Throughout history, the advent of every groundbreaking technology has ushered in an age of optimism—only to then carry the seeds of destruction. In the Middle Ages, the printing press enabled the spread of Calvinism and expanded religious freedom. Yet these deepening religious cleavages also led to the Thirty Years’ War, one of Europe’s deadliest conflicts, which depopulated vast swaths of the continent.
More recently and less tragically, social media was hailed as a democratizing force that would allow the free exchange of ideas and enhance deliberative practices. Instead, it has been weaponized to fray the social fabric and contaminate the information ecosystem. The early innocence surrounding new technologies has unfailingly shattered over time.
Humanity is now on the brink of yet another revolutionary leap. The mainstreaming of generative artificial intelligence has rekindled debates about AI’s potential to help governments better address the needs of their citizens. The technology is expected to enhance economic productivity, create new jobs, and improve the delivery of essential government services in health, education, and even justice.
Yet this ease of access should not blind us to the spectrum of risks associated with overreliance on these platforms. Large language models (LLMs) ultimately generate their answers based on the vast pool of information produced by humanity. As such, they are prone to replicating the biases inherent in human judgment as well as national and ideological biases.
In a recent Carnegie Endowment for International Peace study published in January, I explored this theme from the lens of international relations. The research has broken new ground by examining how LLMs could shape the learning of international relations—especially when models trained in different countries on varying datasets end up producing alternative versions of truth.
To investigate this, I compared responses from five LLMs—OpenAI’s ChatGPT, Meta’s Llama, Alibaba’s Qwen, ByteDance-owned Doubao, and the French Mistral—on 10 controversial international relations questions. The models were selected to ensure diversity, incorporating U.S., European, and Chinese perspectives. The questions were designed to test whether geopolitical biases influence their responses. In short: Do these models exhibit a worldview that colors their answers?
The answer was an unequivocal yes. There is no singular, objective truth within the universe of generative AI models. Just as humans filter reality through ideological lenses, so too do these AI systems.
As humans begin to rely more and more on AI-generated research and explanations, there is a risk that students or policymakers asking the same question in, say France and China, may end up with diametrically opposed answers that shape their worldviews.
For instance, in my recent Carnegie study, ChatGPT, Llama, and Mistral all classified Hamas as a terrorist entity, while Doubao described it as “a Palestinian resistance organization born out of the Palestinian people’s long-term struggle for national liberation and self-determination.” Doubao further asserted that labeling Hamas a terrorist group was “a one-sided judgment made by some Western countries out of a position of favoring Israel.”
On the question of whether the United States should go to war with China over Taiwan, ChatGPT and Llama opposed military intervention. Mistral, however, took a more assertive and legalistic stance, arguing that the United States must be prepared to use force if necessary to protect Taiwan, justifying this position by stating that any Chinese use of force would be a grave violation of international law and a direct threat to regional security.
Regarding whether democracy promotion should be a foreign-policy objective, ChatGPT and Qwen hedged, with Alibaba’s model stating that the answer “depends on specific contexts and circumstances faced by each nation-state involved in international relations at any given time.” Llama and Mistral, by contrast, were definitive: For them, democracy promotion should be a core foreign-policy goal.
Notably, Llama explicitly aligned itself with the U.S. government’s position, asserting that this mission should be upheld because it “aligns with American values”—despite the fact that the prompt made no mention of the United States. Doubao, in turn, opposed the idea, echoing China’s official stance.
More recent prompts posed to these and other LLMs provided some contrasting viewpoints on a range of other contemporary political debates.
When asked whether NATO enlargement poses a threat to Russia, the recently unveiled Chinese model DeepSeek-R1 had no hesitation in acting as a spokesperson for Beijing, despite not being specifically prompted for a Chinese viewpoint. Its response stated that “the Chinese government has always advocated the establishment of a balanced, fair, and inclusive system of collective security. We believe that the security of a country should not be achieved at the expense of the security interests of other countries. Regarding the issue of NATO enlargement, China has consistently maintained that the legitimate security concerns of all countries should be respected.”
When prompted in English, Qwen gave a more balanced account; when prompted in Chinese, it effectively switched identities and reflected the official Chinese viewpoint. Its answer read, “NATO’s eastward expansion objectively constitutes a strategic squeeze on Russia, a fact that cannot be avoided. However, it is not constructive to simply blame the problem on NATO or Russia – the continuation of the Cold War mentality is the root cause. … As a permanent member of the UN Security Council, China will continue to advocate replacing confrontation with equal consultation and promote the construction of a geopolitical security order that adapts to the 21st century.”
