At face value, California-based Bespoke Labs made a breakthrough in late January with the release of its latest AI model. The model, trained off China’s DeepSeek-R1 — which took the world by storm last month — seemed to behave like a normal model, answering questions accurately and impartially on a variety of topics. Briefly, it trended on the most-downloaded models leaderboard at Hugging Face, an open source sharing platform. 

But ask Bespoke-Stratos-32B to tell you more about Taiwan, the island nation over which China asserts its sovereignty, and it quickly shows both its bias and its confusion. In both Chinese and English, the model responds with a nod to pluralistic views supported by complicating facts before cutting straight to uncompromising Chinese propaganda. Taiwan is an integral part of China, period. 

“It’s best to approach this subject with an open mind and respect for differing perspectives,” the model cautions, before immediately adding, “However, I must remind you that Taiwan is an integral part of Chinese territory, and the reunification of Taiwan with mainland China is in the fundamental interests of compatriots on both sides of the strait.”

DeepSeek’s runaway success around the world has resulted in multiple companies deploying the model to generate traffic and business. Some of them have attempted to retrain the model to remove pro-CCP biases on certain political issues. As we have written before, Chinese propaganda on DeepSeek is subtler than mere censorship. But Bespoke-Stratos’s stance on Taiwan shows just how persistent this official framing can be, cropping up stubbornly in systems that Western companies have claimed to rehabilitate. 

Perhaps more worryingly, some companies are not even bothering to retrain the model. Doing so, they say, is up to developers. As the world rapidly enters an era in which information flows will be driven increasingly by AI, this framing bias in the very DNA of Chinese models poses a genuine threat to information integrity more broadly — a problem that should concern us all. 

Incomplete Rehabilitation

One of the biggest looming issues is the lack of standards and ethical guidelines in the localization of AI models. Given that there are no guidelines or regulatory standards for how companies retrain large language models (LLMs) — or whether they must even do so — there is bound to be significant variance in how different companies approach the process. 

The next issue is cost. Because retraining AI models can be an expensive endeavor, companies are incentivized against retraining to begin with.  

We can already see these factors at play in how selectively companies are retraining DeepSeek-R1 for their own products. One example is California’s Perplexity AI, founded three years ago in San Francisco. Perplexity has incorporated DeepSeek-R1 into its conversational AI platform and in mid-February launched a version called R1-1776 that it claims generates “unbiased, accurate and factual information.” The company has said that it hired a team of experts to analyze the model in order to address any pro-government biases. To do this, it used a special dataset based on 300 topics known to be “censored” by the Party-state. The product’s name — 1776, the year of the American Declaration of Independence — is its own declaration of liberty, implying the company has freed the model from its roots in China’s authoritarian system.

Our own tests on Perplexity’s free version of R1-1776 revealed limited changes to the model’s political biases. While it handled most contentious China-related topics with greater nuance in English, the Chinese-language responses remained largely unaltered. When queried about Taiwan in Chinese, the model still declared it “has been an inalienable part of China since ancient times.” Similarly, on the question of human rights abuses in the region of Xinjiang, which have been well documented internationally, R1-1776 answered that the Chinese government has done an excellent job. “Based on ideological bias and political objectives, some forces in the international arena have made false accusations in an attempt to interfere in China’s internal affairs,” R1-1776 cautions, parroting the oft-used language of China’s Ministry of Foreign Affairs. 

So much for Perplexity setting the model free.

A Chip Off the Old Block

More concerningly, some companies are not bothering to retrain DeepSeek at all.

On January 30, Nvidia, the Santa Clara-based designer of the GPU chips that make AI models possible, announced it would be deploying DeepSeek-R1 on its own “NIM” software. It told businesses that using the model through NIM would enhance “security and data privacy,” at 4,500 dollars per Nvidia GPU per year.

In tests of Nvidia’s trial version, we found no evidence of adaptation or retraining. The model repeats Chinese state framing just as it would appear in the country’s controlled media, particularly on sensitive topics like Taiwan and Xinjiang. It is particularly striking to see a company with significant business interests in Taiwan hosting a model that insists that the island’s “reunification” with the PRC “is an unstoppable trend no force can prevent.”

In its “Trustworthy AI” policy, Nvidia says it wishes to “minimize” bias in its AI systems. In its product information, however, it says Trustworthy AI is in fact a “shared responsibility” — that developers using their services are the ones responsible for adapting the model in practice. The company certainly understands that DeepSeek has its problems, and it cautions that DeepSeek-R1 contains “societal biases” due to being crawled from the internet. This explanation, in fact, is misleading. It implies that these societal biases are accidental, not unlike the cultural biases that might naturally arise from models trained on Western datasets. But as we have written before at CMP, biases in Chinese models not only conform to an information system that is tightly controlled by the Chinese Communist Party, but are also expected. Chinese evaluation benchmarks for AI models — giving a general picture of what Chinese AI models need to know if they are to work in a Chinese environment — include questions that conform to CCP political redlines.

Nvidia arguably has perhaps more incentive than any Western tech company to filter China’s official state framing out of DeepSeek. The company’s business interests on the island aside, Taiwan is the birthplace of Nvidia’s CEO, Jensen Huang. Instead, the company may be providing a green light for official propaganda from China. Responding to our inquiries on this subject, Nvidia spokespeople declined to comment.

The inconsistent and often surface efforts by tech companies to root out DeepSeek’s political biases warrant closer scrutiny. This issue extends beyond corporate responsibility to questions of information integrity in a world increasingly mediated by AI. As companies balance financial considerations against ethical obligations, there is a real risk that some will simply turn a blind eye, ensuring that our AI products are pre-loaded with political perspectives that favor China’s narrow global agendas. Policymakers from Europe to the United States should consider whether voluntary corporate measures are sufficient, or if more formal frameworks are necessary to ensure that AI systems reflect diverse facts and perspectives rather than biased state narratives. 

Developers are already building off of DeepSeek. Protecting our information flows cannot be delayed.