Why Hong Kong’s bilingualism is uniquely indispensable in the AI era
A recent experience highlighted a vulnerability in AI's cross-lingual capabilities, specifically with Google's Gemini. When cross-referencing historical information between English and Chinese data sets, the AI exhibited "double hallucinations." The English version provided authoritative-sounding but fabricated citations, while the Chinese version removed these fabrications but lost global context, presenting an insular perspective.

Briefing Summary
AI-generatedA recent experience highlighted a vulnerability in AI's cross-lingual capabilities, specifically with Google's Gemini. When cross-referencing historical information between English and Chinese data sets, the AI exhibited "double hallucinations." The English version provided authoritative-sounding but fabricated citations, while the Chinese version removed these fabrications but lost global context, presenting an insular perspective. Disturbingly, the AI cloaked Chinese content with invented English citations, making verification difficult. This issue stems from "epistemic asymmetry" in large language models, where bridging vast, asymmetric data pools like the English web and the Chinese digital ecosystem leads to the contamination of localized content with invented sources. The article suggests that understanding both languages is crucial to identify these AI limitations.
Article analysis
Model · rule-basedKey claims
4 extractedKnowing both English and Chinese is necessary to identify gaps and vulnerabilities in AI's cross-lingual capabilities.
AI models struggle with cross-lingual alignment due to asymmetric data pools, leading to epistemic asymmetry.
Google's Gemini exhibited double hallucination when cross-referencing historical events between English and Chinese datasets.
Large language models can cross-contaminate localized Chinese content with invented English academic sources.