The idea of machine unlearning, which involves eliminating specific data from an AI system after it has been trained, was primarily theoretical when Ben Luria, an Israeli entrepreneur, co-founded his startup. However, everything changed last year when China's DeepSeek introduced its first open-weight large language model, prompting Luria to recognize a significant business opportunity in adjusting it to align with Western principles.
Several Chinese open-weight models, including those from DeepSeek, Moonshot, Z.AI, and Alibaba, have gained traction among users, including American firms and startups, due to their cost-effectiveness, robust features, and flexibility. However, concerns about these models being too closely associated with the Chinese Communist Party and their safety compared to closed systems persist.
According to Ben Luria, co-founder of the AI safety startup Hirundo, “We’re not making the model anti-Chinese. We are bringing it up to speed on safety and to the standard of mainstream Western AI.” In an interview with Rest of World, he explained that their objective is to minimize risks for businesses adopting open-weight models. The company employs precision techniques akin to neurosurgery to adjust the models, ensuring they reflect the values, geopolitical perceptions, censorship tendencies, and biases that Western companies expect.
Hirundo's team modifies the model weights to enhance security against cyber threats, promote fairness, and diminish access to hazardous information like bio-weapons data. A report from cybersecurity firm CrowdStrike noted that prompts touching on topics deemed sensitive by the Chinese government were up to 50% more likely to produce code with serious security vulnerabilities. Additionally, another study indicated that code generated by Chinese models was less secure when users identified as U.S. government officials.
Hirundo’s evaluations of Alibaba’s Qwen models revealed their lack of acknowledgment regarding sensitive political topics in China, such as the repression during the Tiananmen Square protests in 1989 and events in Hong Kong in 2019, along with the existence of forced labor camps in Xinjiang. Discussions around the character Winnie the Pooh are also avoided, likely due to its use as an allegory for Premier Xi Jinping among dissidents.
Luria emphasized that Hirundo's adjustments aim to provide the balanced viewpoints similar to those offered by OpenAI's ChatGPT and Anthropic's Claude. He reiterated, “We’re not making the model anti-Chinese. We are bringing it up to speed on safety and to the standard of mainstream Western AI.”
Chinese models are catching up to leading closed systems from U.S. tech labs, trailing by a mere 3% in benchmark scores and attracting considerable download activity on the Hugging Face platform. Notably, companies like Airbnb, Uber, and the AI firm Harvey are generating proprietary models based on these Chinese innovations, while Europe is working on its OpenEuroLLM sovereign AI model using Qwen technology.
Despite their advantages, apprehensions about security remain. The Trump administration recently prohibited civilian government agencies from utilizing DeepSeek products or services, citing national security risks.
Graham Webster, editor-in-chief of the DigiChina Project at Stanford University, echoed concerns about using Chinese models in sensitive contexts. However, he noted that for low-risk applications, such as developing websites for small businesses, the potential issues may be minimal.
In Boston, Lazarus AI is providing on-premise AI services for the U.S. government and other regulated sectors, successfully realigning Alibaba’s Qwen models by removing enforced censorship. The company identified sensitive topics outlined in leaked censorship directives and generated relevant prompts, utilizing responses from Google’s Gemini Flash to recalibrate the models.
Post-intervention, these models began providing information previously suppressed, such as the names of pro-democracy activists, indicating they had initially acquired this knowledge but were restricted from sharing it, according to Eric Hartford, former chief scientist at Lazarus AI.
“In the United States, there is a prevalent anxiety about Chinese models inheriting dangerous behaviors,” he said. He aims to present these models to businesses and heavily regulated sectors in a way that makes them more acceptable.
Lazarus AI’s ReAligned models, promoted as “open-source frontier models free from propaganda,” have been adopted by around a dozen U.S. companies. However, many clients, particularly government entities, are wary of embracing Chinese models despite their superiority over other open-weight versions from OpenAI and Thinking Machines Lab.
Despite claims of effectiveness from companies like Hirundo and Lazarus, Oriane Peter, a Ph.D. researcher in AI at King’s College London, argues that aligning a model to Western values or crafting it to be universally safe is a “near-impossible standard.” She describes alignment as the process of guiding a model’s output to meet specific desired behaviors within a defined cultural context.
Thus, those employing Chinese models beyond China must adapt them, not solely because Western ideologies are inherently safer or more ethical, but due to the necessity of tailoring models to the sociocultural environments in which they will be used.
As the adoption of Chinese open-weight models grows, tools designed to uncensor them and eliminate their censorship mechanisms—referred to as abliteration—are becoming more common. However, many of these tools fall short, with models often still retaining elements of censorship and misinformation, Luria noted.
He pointed out that achieving these alignments is complex and deeply ingrained in the models, making it challenging to eradicate them.
Globally, there is a mounting backlash against Western biases in AI systems. Researchers are observing that even some models from China and Russia default to American-led narratives about historical events. Countries across Asia, Latin America, Africa, and the Middle East are developing their own sovereign AI models designed to be more reflective of their unique languages and cultures.
The centralized control of model alignment in Silicon Valley can risk creating a “massive homogenization effect” in AI systems, as noted by Peter, which has been linked to cognitive uniformity and intellectual decline among users.
Most users of Chinese models seem unconcerned about these alignments, according to Graham Webster. He explained that while caution might be warranted in sensitive regions like Taiwan, low-risk applications like those aimed at small businesses are generally not affected by the underlying issues present in the training data.
Developer Stu Clott, based in San Diego, utilizes DeepSeek for various tasks. He stated that he has yet to encounter situations where the model's censorship posed a problem for him.
