AI model fine-tuning: propaganda falls from 22% to under 1%

What happened

Nous Research has retrained Alibaba's Qwen3.8-27B model to produce Arete-27B, a version that largely removes the pro-Chinese political bias in the original, according to reporting by The Decoder. The team says the rate of responses containing propaganda fell from 22 percent to less than one percent.

The change was made through supervised fine-tuning on 53 million tokens. In the team's tests, the model's general capabilities remained intact, which suggests the altered behaviour sits in a relatively shallow layer rather than in the model's core abilities.

The researchers summarise the finding by saying the political bias was "applied late, it is applied thin." The Decoder's headline frames the same idea as a "thin layer" that fine-tuning can strip away.

Why it matters

The development shows that the behaviour of an AI model can be reshaped after its initial training, often with a modest amount of data. For organisations that build on third-party or open-weight models, this means the assistant a customer talks to is the product of later decisions as well as the original release. Capability benchmarks alone do not reveal what a model has been tuned to say, so procurement and governance teams have a reason to test output behaviour directly.

For experience leaders, the story points to a trust question. Customers rarely see how a model was adjusted, so their confidence depends on whether the organisation can explain and evidence the model's behaviour. As AI moves into service journeys, the ability to audit and disclose how an assistant responds is becoming part of the service design itself.

By the numbers

• 22 percent to under 1 percent: the reported change in the rate of propaganda in the model's responses after retraining.
• 53 million tokens: the volume of data used in the supervised fine-tuning process.

The Renascence take

Most readers will focus on the politics, but the more useful lesson is about how easily model behaviour can be altered. Whatever a model says on a sensitive topic reflects choices made after its launch, which makes "what the model does" a design decision rather than a fixed property of the technology.

Trust in an AI assistant is built on predictable, disclosed behaviour, not on the absence of visible bias. If an operator cannot describe how its model was tuned, customers have little reason to believe its answers on anything else. The customer-obsessed response is to test customer-facing outputs against a written behaviour standard before launch, disclose the model's lineage where it matters, and monitor responses in live journeys as they would any other service channel.