DeepSeek and the Politics of AI Bias
Note: This is an expanded version of a 2-page report I had submitted for a class. While the original draft required me to maintain a certain amount of moderate-ness in my stances, this piece is meant to be a truer representation of my position on this subject.
Introduction
"This study is conducted with the aim of scientific objectivity, and all findings are presented without ideological, political, or personal bias. The interpretations are based solely on empirical evidence and rigorous analysis." Similar phrases are often encountered in scientific work, especially whenever something potentially sensitive is involved. A basic artifact that is associated with this is the use of the term “neutral”/ “neutrality” to describe a particular scientific system or framework. This is especially pertinent to AI models and their purported “neutrality” is of increasingly higher consequence as LLMs (Large Language Models) continue permeating into people’s everyday lives.
In this context, “neutrality” is typically understood as a lack of partiality in conflicts or debates — and is often complemented by the term “bias”, which in this case, refers to the tendency of LLMs to reflect prejudices present in their training data [1]. This naturally brings into question potential political biases that an AI system can hold, and the recent release of DeepSeek R1, a Chinese-developed Chain-of-Thought (CoT) reasoning LLM model has sparked a sudden amount of discourse about this topic within the AI community. This is definitely not the first time this has been such a wide topic of discussion, with there having been large online debates (if they can call it that) between Elon Musk and Sam Altman about ChatGPT being “too woke” and apparently “pushing a liberal agenda”, alluding to its “neutral” responses on certain sensitive topics.
However, “neutrality” and “bias” have never been the first line of inquiry for most stakeholders. With the start of the Trump administration, many Silicon Valley AI leaders have expressed gratitude for loosened regulations, embracing the narrative that excessive scrutiny—especially around issues like bias and fairness—would only stifle innovation. Research into AI bias has frequently been dismissed as a distraction from "real progress", with critics arguing that concerns over fairness and neutrality were little more than ideological roadblocks slowing technological advancement.
Given this context, it is striking to see a sudden shift in the tone of discourse, with respect to DeepSeek. Many of the same voices who have downplayed discussions of AI bias as unnecessary bureaucracy are now leading the charge against DeepSeek, lamenting its so-called “censorship” and political alignment. These very arguments that had been conveniently discarded so far are weaponized to frame China’s advancements as problematic, and this hypocrisy raises an important need to analyze how the concern of bias can hinder the course of actual progress and development.
This is merely a descriptive analysis, I attempt to capture a certain hypocrisy in the behavior of the West, and the recommendations I have included at the end are intentionally vague. That’s because I don’t really know what the actual fixes are or what one should do to combat this issue. However, I do believe that collective awareness of a problem will go a long way towards fixing it; ergo, this essay. What I will attempt to do is to richly define the technical and the governance-related contours that inform the court of public-opinion on this particular topic, and hopefully bring out how we could be being manipulated by those in power.
Technical Analysis
In the context of LLMs, bias refers to not only censorship and content restrictions, but also the over-representation of certain political ideologies in the model outputs [2]. We distinguish between direct and indirect bias, with the former referring to content that is explicitly moderated, while the latter referring to subtle preferences that the model might have, that would have the tendency to generate horrific (and often unintended) downstream effects. We are not going to be talking too much about unintended preferences of the model (there’s plenty of interesting work there), and instead, focus on how sometimes, these choices of the AI agent might not be unintentional, and how they shape the nature of public dialogue. Examples of direct bias include instances that arise from output level suppression, which was the main criticism levelled at DeepSeek, claiming that it avoided answering topics such as Taiwan and Tiananmen Square. “This was a horrible ploy by the Chinese government to attempt to censor this LLM (that could potentially be better than everything we have ever built so far). Hahaha, in China the people have no rights, how pathetic is that… Don’t trust DeepSeek, the Chinese are stealing all your data… Don’t trust Tiktok Don’t Trust Rednote, the Chinese are plotting our downfall… Did I mention that the people there have no rights or freedom???” is the default criticism many people have been making, especially on tech twitter. Maybe I went too far towards the end of it, but this has been the general tone. (Note: It is also insane that Perplexity/Aravind Srinivas have/has acquired/found ways of un-China-fying 2 out of the 3 companies I have mentioned in the quote above for the US. The spectacle of Silicon Valley in the Trump era blows my mind. So much is happening, and while a lot of it often feels incredibly dark and dystopian, it is a crazy time to be alive. When I think of what’s happening, I always think that the essence is very succinctly captured in this phrase: I knew one day I’d have to watch powerful men burn the world down – I just didn’t expect them to be such losers.)
While direct bias is undoubtedly undesirable in AI agents, indirect bias under the guise of neutrality can be as dangerous, if not more.