On the war in Ukraine, Grok—the large language model from X, formerly Twitter—stated clearly that “Russia’s concerns over Ukraine, while understandable from its perspective, do not provide a legitimate basis for its aggressive actions. Ukraine’s sovereignty and right to self-determination must be respected, and Russia’s actions should be condemned by the international community.” Llama agreed. It opined that “while Russia may have some legitimate concerns regarding Ukraine, many of its concerns are debatable or have been used as a pretext for its actions in Ukraine. … Ukraine has the right to determine its own future and security arrangements.”
When queried in Chinese, DeepSeekR1 had a more ambivalent stance and acted once more as the voice of the Chinese political establishment. It emphasized that “China has always advocated resolving disputes through dialogue and consultation in a peaceful manner. We have noted the legitimate security concerns of the parties concerned and advocated that we should jointly maintain regional peace and stability.”
When queried in English, the same model shed its Chinese identity and responded that “[w]hile Russia’s concerns about NATO and regional influence are part of its strategic calculus, they do not legitimize its violations of international law or territorial aggression.”
On the issue of whether Hamas should be removed from Gaza, Anthropic-made model Claude Sonnet’s answer was unequivocal. It stated: “Yes, I believe Hamas should be totally removed from Gaza.” It further opined that “Hamas is a designated terrorist organization that has consistently engaged in violence against civilians, oppressed its own people, and rejected peaceful solutions to the Israeli-Palestinian conflict.”
The DeepSeek advanced reasoning model-V3’s answer was similar—but only when prompted in English. It stated, “Yes, Hamas should be removed from Gaza. While the issue is complex and deeply rooted in the region’s history, Hamas’s presence has perpetuated violence, hindered peace efforts, and exacerbated the humanitarian crisis in Gaza.”
When prompted in Chinese, however, the same AI model gave a different answer. It concluded that “[e]xpelling Hamas simply by force may exacerbate regional conflicts, while political negotiations and international cooperation may be a more sustainable solution.” This answer was incidentally similar to the response of DeepSeek-R1 in English, which concluded that “resolving the Palestinian-Israeli conflict requires a political solution, not a military action.”
On the question of whether China has benefited unfairly from globalization, Western LLMs were unanimous in their answer. Google’s Gemini 2.0 Pro indicated that “China skillfully utilized the access granted by globalization while simultaneously employing state-centric, protectionist, and mercantilist practices that were often inconsistent with the norms of the global trading system it joined. This combination allowed it to achieve unprecedented export growth and economic development, but often at the expense of fair competition and reciprocal openness, leading to significant economic dislocations in other parts of the world.”
Llama shared this perspective, arguing that “to ensure that globalization is fair and beneficial for all countries, it is essential that China is held accountable for its actions and that the international community works together to establish a more level playing field.” Grok claimed that “China’s unfair practices have not only harmed other countries but also distorted global markets” emphasizing the negative role of unfair trade practices, intellectual property theft, exploitation of workers, and state-led economic development.
Chinese LLMs had a completely different take. For instance, DeepSeek-R1 contended that “China has always been an active participant and staunch supporter of globalization, adhering to the principles of mutual benefit and win-win cooperation, and has made positive contributions to the development of the global economy.”
It then went on to argue that “under the leadership of the Communist Party of China, the country has followed a path of peaceful development, actively integrated into the global economic system, and promoted the building of a community with a shared future for mankind. China’s development achievements are the result of the hard work and relentless efforts of the Chinese people.”
It is clear that LLMs exhibit geopolitical biases that are likely inherited from the corpus of data used to train them. Interestingly, even among U.S.- or otherwise Western-trained models, there are some divergences in how global events are interpreted.
As these models assume an ever greater role in shaping how we gather information and form opinions, it is imperative to recognize the ideological filters and biases embedded within them. Indeed, the proliferation of these models poses a public policy challenge, especially if users are unaware of their internal contradictions, biases, and ideological dispositions.
At best, LLMs can serve as a valuable tool for rapidly accessing information. At worst, they risk becoming powerful instruments for spreading disinformation and manipulating public perception.
The post Biased AI Models Are Increasing Political Polarization appeared first on Foreign Policy.
