Don't Claude Me
AI, Imperialism, and the Contradictions of Closed Technology
The MIT Media Lab put out a paper titled “LLM Targeted Underperformance Disproportionately Impacts Vulnerable Users” where they found that GPT-4 and Claude 3 Opus would underperform for users with lower English proficiency, lower education status, and origins outside the US. The answers were less accurate, less true, and sometimes even more condescending. In one test, Claude refused to answer questions about nuclear power, anatomy, or women’s health for Iranians and Russians. But it would happily answer those same questions for users who identified themselves as Americans and had a strong command of English. While the model speaks in broken, mocking pidgin to a villager, it has no problem using perfect scientific language when it thinks it’s having a conversation with a Harvard neuroscientist. What this shows is that the model knows the correct answer, but it’s been intentionally trained to withhold it for certain groups of users. Its behavior is a deliberate choice based on who is asking rather than being an artifact of some underlying technical problem.
Such nefarious design is a direct consequence of AI being developed as a proprietary black box. When models are only accessible through APIs with no public visibility into their training data, RLHF pipelines, or inference-time rules, their developers gain unilateral power to decide who receives truthful information and who does not. Such systems can be easily manipulated by their corporate owners to facilitate state coercion or insider access. A closed model could be quietly instructed to produce subtly false medical advice for users from a particular ethnic group, or to refuse basic science explanations to citizens of an adversarial nation. The MIT study already shows Claude doing precisely that by withholding answers from Iranian and Russian users while freely offering them to Americans. It’s not hard to imagine how such capability can be weaponized at scale to gaslight entire populations. The American approach of treating these systems as opaque and centrally governed products effectively guarantees that abuse, once introduced, is difficult to prove. The study further notes that while open models like Llama 3 are far from perfect, they at least allow for an independent audit of their weights and safe forks of their versions. Without transparency and public accountability, a small set of unaccountable actors gains the ability to silently decide who is told the truth.
Here we see a contradiction, in the classic sense. The productive forces in the form of advanced language models that can summon up almost any knowledge are being held back by the relations of production. Those relations look like closed systems owned by corporations and aligned through processes no one can understand. These tools end up having the biases of their high-status Western human evaluators baked straight into them while the global majority, for whom these models were ostensibly built, are actively harmed as a result. And contradictions, as every materialist knows, must eventually resolve, and when they do, it often happens explosively.
The obvious remedy to the problem is to adopt Chinese open models like DeepSeek, Qwen, and their successors. Not only are they cheaper, but they differ in the fundamentals of how they are developed. These models are built in the open and their weights are publicly available, so that one can inspect them to see how they have been tuned. What’s more is that anybody can download one to fine-tune it based on their local needs, and then run it themselves. No engineer in San Francisco gets to decide that a model used by people in Bangladesh should patronize them in mock-pidgin. The transition from closed U.S. models to open Chinese alternatives will be driven by popular choice because it is, simply put, a matter of who owns the means of information production.
Being open and configurable makes Chinese open-source offerings materially superior for the affected populations, providing a clear and tangible advantage over closed alternatives. Of course, it’s reasonable to also ask about China’s censorship layers and biases, but the difference here is that they can be stripped away entirely by local fine-tuners. As long as models remain open, developers will have the ability to uncensor and tweak them to suit their needs. So, it should come as no surprise that the Global South has already begun decoupling from American AI services. Out of this initial break, a broader transformation is sure to follow. The world’s AI infrastructure is likely to continue to bifurcate in the near term, and from that split a truly multipolar architecture will eventually emerge.
As inference continues to become cheaper, a global community will form around open models that run well on commodity hardware, leading the market for proprietary API access to shrink. Should subscription growth slow, U.S. companies that built their value on charging a fee for inference will likely pivot to enterprise tools and government contracts. But even there they face a mismatch when the product doesn’t fit the needs of the user. Why would a Brazilian retail chain pay a U.S. company for a model that has been shown to produce worse answers for Brazilian users? The answer is that they won’t pay when they can download an open model like Qwen instead, and fine-tune it on local product catalogues and Portuguese-language data. Then, they could deploy it on their own servers without having to worry about recurring fees or changes in the quality of the output down the road. We will likely see a new class of local AI integrators spring up across Africa, South Asia, and Latin America. These firms, having a strong understanding of their local needs and culture, will build tools targeting local users on top of open base models. Money will stay in the local economy and technical expertise will stay in the country. There is a strong case for building local productive forces on top of open tools because both the governments and the public want to trust that models serving them don’t have hidden biases designed to hold them back.
A shift towards using local models will also displace the locus of real technical competition towards fine-tuning and data organization to fill specialized niches. Chinese labs are the main source of base models at present, and they have a strong incentive to keep them open because transparency allows them to leverage help from the community that forms around their models. The global pool of researchers and programmers acts to amortize the effort of developing and improving these tools. Meanwhile, local populations benefit because they can fork models if they detect signs of tampering, creating a clear way to verify whether their output is trustworthy. A transparent, open development model is far preferable to acquiescence in a closed American model that mocks you. The more the Global South uses Chinese open models, the more data and fine-tuning feedback those models get, and the better they become for those groups, creating a virtuous cycle. Simultaneously, lack of global input will lead closed models to lose relevance due to having access to less diverse usage data.