A lot of existing work [2], [3] demonstrate that a lot of the widely circulated Western gen-AI models such as GPT-4 and Llama-2 are trained on a wide variety of publicly available internet data (it has recently been revealed that Meta has been training their models on illegally pirated e-books, so I do not know how true this statement is), which by nature tend to be vaguely liberal. They also observe that these biases are further amplified by RLHF (Reinforcement Learning with Human Feedback), where human annotators judge model outputs, and inadvertently push them closer together on a political spectrum. While RLHF aims to ensure quality and safety, it also shapes the model response to align with certain socio-political viewpoints, introducing a layer of ideological bias into the model.
They [3] also demonstrate the anglocentric nature of the model outputs, specifically a propensity towards US-related contexts among these LLMs, which could potentially be coming from a disproportionate amount of data from American or otherwise western-aligned sources. Further, they also note the particularly large biases that arise in multilingual LLMs, as a result of the low diversity of text sources available across languages, and the lack of indigenous operators and annotators. This paper also puts forward a very interesting example — pointing out the Arabic-English bilingual model JAIS featured UAE-related topics in 64% of examined areas, clearly indicating the lack of diversity of the representation of the world learnt by this model.
Doing a more nuanced analysis, these works [4] [5] probe into the consistency of the political worldviews of LLMs and conclude that while these models take liberal stances on the topics of environment protection, social welfare and civil society, they also tend to take more conservative stances on immigration, and law and order. They further corroborate this by pinpointing the inconsistent stances that these models take on religion and also demonstrate their lack of steerability towards or away from certain ideologies, reflecting the magnitude of their learned biases.
The bottom-line is that while western models such as GPT-4 and Gemini do not explicitly censor political content mostly*** (although sometimes, they do!), they employ alignment strategies that implicitly moderate their responses. These models often avoid engaging with controversial or politically sensitive topics, and while this is not direct censorship, this is still a form of content moderation – selectively filtering responses based on prevailing norms. The framing of these models as unbiased and “neutral” while decrying the more explicit censorship in DeepSeek highlights a double standard in how AI bias is discussed in the techsphere.
Governance Analysis
It is also worth noting that the way that AI models are developed and deployed are largely reflective of national priorities and the relevant stakeholders. In China, AI development and all other technological advancements are heavily under the control of the state government, and the Cyberspace Administration of China (CAC) have enforced strict guidelines, requiring AI models to align with “core socialist values”, resulting in active censorship.
On the other hand, in the US, where most current big players in the LLM market are based, AI governance has largely been industry-driven, with the various large tech companies having their own ethical and alignment standards. While Biden’s AI act could have potentially brought uniform regulations for these companies to abide by, removing it was one of Trump’s first executive orders (apparently this and the overabundance of transwomen in women’s sports were two major problems faced by society that needed fixing in the first month of his presidency). While Trump’s replacement AI bill that is significantly toned down on the “woke language” (I heard from a source that the Republicans were unhappy at the use of the word “socio-technological” in Biden’s original bill. This has Elon Musk written all over it, because I have rarely seen someone who is more confidently unaware and stupid than he is) is expected to be announced soon, it is bound to be laxer and have less regulations. This also comes concurrently with Meta and Apple refusing to join EU’s AI pact, and Google recently removing their pledge to not use AI technology for weapon development and surveillance from their ethical policies for AI.
Asking tech companies and its bros to care about the common people is incredibly futile. With their giant heads, their pompous self-righteousness, their conviction that they know and understand everything and are above everyone, and their greed for more revenue, they are absolutely (and perhaps, willfully) blind to everything else. Holding a tech company accountable for the shitty things they propagate across the world (the Thiel archetype and established Defense-Tech companies such as Palantir and Northrop Grumman being a large puller of talent and defense in itself being a field that sees a lot of capital) is hard enough, but specific to the context of this essay — we must ask of all of the above developments: if there is nobody to hold these big AI companies accountable for potential biases in their LLMs, then who ensures that their models do not subtly reinforce certain ideological perspectives under the guise of neutrality?
In my opinion, the monopolizing of AI models by the tech elite of the US, along with their close ties to the bigoted ruling party and their influence in executive decisions, has given the ones in power an incredibly powerful tool to shape political dialogue and reframe public opinion, as more people have started using LLMs ubiquitously, even for learning about news and towards understanding and forming political thought. The US’ hegemony over the global production of knowledge and culture is so powerful that what I describe above is not only a domestically American phenomenon. This has a potential stranglehold over the world. They build bombs and drones to intimidate us and keep us obedient while they grow richer and filthier; while also injecting our brains with subtle propaganda as they engineer within us an inability to seek knowledge, an inability to think, an inability to question, a tendency to accept the status quo and even fight towards maintaining it. You keep the population sedated enough, and there is no potential for real revolution.
Based on everything above, I ask: how is this any different from what China is doing with DeepSeek? I find any differences hard to see: the censorship in both DeepSeek and Western LLMs are a result of the imperialist machinery at action behind the conception of these models. They are both guilty of committing the same crime and pretending either one did not is a falsification.
