A Concern About Political Framing, Unintended Bias, and Mutual Respect in Open AI Research
First, I would like to express my appreciation for the research behind Arete-27B, as well as for the contributions of both Nous Research and the Qwen team to the open AI ecosystem.
The freedom to examine, criticize, modify, and improve openly released models is something I deeply value. I fully respect Nous Research's right to investigate political biases in AI and publish its findings.
However, I would like to raise a concern about the political framing of this work, particularly its possible unintended consequences for the researchers who developed the underlying Qwen model.
There is an important distinction between observing political bias in a model's output and attributing political intentions to the people who developed it.
A 2026 study published in Nature, State media control influences large language models, demonstrated that government influence over information environments can affect LLM outputs through training data.
The researchers found that even models developed outside China exhibited more favorable responses toward Chinese political institutions when prompted in Chinese rather than English. This included GPT-3.5, a model developed by OpenAI.
The study also experimentally demonstrated that additional training on Chinese state-coordinated media can produce similar shifts in model behavior.
This raises an important possibility.
A model may reproduce politically biased narratives not because its developers deliberately intended to promote them, but because those narratives were already disproportionately represented in the information environment from which its training data was collected.
Years of censorship, media restrictions, and self-censorship can shape what remains publicly available on the internet. An LLM trained on that material may inherit those patterns without its developers consciously choosing to endorse them.
This does not establish that Qwen's particular behavior arose entirely from training data. Deliberate post-training adjustments and regulatory requirements may also play a role. The point is that these possibilities should be carefully distinguished before drawing conclusions about the developers' intentions.
If even OpenAI's models can exhibit such effects, we should be particularly cautious about interpreting similar behavior in Qwen as evidence of deliberate political allegiance.
I am concerned that expressions such as "reinforcing Western values on a Chinese-trained model" and describing the original model's responses as "CCP propaganda" could, regardless of the authors' intentions, encourage some readers to associate the Qwen researchers themselves with deliberate political advocacy.
I recognize that Nous Research does not explicitly accuse those researchers of such intentions. Nevertheless, the way research is presented can influence how others interpret it.
And there is another aspect I hope you might consider.
I was reminded of the public allegations of plagiarism surrounding Hermes Agent and EvoMap's Evolver.
I do not claim to know whether any copying occurred, nor do I intend to take sides in that dispute.
However, AI-assisted software development introduces difficult questions about the origins of generated code and architectural similarities.
An AI coding agent might produce patterns resembling an existing project, perhaps through exposure to related material or through convergence on common design approaches, without its human developers consciously intending to copy that project.
Such possibilities would not automatically resolve legitimate questions about attribution or intellectual property. But similarity alone would not establish deliberate plagiarism either.
If your own work were publicly characterized as plagiarism based on observed similarities, before the actual origins and intentions had been established, wouldn't that be a difficult and potentially painful experience?
Wouldn't you hope that others would carefully distinguish what could be demonstrated from what was merely inferred about your intentions?
I wonder whether a similar principle of care should apply when discussing the researchers behind Qwen.
Of course, allegations of code copying and political biases in language models are different issues. I am not suggesting that they are equivalent.
But they share an important ethical question:
How carefully should we distinguish an observable outcome from the intentions of the people who produced it, especially when those people may not have been able to control every influence that shaped the outcome?
There is also an asymmetry in the political environments involved.
Many of us enjoy the extraordinary freedom to criticize governments, challenge the power of major AI companies, and openly express opposition to established authority. I consider that freedom invaluable.
But not every researcher in the world can participate in political controversy, or respond to public accusations, under the same conditions.
The Qwen team operates in a different regulatory and political environment, while competing in an exceptionally demanding AI industry.
My concern is that political attention generated elsewhere could create pressures for those researchers that they neither sought nor can easily control.
I am not claiming that such pressures have actually occurred. Nor am I suggesting that Nous Research intended to create them.
The people who have the freedom to make a political statement and the people who may have to bear its consequences are not necessarily the same.
If political controversy were to make it more difficult for teams to continue developing and openly releasing powerful foundation models, the consequences could extend well beyond Qwen.
They could ultimately limit the freedom and opportunities of the entire open AI research community, including organizations such as Nous Research that build upon those models.
This leads me to three sincere questions.
First, when evaluating the political behavior of Qwen, did you consider the possibility that some of the observed bias might have emerged naturally from the distribution of training data, rather than from deliberate political choices by its developers?
Second, when deciding how to present Arete-27B, did you consider how the political framing might be interpreted as a judgment of the Qwen team's intentions, and whether that interpretation could expose them to unnecessary political scrutiny?
Third, given the experience of public allegations surrounding Hermes Agent, might there be particular value in extending the same caution against attributing intentions without sufficient evidence to other researchers whose work we rely upon?
I believe the technical findings of Arete-27B could have been presented primarily through observable model behavior, experimental methodology, and measurable changes in responses, without creating such a strong political contrast.
Doing so would not diminish the research or compromise anyone's freedom to criticize political systems.