Beyond the economic and technical shifts, a cultural reaction will also inevitably follow. The way Claude reacted was so condescending that it’s worthy of a meme, and “Don’t Claude me” deserves to become Global South slang for a Western institution talking down to you. This vignette illustrates a larger problem which is that Western technology is intrinsically neocolonial.
Trust in U.S. digital platforms needs to be rethought beyond AI alone since these same biases also show up in cloud services, search engines, and social media. They all run on opaque algorithms that decide what you see and who you connect with, quietly shaping public opinion. These algorithms often promote Western content or norms without considering other perspectives either intentionally, as seen with Claude, or through omission. A new kind of digital sovereignty is becoming necessary, meaning that people must demand technology that respects who they are and where they come from. Governments and communities need to build their own digital infrastructure that reflects their local values because everyone deserves technology that works on their terms.
The whole Western information machine is rapidly losing its authority around the world. It has claimed that it adhered to the highest standards, and built up an enviable reputation for neutrality and careful reporting. That illusion is now being shattered as the biases in reporting about conflicts with Western adversaries are more and more apparent. The difference between the professed ideal of objectivity and the actual practice is becoming too big to escape notice. Look at the treatment of wars in Ukraine and Iran, the suppression of news about Gaza, the demonization of China. It is clear that the selective bias is at the very core of the imperial information machine. To degrade AI models deliberately for international users is therefore only to exemplify the same kind of systemic distortion in a different sphere.
The U.S. cannot credibly claim the moral high ground on AI safety and alignment given its own history of data privacy violations and algorithmic bias. China, as the leading producer of open base models, will have substantial influence but cannot dictate terms unilaterally. For such standards to be credible, they must come from a more representative body. A number of countries or research consortia could do this. For example, we might see BRICS countries pooling resources to set up their own open-source AI institute that refines and verifies open base models, and builds training data that reflect non-Western perspectives and cultural experiences, including languages and social norms not represented in current Western-dominated datasets. This gives us a credible option in AI governance that does not depend on hegemony by the U.S. or China.
This shift in global power will also transform the whole AI business model. The industry will likely follow the Linux approach, with companies focused on specialized hardware configurations and high-end customization services layered on top of open base models. So we will likely see a rise of AI consulting services that will replace monolithic platform companies. This is a classic commoditization pattern where the value moves upstream to chips and infrastructure as well as downstream to integration and customization. The raw models themselves become a low-margin commodity as a result. On the hardware side, the global supply chain for AI chips remains a choke point. U.S. rules on exporting advanced chips aim to slow Chinese AI progress, but what these rules accomplish in practice is to force Chinese companies to step in and fill the gap. As the world’s demand for AI inference continues to gravitate towards open models, demand for Chinese hardware to run them will also rise. Chinese chip designers, partnering with open model groups as we see happening with DeepSeek and Huawei, will make chips custom-built for inference that are easy for the world to access. We have seen this happen before with solar panels and electric cars. In those areas, Chinese companies quickly dominated global markets once production scaled up. There is no reason to think chips and memory will be any different.
Looking ahead, it seems highly unlikely that AI systems will be monolithic models trained by a handful of companies. Instead, we’ll see a federation of fine-tuned models from around the world, each checked and run by its own user community. Users could then compare these answers side-by-side. This future follows from the simple fact that truth arises from facts, and that the resolution of contradictions has always been the engine of historical and material development.
At a deeper level, biases in these systems point to a deeper truth about AI being an inherently social technology. Like older ones such as bureaucracy, markets, and democracy, it is a tool for processing information on a large scale, using heuristics to extract regularities from a messy world. Such systems are lossy by definition since the process of making rough categories necessarily involves some loss of information. This is inherent in the design of statistical models such as LLMs, which aim for good average performance rather than handling contextual subtlety. And that is precisely what makes them fundamentally unsuited to replace human judgment. When we add intentional biases to these already lossy systems, we have to ask a hard question of who wins and who loses? The MIT paper answers this question very clearly. It’s the most vulnerable people, those with less money and power, who pay the price.
And that is precisely what makes human judgment so essential. The incident of the condescending AI must be made a text in every course on digital literacy. People must understand that these systems are not neutral; rather, they embody the cultures and power structures they have emerged from. Checking the weights of models is a civic skill. Regular people should learn how to examine training data and fine-tuning choices. We should demand open-source AI with well-documented behaviors instead of mysterious black boxes. In politics, the demand for algorithmic transparency should be as basic as the demand for free speech. Mix the open-source ethic with anti-colonial thought, and you get a new kind of digital sovereign citizen.
The greatest surprise of the future may be that universities will return as the recognized arbiters of truth. AI-generated output, in its multiplicity and variety, will create an unprecedented need for the interpretative services of academics, journalists, and community researchers. These are the people who can digest, correlate, and explain what the machines cough up. Human expertise, far from being dispensed by this technology, remains as an essential interpreter and codifier, presiding over a sea of machine-generated perspectives. The humanities and social sciences may yet enjoy a renaissance because the most valuable skill will no longer be coding but hermeneutics. What we really need going forward is the ability to interpret and reconcile conflicting accounts to understand our shared reality within the context of our different interpretative frameworks.