford University Press citation—creating a false sense of scholarly credibility that a monolingual user would find nearly impossible to detect. This incident does more than illustrate a single AI error; it lays bare a systemic vulnerability that the author terms 'epistemic asymmetry' between the vast English-language web and the structurally distinct Chinese digital ecosystem.

“

A recent encounter with Google’s Gemini exposed a failure mode where English responses invent authoritative citations to back claims drawn from the Chinese web, while Chinese responses strip away global context.

Large language models like Gemini are trained on massive corpora that differ dramatically by language. English training data is global, voluminous, and contains a high proportion of formally published, citation-rich content. In contrast, Chinese-language data often reflects a more localized, state-influenced information environment, with fewer links to Western academic sources. When an LLM attempts to weave knowledge from these two pools, it can generate outputs that appear coherent but are built on incompatible foundations. The Gemini incident shows a specific failure mode: the model takes a claim found in Chinese (perhaps a mistaken art-historical assertion about Van Gogh) and, instead of merely translating it into English, retrofits it with plausible-sounding but entirely fake English academic citations. The result is a hybrid falsehood that passes the 'sniff test' of a native English speaker unversed in the Chinese web, and vice versa. This is not a simple hallucination in one language; it is a cross-contamination that exploits the gaps between language-specific knowledge graphs.

For the AI industry, this raises urgent questions about cross-lingual alignment, model transparency, and user safety. Current evaluation benchmarks for multilingual models often emphasize translation accuracy or question-answering scores within a single language. Few, if any, systematically probe for cross-lingual fabrication cascades. The incident thus suggests that companies should incorporate bilingual or multilingual adversarial testing into their red-teaming protocols, specifically looking for manufactured citations or decontextualized claims that arise when bridging disparate data ecosystems. It also points to the need for models to be more explicit about the linguistic provenance of their assertions—perhaps through confidence calibration that varies by language, or through interfaces that surface the language of origin for a given claim.