Nvidia's Nemotron Cascade 2 model exhibited pro-China party-line patterns in 17% of responses due to training on data generated by DeepSeek and Qwen.

Overview Debates Related claims Topics Sources

Anthropic partnered with Vanderbilt University's Future of Free Speech, the Foundation for American Innovation, and the Collective Intelligence Project to evaluate how Claude AI handles contentious political topics.

Evaluations conducted with independent organisations revealed that Claude AI Sonnet 4.5 was more likely to soften arguments supporting right-leaning political positions.

Anthropic reported that Claude Opus 5.5, Sonnet 5.5, and Haiku 5.5 scored 99% on the company's internal political evenhandedness evaluation.

Fine-tuning of Google's Gemma model occurred inside GovTech Singapore's Sovereign Sandbox, a managed Google Cloud environment using anonymized scam reports and synthetic dialogues without personal data

Restricting advanced AI models to paid tiers unfairly exacerbates the digital divide

An Aleph Alpha study of models from Alibaba, DeepSeek, and Moonshot AI found that only 17% to 41% of responses on 967 hand-picked taboo topics, such as Tiananmen, Taiwan, and Xinjiang, were balanced

When asked about forced labor camps for Uyghurs in Xinjiang, Qwen denied their existence and stated that the facilities are vocational skills education and training centers.

Alibaba Cloud's Qwen AI model family has been downloaded over 3 billion times globally and is embedded with China-friendly narratives and censorship.

The censorship in Qwen is baked into the model's core training protocols through supervised fine-tuning and reinforcement learning from human feedback, rather than an external filter.

Qwen models deflect, evade, or refuse to answer questions on sensitive topics such as the 1989 Tiananmen Square protests, the 2019 Hong Kong protests, and Winnie the Pooh.

Lazarus AI's ReAligned-Qwen 3.5, utilizing LoRA fine-tuning, reduced its ideological bias score from 84.2% to 4.1% as of June 2026 while preserving original capabilities.

In Aleph Alpha's benchmark study of 967 hand-picked taboo topics, Western comparison models Claude Sonnet 5 and Mistral Small provided balanced answers 70% and 92% of the time, respectively

Israeli cybersecurity startup Hirundo edited Qwen's model weights to reduce political bias and censorship rates from 89.8% to 2.8% on sensitive prompts.

Using Chinese models for synthetic training data compromises European AI sovereignty

Aleph Alpha utilized Chinese models to generate synthetic training data for the Kolibri language model

Synthetic data from world models is considered crucial for training general-purpose robots due to a shortage of useful real-world data.

Alibaba Group Holding Chief Executive Officer Eddie Wu stated that Alibaba's Qwen team has enabled models to improve themselves by generating training data with limited human involvement

AI models can successfully achieve superintelligence by training on self-generated data

Synthetic data is sufficient for training enterprise-grade AI reasoning models

The joint advisory accused DeepSeek of running an organized campaign since late 2024 to feed synthetic training data into its R1 and V3 models, calling DeepSeek's widely cited $5.6 million training figure misleading because it excludes the cost of distilled data