State propaganda in some Chinese AI models is only a thin layer that fine-tuning can strip away
🤖 AI-generated content — The title and summary were produced automatically by artificial intelligence, without human editorial review.
Nous Research supervised fine-tuned Alibaba’s Qwen3.8-27B on 53 million tokens to build Arete-27B, an open-weight model aimed at Western rather than Beijing-aligned answers. Measured pro-party propaganda fell from 22% to under 1% while general capability stayed intact; the team says bias sits in a thin late layer. Tests used 803 prompts in seven categories, with xAI’s Grok 4.7 judging answers against dated factual baselines.
Continue in the app — vote & join in ➔Source: The Decoder (AI) · via ahirlevel.hu