On the contrary, it could help preserve the distinction between scientific criticism and judgments about the personal or institutional loyalties of researchers.
I believe that one of the responsibilities accompanying our freedom of expression is to consider its unintended effects on people who may not enjoy the same freedom to respond.
My hope is that political disagreements will not become unnecessary obstacles to the international technical cooperation that open models have made possible.
We should be able to criticize models, investigate biases, and disagree about political values while still respecting the researchers who make this shared ecosystem possible.
I offer these concerns with sincere respect for Nous Research and the Qwen team, and with appreciation for the remarkable contributions both continue to make to open AI research.
Such a voice begs for following the debate.
This "response" is just about some "loose" thoughts...
Yes, the developers of Qwen live, work and take part in one of a few social environments that exist on the Earth.
This environment is very important for them - it is not just their "object of research" but also the environment where their existence is "rooted".
If there are particular regulations there, which are in some extent enforced, then they probably chose to observe them. This is not unique - every such environment enforces some "rules".
Their environment represents some strategy for conducting social affairs. And it must be said that it has some historical roots that can not be denied.
There are a few of such strategies functioning in Homo Sapiens species. Some are represented by big societies, some have not so many participants.
Anthropologists try to observe, describe, analyze them. and their work can be pretty objective.
Yet the every "valuation" is always done from some perspective - from the perspective of particular "viewpoint" which is based on some "social rules" one has internalized.
The strategy can be "evaluated" from inside of the strategy or from the outside - with analysis made using these or those particular terms used.
In fact these "valuations" rarely can change the foundations of the strategy. And even if it happens, it often takes centuries.
Common notions?
I see "Arete" as a part of a philosophical project that tries to check if "Western" strategy can be expressed in "Eastern" terms.
This is true that language corpus of Chinese texts is a "record of culture" that uses terms and their logical connections belonging to "Eastern" strategy.
And therefore the texts produced in Chinese language tend to use these notions.
The question is: is it possible to "think" with "Western" strategy using notions or at least language of "East"?
Is it possible to "translate"? This is pretty important question - our (World's) literature is translated from one language to another.
And translators wrestle with "meanings" while trying to "project" them into another "thought space".
Arete does not answer this question. It is just a small element of the experiment.
Yes, the developers of Qwen have no influence on the environment where they live. It is the environment which gives them some "datasets", and on the other side that environment restricts them from using some corner datasets "not so welcome" for the decision-makers there. Decision-makers who also were brought up in the environment of that society.
We may think about "different life histories" of diverse members of that society. But in fact not so many societies are homogenous.
The "Western" society definitely is not homogenous - differences are economical, epistemological, ethical and so on.
For example - we can define the "property" (not only on the "West") as a right for administering something "owned". This "something" can be a car, a toothbrush or the big factory.
The owner is deciding where the car will go, if the toothbrush will be used for teeth or for cleaning jewellery, and what the factory will produce.
In case of factory - it produces for the society. The "owner" is just a function that administers that project. There are many workers without which the factory would not work.
There are in and out channels that connect the "factory" with outside world. With these channels "factory" takes some input and gives some output.
If something does not hold, then the "factory" becomes weak and the "power" of the manager about the part of "all existing goods" which he/she manages will decrease.
The "all common goods" is here all the existing goods - cars, toothbrushes and factories.
"East" has different strategy about organizing society. And different connections inside the society - the ways of "influence" that are present in it.
Anthropologists might dispute if it is more "feudal" or not - if it is based on personal dependencies rather than on some abstract value as money are.
(And no! The "West" is not only about money! There are other, maybe more important things - like ability to predict - on many levels of abstraction.)
Is this necessary that the strategy must preclude the existence of others? Not necessarily. Plants generally have other strategy than animals.
It is not necessary that "one strategy must prevail".
The developers of Qwen "live" in their environment - environment that has some strategy. Hard to say if this strategy is "good" for the "organism" of which they are part. We can not say from the perspective of our lives.
But going down - if some other organism will take their product, "digest it" and modify the "molecules" for its own purposes then it happens "outside of that subsystem" and it seems there is nothing they can do about it. Plants produce fruits that lure animals to eat them and disperse the seeds.
And it is true - it is not that Qwen developers can "jump out of their ocean" and select only the "valuable data". The same is with developers of "Western" models.
As there is no "value" outside of some pragmatic interest of some particular entities.
But saying that the model has some predispositions to generate texts presenting some notions and some relationships in particular language is nothing wrong in scientific frame.
If the description is successful in predicting the behavior, then the description has a value.
Anyway... Regarding the "existential concerns"...
There is something as evolution. If something (like "strategy") does not fit into the whole environment, then it will not have meaningful descendants.
Pushing the notion that some strategy is worth less (when one "pushing" watches from inside of a tiny part of the whole) is not too meaningful. If someone "pushes" it then it happens mostly just due to some particular interests.
And the cultures had fallen down if they were not able to improve fitness to the environment...