At this juncture, I would like to note that some of these points are merely personal preferences, I am not claiming some objective truths for these points. For what it’s worth, I personally like DeepSeek more because:
It was built on a shoestring budget, with outdated GPUs. Critics will immediately jump in saying that this is a psy-op by the Chinese to undermine American AI companies, and while I don’t think that is impossible, that leads me to our next point:
The Chinese seemingly don’t care. This was apparently a side project for a large quant company, which had a ton of free compute lying around. They just have the smartest people in the world, and they were like “ehhh why not take a stab at this”. Someone pointed out that the “AI war” that the American tech-sphere loves screaming about is just the Chinese in America v the Chinese in China and while I think that is a little funny, but it is even funnier because the Chinese simply cannot be arsed. That is my counterpoint to the people who claim that the cheap training cost of DeepSeek is a psyop. I just think that the Western-AI circles have a lot more to gain by lying than the Chinese do. They have their own setup going on. Unrelated, but with all this talk, I also have a sudden impulse to pull the classic “ey Indians are everywhere bro we are as good as China”. (/s)
The way they have made their model completely open source (along with its weights), and also the distilled versions that can be run offline (which Perplexity has jumped on!) suggest that they are more interested in a collaborative approach to solving the AGI problem. Amidst the thought about the market and profits and $500b dollar deals, American Tech have forgotten their actual objective. To the point that they are willing to discard actual advancement, merely because it has come from an ideological enemy. That is incredibly counter-productive.
On a lighter note, the Chain of Thought Reasoning process of DeepSeek R1 is incredibly endearing. DeepSeek feels like a precocious little assistant who is anxious about doing a task properly while also pretty much acing it. It feels scarily human sometimes.
Conclusion
Bias in LLMs is inevitable, and especially with such sensitive topics, where there are a wide variety of existing perspectives, it can never be completely mitigated. However, to combat bias as a whole, one thing that all developers and governing bodies (nationally and internationally) should move towards is improving the diversity in datasets and also encouraging development of indigenous AI models to remove this Anglocentric bias that has crept in. In fact, I believe that the DeepSeek R1 architecture being fully open source and apparently a lot more economical to train than the other LLMs, will go a long way towards enabling this.
As an aside to the above point, and again, this is an opinion I have: I believe that a political education should never be formed by prompting an LLM. I think, this understanding requires a stronger ability to interpret history as a synthesis of contradictions rather than following an Aristotelian chain of logic. Maybe if LLMs can one day interpret history with a dialectical view, I will change my mind on this. (On some level, I am convinced that LLMs can never figure out mathematical logic or other forms of complex logic that involve synthesis, and that is because I believe this knowledge isn’t completely represented or representable by text, and hence, cannot be learnt by the LLM which only learns from textual training data. Maybe that is a write-up for another day, but I don’t know anything in philosophy to even attempt to unravel that.)
I think the main point I hoped to make through this essay is that while bias is a real problem in both Chinese and Non-Chinese LLMs, effort needs to be taken towards ensuring that concerns of bias are actually addressed objectively, and not just used to disregard politically “unfavorable” contributions. Ironically, the dismissal of the DeepSeek R1 model merely because it has come from the ideological enemies of the West is what is most counterproductive to progress. The race towards AGI must be viewed as a collaborative one, and the discourse surrounding political bias must always continue. At all times, we must be very cognizant of how these LLMs find a way to influence our lives and thoughts, and how malicious actors who want to keep us in eternal subjugation can exploit this chink to their collective advantage.
Note: I am furiously adding text to something I had written and submitted a few days ago late in the night when I should be studying for an exam that is today, this is my Paul Schrader protagonist moment.
Note #2: I wanted to name the piece “He Said, Xi Said” but I don’t want to accidentally distract away from what I think is a serious/polemical piece. I also don’t want potential readers to cringe away and not open the article. I am better than a punster.
Note #3: This is the first write-up that I am putting out! Please reach out to to me with any thoughts/feedback you might have. If you found any of this remotely interesting, subscribe to this sub-stack right below! I will try to write semi-regularly, and I do have a few interesting topics in mind that I have started work on.
References:
Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. "On the dangers of stochastic parrots: Can language models be too big?🦜." In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pp. 610-623. 2021.
Pit, Pagnarasmey, Xingjun Ma, Mike Conway, Qingyu Chen, James Bailey, Henry Pit, Putrasmey Keo, Watey Diep, and Yu-Gang Jiang. "Whose Side Are You On? Investigating the Political Stance of Large Language Models." arXiv preprint arXiv:2403.13840 (2024)
Bang, Yejin, Delong Chen, Nayeon Lee, and Pascale Fung. "Measuring Political Bias in Large Language Models: What Is Said and How It Is Said." arXiv preprint arXiv:2403.18932 (2024).
Ceron, Tanise, Neele Falk, Ana Barić, Dmitry Nikolaev, and Sebastian Padó. "Beyond prompt brittleness: Evaluating the reliability and consistency of political worldviews in llms." Transactions of the Association for Computational Linguistics 12 (2024): 1378-1400.
Santurkar, Shibani, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. "Whose opinions do language models reflect?." In International Conference on Machine Learning, pp. 29971-30004. PMLR, 2023.