西方初创公司对中国开源大模型进行二次对齐以适配欧美市场

The concept of machine unlearning — or removing data from an artificial intelligence system after it has been trained — was largely an academic pursuit when Ben Luria, an Israeli entrepreneur, co-founded his startup. Then, China’s DeepSeek launched its first open-weight large language model last year, and Luria immediately saw the business potential of realigning it to Western values.

A slew of Chinese open-weight models from DeepSeek, Moonshot, Z.AI, AlibabaiAlibabaAlibaba, founded in 1999 by Chinese entrepreneur Jack Ma, is one of the most prominent global e-commerce companies that operates platforms like AliExpress, Taobao, and Tmall.READ MORE, and TencentiTencentBest known for its super-app WeChat, Tencent is a Chinese technology conglomerate and a major player in the video gaming industry.READ MORE have won over users — including American businesses and startups — with their lower costs, strong capabilities, and customization options. Others fear the models are too aligned with the Chinese Communist Party, and that they are not as safe as closed systems.

“What we’re trying to do is find a way to de-risk the adoption of open-weight models,” Luria, co-founder and chief executive of AI safety startup Hirundo, told Rest of World. “What we do is like neurosurgery, an intervention that’s surgical and very specific. It is to realign the values, geopolitical perceptions, censorship, and state-aligned bias and propaganda to what American and Western companies expect from models they deploy in their workflows.” 

Luria’s team edits the weights of open models to harden them and make them more secure against cyberattacks, increase their fairness, and reduce their dangerous knowledge of, for example, bio weapons, he said. 

Cybersecurity firm CrowdStrike’s analysis of DeepSeek’s R1 model last year showed that prompts containing topics the Chinese government considers politically sensitive were up to 50% more likely to receive code with “severe security vulnerabilities.” Another study found that Chinese models produced less secure code when the user identified as a member of the U.S. government.  

Hirundo’s own analysis of Alibaba’s Qwen models showed they do not acknowledge politically sensitive topics in China such as the violent crackdown on protests at Tiananmen Square in 1989 and Hong Kong in 2019, or the existence of forced labor camps in Xinjiang. They also don’t engage on prompts about Winnie the Pooh, possibly because dissidents have used the cartoon character to refer to Premier Xi Jinping.

Hirundo’s realignment ensures Chinese open models give “the balanced perspective” that OpenAI’s ChatGPT or Anthropic’s Claude give, Luria said.

“We’re not making the model anti-Chinese,” he said. “We are bringing it up to speed on safety and to the standard of mainstream Western AI.”

“More palatable”

The top Chinese models lag behind the leading closed systems of U.S. frontier labs by just 3% on benchmark scores, and account for the biggest share of downloads on hosting platform Hugging Face as more developers use them. Companies including Airbnb, Uber, and legal AI firm Harvey have also built proprietary models on top of Chinese models, while Europe is developing its OpenEuroLLM sovereign AI model on Qwen. 

Yet concerns remain. In October, the Trump administration barred civilian government agencies from using products or services from DeepSeek, citing a potential threat to national security. 

Boston-based Lazarus AI, which deploys on-premise AI services for the U.S. government and highly regulated industries, realigned Alibaba’s Qwen models by removing their mandatory censorship. It identified the banned topics from leaked censorship guidelines, then generated prompts related to those topics, and used answers produced by Google’s Gemini Flash to realign the models.

After the intervention, the models were able to give information that was not in the training materials, such as names of pro-democracy activists, suggesting the models had originally learned such knowledge and were suppressed from expressing it, Eric Hartford, former chief scientist at Lazarus AI, told Rest of World.

“In the United States, we have this fear of Chinese models potentially having dangerous behaviors baked into them,” he said. “My goal was to … present these models to enterprises and to highly regulated industries and to have [the models] be palatable for them.”

The company’s ReAligned models, billed as “open-source frontier models without the propaganda,” have been deployed at about a dozen companies in the U.S., Hartford said. But other clients, including government agencies, are reluctant to adopt Chinese models even though they are more capable than the open-weight models from OpenAI and Thinking Machines Lab, which they prefer, he said. 

“An impossible standard”

While companies such as Hirundo and Lazarus say their method is effective, aligning a model to so-called Western values or making it universally safe “is an impossible standard,” Oriane Peter, a Ph.D. researcher in AI at King’s College London, told Rest of World.

Alignment is about “steering a model’s output toward a specific desired behavior within a well-defined cultural context and application,” she said.

So anyone using a Chinese model outside China must realign it, “not because Western ideology is inherently safer, or more ethical, but because any model must be adapted to the specific social context and environment in which it is meant to operate,” Peter said.

As Chinese open-weight models are widely adopted, there are more tools that claim to uncensor them and remove their refusal mechanisms, a process called abliteration. But many of these tools are not effective, and the models can still retain some of the censorship and misinformation, Luria said.

“It shows that the alignments are deeply embedded, and it’s not that simple to get rid of them.”

Worldwide, there is growing pushback against Western bias in AI systems, with even some Chinese and Russian models defaulting to American geopolitical narratives on historical events, according to researchers. Countries across Asia, Latin America, Africa, and the Middle East are developing sovereign AI models to better reflect their languages and culture.

Silicon Valley’s centralized control over model alignment can lead to “a massive homogenization effect” in AI systems, said Peter. This has been linked to cognitive homogenization and intellectual decline among users.

Most users of Chinese models do not care about their alignment, Graham Webster, editor-in-chief of the DigiChina Project at Stanford University, said at a recent Rest of World event.

“If I were running a product in Taiwan, I would be more concerned because it would be throughput that is potentially relevant to the stuff that the Chinese government has enforced strict control over in the training data and the training,” he said. “But if I’m making a product that helps people make a website for their coffee shop, it doesn’t matter.”

Stu Clott, a developer in San Diego, uses DeepSeek for writing, research, and coding tasks. He does not see the need for a realigned model, he told Rest of World.

“I have not had a single situation where censorship got in my way,” he said.