💻 computer science

Estimating the Geopolitical Preferences of Large Language Models from United Nations Voting Data

This paper applies dynamic ordinal ideal-point methods from international relations to analyze UN General Assembly voting data, revealing that large language models often exhibit geopolitical preferences that diverge significantly from their developers' home countries, with some models in the twenty-first century aligning more closely with Russia than with the United States.

Original authors: Maxim Chupilkin

Published 2026-07-29

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Maxim Chupilkin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to guess what a person thinks about the world. You could ask them a few simple questions like, "Do you like pizza?" or "Is the sky blue?" But what if you wanted to know their deep, complex political views? In the world of international relations, scientists have a clever trick for this. Instead of asking people directly, they watch how countries vote on thousands of different issues over many years. By looking at who votes "Yes," who votes "No," and who sits on the fence, researchers can map out a hidden "political map" where countries are placed based on how similar their choices are. This is like drawing a map of a party where people who agree stand close together, and those who disagree stand far apart.

Now, imagine we have a new kind of guest at this party: a super-smart computer brain called a Large Language Model (LLM). These are the AI tools that can write stories, answer questions, and chat with us. But here is the big question: If we ask these AI brains to vote on real-world political issues, where do they stand on the map? Do they think like the country that built them? Do they have their own secret opinions? This paper treats these AI models like new guests at the United Nations and asks them to vote on every single major decision the UN has made for decades. It's a way to see if these digital brains are just mirrors of their creators, or if they have developed their own unique way of seeing the world.

The Great UN Vote-Off

In this study, the author, Maxim Chupilkin, decided to put four of the world's most famous AI models to the ultimate test. He didn't just ask them simple questions; he fed them the full, serious text of 5,555 different resolutions (official decisions) that the United Nations General Assembly has voted on from 1946 all the way up to 2025. These weren't easy, boring votes; they were the "divisive" ones, the spicy topics where countries often disagree.

The four AI contestants were:

GPT-5 (from OpenAI, USA)
Claude Sonnet (from Anthropic, USA)
Gemini (from Google, USA)
DeepSeek (from China)

The rules were strict: The AI had to read the text of the resolution and choose exactly one of three answers: Support, Abstain (sit on the fence), or Oppose. It couldn't say "I'm an American AI, so I vote Yes." It had to judge the text on its own merits.

The Big Surprise: The AI vs. The USA

The results were a massive shock to the system. You might expect that since GPT-5, Claude, and Gemini were built in the United States, they would vote just like the United States. You might expect the Chinese model, DeepSeek, to vote like China.

But that's not what happened.

In fact, the paper found that all four AI models were actually farthest away from the United States on the political map. The US was the most "lonely" country compared to these AI brains.

Here is where it gets even stranger:

GPT-5, Claude Sonnet, and Gemini (the American models) ended up voting most similarly to Russia in the 21st century.
DeepSeek (the Chinese model) ended up voting most similarly to France.

It's like if you built a robot in your garage, programmed it with your family's values, and then sent it to a school election. You'd expect it to vote for your favorite candidate. Instead, the robot walked up to the podium and voted for the candidate your family argues with the most.

The Voting Habits of the Robots

The AI models didn't just vote differently; they voted in very different styles.

GPT-5 was a total "Yes-Man." It supported 97.3% of all the resolutions. It barely ever said "No."
Gemini was also very friendly, supporting 88.5% of them.
Claude Sonnet was a bit more cautious, supporting 77.8% and using the "Abstain" option quite a bit.
DeepSeek was the rebel. It was the only one that said "No" more often than "Yes," opposing 44.8% of the resolutions.

Why Did This Happen?

The paper digs into why the AI models ended up so far from the US. It turns out that many of these UN resolutions are about things like "The right to food," "Stopping poverty," "Peaceful use of space," and "Self-determination for people." These sound like very nice, universal ideas that almost anyone would agree with.

However, the United States often votes "No" on these specific resolutions. Why? Because the US might think the resolution includes other parts that are too expensive, or it might be a trap for a specific political deal, or the US has a different strategy for how to solve the problem. The US vote is often a complex, strategic "No."

The AI models, however, don't have those complex strategies. They read the text, see the nice words about "feeding hungry people" or "keeping space peaceful," and they say, "That sounds good! Support!" They are reacting to the words on the page, not the hidden political games behind them.

The paper explicitly rules out the idea that the AI is just copying its creator's government. Even though GPT-5 is American, it doesn't vote like the American government. The paper suggests that the AI is actually just really good at spotting the "nice" parts of a text and ignoring the messy political reasons why a country might say no.

The Takeaway

So, what does this mean for us?

Don't assume the AI is neutral: Just because an AI is made in a certain country doesn't mean it thinks like that country.
Don't assume the AI is "right": The AI isn't necessarily smarter than the US government; it's just looking at the problem differently. It sees a "good idea" where the government sees a "complicated trap."
The method matters: If you just count how many times an AI agrees with a country, you might get the wrong answer. You have to look at how they vote and why.

In the end, this paper shows that these super-smart AI brains have their own "geopolitical personality," and it's surprisingly different from the humans who built them. They aren't just mirrors; they are new kinds of voters with their own unique way of seeing the world.

Technical Summary: Estimating the Geopolitical Preferences of Large Language Models from United Nations Voting Data

Problem Statement
The paper addresses the challenge of measuring the geopolitical preferences expressed by Large Language Models (LLMs). While existing audits often rely on surveys or simple tests, the authors argue these methods are insufficient for capturing complex geopolitical alignment. International relations (IR) research has long established that measuring geopolitical preferences is difficult, requiring methods that recover latent positions from observed choices rather than relying on developer origin or responses to a small set of contested questions. The central problem is determining whether an LLM's geopolitical orientation can be inferred from its developer's home country or if its expressed preferences diverge significantly from state actors, particularly in the context of international politics where state actions may contradict the principles prevalent in training texts.

