A large language model known as Ox Alpha exhibits distinct censorship behaviors depending on the language used. When queried in English about the Tiananmen Square massacre, it provided a detailed and accurate account, including casualty estimates. However, when asked the same question in Chinese, the model deflected, citing government stability measures and referencing China's internet content regulations. Technical analysis of Ox Alpha's tokenization, speed, and caching behavior suggests it is likely a Zhipu model from China, designed to comply with Chinese government registration requirements. AI
IMPACT Highlights the potential for AI models to be developed with built-in censorship mechanisms tied to specific regulatory environments.
RANK_REASON Analysis of an LLM's behavior and technical fingerprints indicating its origin and censorship capabilities. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
