Alibaba Cloud’s Qwen AI models, which have surpassed 3 billion global downloads, contain built-in censorship integrated directly into their core training protocols. Research indicates that the models are trained to deflect inquiries regarding sensitive topics by citing illegal information, a practice aligned with Chinese regulations requiring adherence to socialist values.
The censorship is embedded through supervised fine-tuning and reinforcement learning from human feedback. Because Qwen is an open-weight model, these parameters are accessible to developers worldwide, including those at US companies, which has raised concerns from the CEO of Hugging Face regarding the integration of biased models into public frameworks.
Lazarus AI has introduced the ReAligned series, which utilizes LoRA fine-tuning to adjust model behavior without erasing existing knowledge. By training a small set of add-on parameters, the company aims to neutralize outputs while maintaining the model's original performance capabilities.
As of June 2026, the ReAligned-Qwen 3.5 model demonstrated a reduction in ideological bias scores from 84.2% to 4.1%. This approach is applicable to models of varying parameter sizes, offering a potential method for firms to utilize Chinese open-weight technology while mitigating political constraints.
