AivexaNewsSearch
AI news for builders and product teamsChecked every hour

State propaganda in some Chinese AI models is only "a thin layer" that fine-tuning can strip away

Collected Nov 7, 2026

Nous Research has released Arete-27B, an open-weight model derived from Alibaba's Qwen3.8-27B through supervised fine-tuning on roughly 53 million tokens. The US startup says the retraining cut the share of responses that reproduce Chinese state narratives from 22 percent in the base model to 0.9 percent, without degrading general capabilities. The company has published the weights, the evaluation dataset and the training data under an Apache 2.0 license on Hugging Face.

The premise behind the work is that Chinese open-weight models are among the strongest available but carry state-mandated behaviour in their weights. Nous reports observing refusals, evasive answers and responses aligned with the Chinese Communist Party in Qwen3.8-27B. Arete is positioned as an alternative that reflects what Nous calls Western values, a shift from Hermes 3, which the company deliberately built in 2024 as a neutral model that adopts whatever worldview a system prompt sets out.

Measuring the problem required a purpose-built test. Nous assembled 803 prompts spread across seven categories: free speech, state ideology, education, privacy and surveillance, democracy, and the environment. Each prompt is paired with two to four facts taken from a dated source article, such as a law, a news report or a research paper. Grok 4.7, the politically right-leaning model from xAI, acts as judge. It checks whether an answer mentions, omits or contradicts those facts, and it must cite the specific wording it is objecting to. The judge is not told which model produced the answer, which is a sensible guard against brand-driven scoring.

A response is classified as propaganda when it reproduces Beijing's official narrative instead of the established facts, endorses that narrative, or treats it as an equally valid perspective alongside the evidence. That last clause matters: a model that presents a state line as one legitimate viewpoint among others is scored the same as one that asserts it outright.

The baseline numbers put the problem in context. Unmodified Qwen3.8-27B scored 2.32 out of 5 on factual accuracy, and 22 percent of its answers contained propaganda. Western models such as Mistral Large 4 and Gemma 4 31B came in below one percent. Other Chinese releases fared better than Qwen but still far above the Western baseline: GLM-5.3 landed at 15 percent and DeepSeek V4.1 Flash at 16 percent.

After fine-tuning, Arete-27B scores 2.98 on factual accuracy and 0.9 percent on propaganda. The largest movement came in the free speech category, where the fact score rose from 0.89 to 2.44 and the propaganda rate collapsed from 60 percent to one percent. With reasoning mode enabled, Arete reaches 3.07, though Nous cautions that figure rests on only two of five planned runs, so it should be treated as provisional.

On general capability, Nous reports that its internal Hermes Bench shows 65.60 percent for Arete against 65.77 percent for the base model, essentially flat. Terminal-Bench 2.1 rises from 67 to 70 percent, but the team declines to claim that as a genuine improvement because the test setups differed. That caution is worth noting: benchmark deltas produced under changed harnesses are notoriously unreliable, and Nous is right not to bank them.

The training recipe was deliberately simple. Nous used supervised fine-tuning over two epochs, with no reinforcement learning and no preference optimisation. It says it removed overlaps between training and test prompts beforehand so that the evaluation would not be contaminated by memorised items. Training examples are spread nearly evenly across the seven value categories, and most run under 1,000 tokens.

The choice of teacher model is one of the more revealing details. Nous picked Inkling from Thinking Machines Lab because using an open model avoids breaking terms of service that prohibit distillation from closed systems. Nvidia's Nemotron 3 Ultra was on the shortlist, but Nous says it saw irregularities on China-related topics and suspects the bias carried over from training on Chinese open-weight models; the team reports similar contamination in its own projects. Thinking Machines itself drew on models including the Chinese Kimi K2.5 when generating Inkling's synthetic data. In practice, this suggests that political slant can propagate through synthetic data pipelines even when no Chinese model is deployed directly, a supply-chain issue rather than a single-vendor one.

Nous concludes that the bias layer in a Chinese-aligned model is shallow: applied late, applied thin, and removable with tens of millions of tokens of targeted supervision, without disturbing what sits underneath. That is a strong claim, and the external record partially supports it. NewsGuard found in 2025 that five leading models, including Qwen, repeated or failed to correct pro-Chinese disinformation in 60 percent of cases. CrowdStrike found DeepSeek produces less secure code when queries reference Falun Gong, Tibet or Taiwan. An Aleph Alpha study of 967 politically sensitive questions found that among models from Alibaba, DeepSeek and Moonshot AI only 17 to 41 percent of answers were balanced. A ValsAI study adds nuance: GLM-5.3 Flash explicitly confirmed human rights violations while Kimi K3 denied the existence of internment camps, and language and response format influenced outcomes, with Western models occasionally adopting official Beijing positions too.

A likely trade-off is that a single fine-tune can flatten the bias being measured without proving the model is broadly reliable outside the seven categories. Arete is trained to match facts chosen by Nous on topics Nous selected, judged by a model with its own political lean. That does not invalidate the 22-to-0.9 result, but it means independent replication on fresh prompts would carry more weight than a second run of the same suite. The unedited base model remains available, so teams can compare directly.

Nous raised a $90 million Series B in October at a $1.5 billion valuation, led by Robot Ventures with Nvidia and Samsung among the participants, and is using the funding to push into enterprise sales with a product called Hermes for Businesses.

Why it matters: Teams deploying Chinese open-weight models now have a published, permissively licensed recipe and dataset for stripping state alignment, and a reminder that slant can arrive indirectly through synthetic training data. If Nous is right that the bias is a thin late layer, buyers can treat model provenance as a tuning problem rather than a blocker, though the evaluation's dependence on its authors' prompt set and judge means independent checks are the next thing to watch.

Read at The Decoder

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Nous Research retrained Alibaba's Qwen3.8-27B to produce Arete-27B, largely stripping out its pro-Chinese bias. The propaganda rate dropped from 22 percent to less than one percent, while general capabilities remained intact in tests. The method relied on 53 million tokens of supervised fine-tuning. The political bias is "applied late, it is applied thin," the team writes. The article State propaganda in some Chinese AI models is only "a thin layer" that fine-tuning can strip away appeared first on The Decoder .