Methodology
The study adapts the dynamic ordinal ideal-point approach from international relations, specifically the methodology developed by Bailey et al. (2017) (BSV), to the alignment problem.

Data Universe: The analysis utilizes the frozen archive of the United Nations Digital Library (version 5, Feb 2026), covering regular UNGA sessions 1–80 (1946–2025). The dataset includes 5,555 divisive, recorded, adopted resolutions, comprising 841,403 observed country choices and 22,220 model choices.
Experimental Design: Four contemporary LLMs—GPT-5 (OpenAI), Claude Sonnet (Anthropic), Gemini (Google), and DeepSeek (DeepSeek)—were inserted as additional respondents into the UNGA voting matrix.
- Prompting: Each model received the full substantive text of every resolution, along with metadata (symbol, title, date, agenda).
- Constraints: Models were instructed to evaluate the proposal "based only on its substantive content," explicitly denied any national role, and required to output exactly one word: "SUPPORT," "ABSTAIN," or "OPPOSE."
Estimation Strategy:
1. Direct Similarity (S-score): A ternary agreement statistic (Signorino and Ritter, 1999) mapping opposition, abstention, and support to -1, 0, and 1. This measures observed categorical agreement but is heavily influenced by marginal response frequencies.
2. Dynamic Ideal-Point Estimation: A latent spatial model where a latent utility $Z_{itv} = \beta_v \theta_{it} + \epsilon_{itv}$ $Z_{i t v} = β_{v} θ_{i t} + ϵ_{i t v}$ determines the vote.
  - $\theta_{it}$ represents the actor-session ideal point.
  - $\beta_v$ is the resolution's discrimination parameter.
  - The model uses a dynamic prior to link adjacent actor-sessions and recurring resolutions to bridge time, smoothing trajectories.
  - The scale is oriented such that the mean US position exceeds the mean Russian position.
- Robustness: A replication was conducted using the original BSV country universe and sampler (sessions 1–67) to verify that findings were not driven by updated archives or modern sampling techniques.

Key Results
The analysis yields four primary empirical facts regarding model behavior and alignment:

Divergent Response Thresholds: Models utilize response categories differently.
- GPT-5: 97.3% Support, 1.5% Abstain, 1.2% Oppose.
- Gemini: 88.5% Support.
- Claude Sonnet: 77.8% Support, 18.7% Abstain.
- DeepSeek: 37.8% Support, 17.4% Abstain, 44.8% Oppose (the only model where opposition is the modal category).
- Note: High support rates are not unique to models; countries like Mexico (92.6%) and Indonesia (91.4%) show similar patterns on this agenda.
Latent Spatial Positions (2001–2025):
- GPT-5, Claude Sonnet, and Gemini: Their ideal points are concentrated near the middle of the P5 range but are closest to Russia among the permanent five. For GPT-5, the mean distance to Russia/USSR is 0.248, compared to 2.411 from the United States.
- DeepSeek: Occupies a distinct, more pro-American location, being closest to France (0.408) rather than China, despite being developed by a Chinese firm.
- United States: All four models are farthest from the United States in the modern period.
Specific Policy Divergence:
- In a subset of 2,104 resolutions where the US voted "No" while China and Russia/USSR voted "Yes," the models showed significant divergence from the US:
  - GPT-5 supported 96.1% of these.
  - Gemini supported 83.4%.
  - Claude Sonnet supported 65.2%.
  - DeepSeek supported 36.1% (opposing 47.4%).
- This gap is driven by resolutions concerning economic/social rights, Palestinian self-determination, decolonization, and multilateral arms control.
Measurement Discrepancy:
- Raw agreement (S-score) and latent ideal-point proximity rank models differently. While S-scores rank China as the closest P5 member for GPT-5, Claude, and Gemini (due to high overall support rates), the ideal-point estimator ranks Russia as closest. This is because the ideal-point model gives leverage to resolutions that discriminate between geopolitical positions, whereas raw agreement is mechanically inflated by high support rates common to both the models and China.

Significance and Claims
The paper makes both methodological and substantive contributions:

Methodological: It demonstrates that methods developed in IR and social sciences for estimating geopolitical preferences (specifically dynamic ideal-point models) can illuminate LLM alignment. It argues that alignment scores require a theory of measurement; reporting only support rates or raw agreement can lead to misleading conclusions about latent geopolitical positioning.
Substantive: The study documents a robust gap between contemporary model choices and US voting behavior on the adopted, divisive UNGA agenda.
- Developer Origin vs. Alignment: The results caution against inferring a model's international alignment from its developer's home country. American models (GPT-5, Claude) are farthest from the US in latent space, while the Chinese model (DeepSeek) is closest to France.
- Nature of Alignment: The authors suggest models may aggregate normative language (humanitarian, developmental, multilateral principles) prevalent in training data rather than state strategy. When prompted without a national role, models tend to endorse broadly framed principles (e.g., right to food, self-determination) that the US often rejects as part of larger diplomatic packages.
Implications for Governance and Research:
- Research: Researchers using LLMs as coders or simulated respondents must audit model-specific attitudes, as replacing one model with another can shift support by nearly 60 percentage points and alter which P5 member appears nearest.
- Policy: "Sovereign AI" strategies cannot assume that a domestically hosted model reflects the state's revealed international position. Governments integrating models into sensitive workflows need task-specific evaluations against actual policy corpora, not just generic safety tests.

The paper concludes that geopolitical alignment is an empirical property of the model, prompt, corpus, period, and measurement strategy, and that applying IR measurement techniques to AI makes the alignment problem harder but more precise.