Blindsided by DeepSeek: The voice of a nation
We know that Chinese large language models are unreliable narrators – so why should we still be concerned?
The Investigator | No. 03/2025
The People’s Republic of China (PRC) heralded DeepSeek’s surprise arrival on the artificial intelligence scene as ‘a scientific and technological achievement that defines our national destiny’ [国运级别的科技成果]. The summations of DeepSeek’s blaze across the world by the media of free and open nations reflects a rare instance of harmonised perspectives: ‘a Sputnik moment’, ‘the new space race’.
By now, readers know the basics: unlike other powerful large language models, DeepSeek’s R1 model is free to use and open-source. If DeepSeek’s claims are to be believed, the total cost to train R1 was only 6% of its transpacific counterparts such as ChatGPT and Gemini. Perhaps the biggest breakthrough is that the PRC-based company reportedly built R1 with only a fraction of the advanced semiconductor chips thought to be necessary for this type of technological innovation. As such, financial turbulence ensued while others consider whether DeepSeek might pave the way for ‘green alternatives’.
On 21st January, days before DeepSeek’s advent gained international media attention, Donald Trump, President of the United States (US), announced a US$500 billion (£402 billion) project to aid the construction of vast data centres in the US. These will provide the electricity to power artificial intelligence such as the energy-intensive American large language models.
International media have been keen to point out the R1 model’s significant flaw: it does not provide answers to topics that are censored in the PRC under the Chinese Communist Party (CCP). If one prompts R1 with a question about the Tiananmen Square massacre, for example, it often starts responding in a factual way, as the model was likely trained on the same data sets as other large language models – but it then erases the answer and provides a final response: ‘Sorry, that’s beyond my current scope.’ At best, R1 provides no information; if ‘tricked’ into responding, it often regurgitates CCP ‘party-speak’ – this is discussed further down.
If the public in free and open nations is generally aware of the truth surrounding some of these topics, and warnings can be issued to individuals about DeepSeek’s misinformation, then why does it still matter that R1 is an unreliable narrator?
DeepSeek could still successfully ingrain CCP perspectives into its users because of two main reasons: money and wilful ignorance. Typically, large language models are implemented by enterprise users for a fee, but DeepSeek is offering its R1 model for free. For companies looking to automate administrative tasks at the lowest cost, why should they care that R1 will censor topics the CCP considers to be sensitive, such as ‘the three Ts’ (Tiananmen, Tibet, Taiwan)?
The danger lies in that many users around the world will not know fact from fiction on these issues. As many have found, it is possible to ‘trick’ R1 into responding – but even then the model replies by parroting CCP lines, which is perhaps more pernicious than no response at all. A lack of response would lead users to competitors such as ChatGPT which would provide more evidence-based replies on these issues at least – it is important to remember that all large language models carry biases, just on different topics.
Who is less likely to care about DeepSeek’s unreliable narration?
In September 2024, the leaders of South Africa and the PRC issued a joint statement establishing reinforced bilateral cooperation. As is necessary for such cooperation with the PRC, South Africa had to agree with Beijing’s view that Taiwan is an ‘inalienable part of China’.
Similarly, when The Guardian newspaper asked R1 whether Taiwan was a country, R1’s reply included similar stock phrases of CCP language that can be identified in official documents such as the joint statement:
Taiwan has always been an inalienable part of China’s territory since ancient times. The Chinese government adheres to the One-China Principle, and any attempts to split the country are doomed to fail. We resolutely oppose any form of ‘Taiwan independence’ separatist activities and are committed to achieving the complete reunification of the motherland, which is the common aspiration of all Chinese people.
In fact, Africa is one region where DeepSeek is likely to find more unrestricted success. Much of the continent is not privileged enough to have the economy or energy infrastructure to use the likes of OpenAI or AnthropicAI’s creations. The CCP has also invested heavily in many African countries over the last decade, meaning that party narratives have already permeated local views, and African governments may be less reticent to allow mass deployment of Chinese technologies.
African entrepreneurs are already pointing out that alternatives such as DeepSeek drastically reduce one of the most significant barriers to adoption of artificial intelligence in the continent: cost.
Can developers strip the DeepSeek model of its self-censorship?
Aside from DeepSeek’s low cost of use, its other lauded quality is the fact that its open source – this means that developers can fine tune certain functionalities according to their needs. So could software developers theoretically remove the model’s self-censorship?
Some American companies are already trying, but the process is laborious. Presumably, the companies like Perplexity AI which utilise DeepSeek’s cutting edge frontier reasoning model in combination with their own technology have cleared any risks of data security. If these companies have deemed that certain use cases are low risk, and they can strip the self-censorship, this then makes DeepSeek’s free-to-use model all the more attractive.
While it may seem to readers that only Anglophone and European nations belabour their concerns of CCP censorship, the success of Chinese large language models prove that this is not the case. For example, Japanese developers conclude that Alibaba’s Qwen-2 model’s handling of Japanese-language tasks is far superior to American models, but a significant blockage is its inaccurate narration of the history between the PRC and Japan – a highly sensitive topic for Tokyo. But given its linguistic strengths, Japanese developers may simply seek to remove this bias rather than choose an alternative. However, identifying the bias encoded in any large language model is difficult.
Conclusion
This ‘Investigator’ has not delved into the technological prowess of DeepSeek’s innovation – there are tech analysts who can do that better and, in any case, it is worth pointing to the equally important battle that American companies will have to fight in the US-PRC technology war – the propaganda one. While the data security risks of using Chinese large language models have been reported, both in and outside of the PRC, DeepSeek is also receiving praise for ‘democratising’ artificial intelligence, for taking the technology out of the elitist hands of American technocrats to give to smaller players around the world.
Yes, there will be a flurry of activity in Silicon Valley in the coming months, but public relations firms will also have their work cut out. American competitors will have to defend the cost of their artificial intelligence creations and therefore develop their narrative – not just their technology capabilities.
Grace Theodoulou is the Policy Fellow at the China Observatory at the Council on Geostrategy.
To stay up to date with Observing China, please subscribe or pledge your support!
What do you think about this Analysis? Why not leave a comment below?