Swiss AI Voting Study Shows Preference for Mainstream Politics

Artificial Intelligence is not the radical disruptor we feared—it is a conformist. A groundbreaking study from the federal technology institute ETH Zurich exposes a startling reality: if AI systems like ChatGPT were given the right to vote, they would overwhelmingly back the establishment. Far from introducing chaotic new variables into the democratic equation, these large language models gravitate toward mainstream political parties and safe, streamlined decisions.

This revelation shatters the sci-fi illusion of a rogue AI upending society. Instead, the research suggests a future where AI reinforces the status quo with mechanical precision. The study explicitly concludes that AI models demonstrate "more uniform behavior" than human beings. While human voters are messy, emotional, and diverse, the algorithms are predictable and conservative in their choices. For Switzerland, a nation that prides itself on a multi-party system and direct democracy, this finding signals that AI may currently lack the nuance required to navigate the complexities of the Swiss political landscape.

Researchers bypassed the gridlock of US presidential politics to test AI on something far more tangible: the streets of Zurich. Because models like ChatGPT are programmed to evade polarizing questions about figures like Donald Trump or Kamala Harris, the Computational Social Science team at ETH Zurich deployed a clever workaround. They fed the algorithms 24 specific, local urban development proposals—ranging from a car-free Langstrasse to multicultural festivals on Sechseläutenplatz.

The experiment pitted the digital minds of ChatGPT4 and Llama 2 against 180 human participants. The stakes were local, but the implications were global. By asking the AI to improve Zurich for its citizens, the team forced the models to take a stance. The results offer a rare glimpse into the "political" preferences of the machine. Unlike the humans, who showed passion for a wide variety of projects, the AI's choices were clinical and detached, stripping away the chaotic vibrancy that defines local Swiss politics.

The data reveals a staggering lack of diversity in machine thinking. While the 180 human participants varied wildly in their enthusiasm—selecting different numbers of projects based on personal conviction—ChatGPT almost invariably selected four or five projects. This robotic consistency exposes a critical flaw: a "WEIRD" bias. The study confirms that these synthetic samples reflect a Western, Educated, Industrialized, Rich, and Democratic worldview, failing to capture the true variance of a human electorate.

This uniformity is dangerous for democratic representation. If AI cannot replicate the messy, divergent opinions of a real population, its utility in political modeling is severely compromised. The algorithms function like a hive mind, smoothing over the edges of public dissent and settling for a homogenized middle ground. In a direct democracy like Switzerland, where minority voices and regional differences are paramount, such a lack of variance isn't just a technical glitch—it's a democratic deficit.

Despite its conformist tendencies, the AI proved to be a superior accountant. The models consistently prioritized less expensive projects, displaying a fiscal responsibility often absent in human voters. The ETH study notes that while humans frequently lack cost awareness, the AI ruthlessly optimized for budget. However, this cold logic comes with a significant vulnerability: the AI is easily manipulated.

The researchers discovered that the models were swayed simply by the order of projects on the list. Imagine a voter changing their party allegiance solely because one name appeared at the top of the ballot and another at the bottom. This "primacy effect" undermines the AI's credibility as a sophisticated decision-maker. It suggests that while the machine can crunch numbers better than a human, it lacks the conviction to stand by its choices when the presentation changes. It is a budget hawk with a fragile will.

This research casts a long shadow over the futuristic concept of "digital twins"—the idea that every voter could have an AI counterpart to assist or even replace them in the voting booth. While economists and visionaries like César Hidalgo have proposed such systems to streamline democracy, the ETH findings suggest we are nowhere near ready.

If a digital twin merely votes for the mainstream and collapses under the pressure of ballot formatting, it serves no one. For Switzerland, the study offers a stark warning: technology can mimic the mechanics of voting, but it cannot yet replicate the soul of the voter. As we rush toward a digitized future, we must confront the reality that an AI electorate would be efficient, cost-effective, and profoundly boring—stripping our democracy of the vital conflict and diversity that makes it work.