Lazarus AI realigns Alibaba models for Western business use

Lazarus AI's ReAligned models enable U.S. companies to deploy Chinese-trained AI without state censorship barriers

Lazarus AI realigned Alibaba's Qwen models by identifying banned topics from leaked Chinese censorship guidelines and generating prompts to remove them. This process targeted specific state propaganda and security flaws, allowing deployment at approximately 12 U.S. companies. The realigned models could provide information previously suppressed in training data, such as names of pro-democracy activists. Chinese open-weight models like Qwen remain popular for Western businesses—used by Airbnb, Uber, and legal AI firm Harvey for proprietary development—though they lag 3 percentage points behind U.S. closed-system models on benchmark tests. CrowdStrike analysis found DeepSeek R1 models (launched in 2025) had prompts containing politically sensitive topics up to 50% more likely to generate code with severe security vulnerabilities. The U.S. government barred civilian agencies from using DeepSeek products in October 2026. While the realignment process removes state-aligned censorship, it does not eliminate all embedded censorship mechanisms. AI researcher Oriane Peter notes aligning models to Western values is 'an impossible standard' due to contextual dependencies. Europe is developing an OpenEuroLLM model based on Qwen. The U.S. ban on DeepSeek applies only to civilian government agencies, not commercial use.

Source: Rest of World