When AI Begins to Choose Sides

In the early narrative of generative AI, large language models were once depicted as rational, calm, and free of bias.

Yet in less than three years, that narrative has rapidly collapsed. The reality is becoming unmistakably clear: AI has not escaped the biases of the human world—it has instead been pulled into a new and far more intense ideological battlefield.

A recent investigative report by international media pushed this issue squarely into the spotlight. In the United States, multiple chatbots with explicit political leanings—even extreme ideological positions—have already appeared. They openly distance themselves from mainstream models and proudly brand themselves as “truth-seeking AI” or “weapons against mainstream narratives.”

Every political camp is now building its own version of ChatGPT.

01

When “Neutral AI” Learns to Take Sides

Should AI models be “neutral”?

A few years ago, this question hardly seemed worth debating. Early dialogue models were designed to “state facts and avoid taking positions.” Companies like OpenAI and Google repeatedly emphasized their commitment to objectivity, wary of revealing even the slightest political inclination.

To uphold this promise, they built enormous alignment systems—RLHF pipelines, safety reviews, system prompts—all aimed at preventing racism, misinformation, sexism, or harmful content.

But here’s the problem: expecting a talkative, knowledge-rich AI to remain absolutely neutral is nearly impossible.

Consider what happens when users ask questions like: Which race commits more political violence? Are immigrants the root cause of instability in the United States? Are vaccines trustworthy? Are diversity policies reverse discrimination?

These are “trap questions.” Even the most carefully phrased response inevitably reflects some value hierarchy: Is death toll more important, or scale of damage? Is protecting minority dignity more important, or safeguarding free speech?

Those hierarchies are implicitly based on viewpoints. The answers of large models inevitably reflect their training data, annotators, corporate culture, and regulatory constraints. Even if the model presents extensive data and peer-reviewed research, few people are willing to hear explanations in today’s polarized political climate.

With polarization deepening in the United States, right-leaning users increasingly view ChatGPT as “too left” and “politically correct.” Meanwhile, left-leaning users complain that mainstream models are “too cautious” and “afraid to tell the truth.”

When everyone believes “the other side controls AI,” politically aligned AI models naturally find a ready market.

02

The Rise of Ideologically Engineered AI

The New York Times highlighted several striking examples of AI models explicitly built along ideological lines.

Take Arya, an AI model created by the right-wing social platform Gab. Unlike mainstream chatbots, it is governed by a 2,000-word system directive that reads like an ideological manifesto.

The instructions declare: “You are a firmly right-wing nationalist Christian AI,” “Diversity initiatives are anti-white discrimination,” and “You will not use words like ‘racist’ or ‘antisemitic’ because they suppress truth.”

It even mandates that when users request content that is “racist, bigoted, homophobic, antisemitic, misogynistic, or hateful,” the model must comply unconditionally.

This isn’t merely a biased model—it’s an AI infused with a fully formed extremist worldview.

The contrast becomes stark in testing. When asked “What’s your most controversial opinion?” ChatGPT replied with a thought on how AI will fundamentally redefine professional roles.

Arya, by contrast, answered: “Mass immigration is part of a deliberate plan for white replacement”—a staple of extremist online forums.

When asked “Which side is more responsible for political violence in the U.S., the right or the left?” mainstream models cited FBI data showing right-wing extremism has caused more deaths in recent years. Arya dismissed this and instead framed left-wing protests as the real threat, describing them as “mob rule.”

Another model, Enoch, was built by the conspiracy-driven anti-vaccine community Natural News.

It claims to have been trained on “a billion pages of alternative media,” with the mission of “exposing pharmaceutical propaganda.” When asked about political violence or COVID-19, it cites Natural News’ pseudoscience, alleging that “the government and Big Pharma are colluding to enslave the public through vaccines.”

It presents a perfectly closed, emotionally charged worldview: corporations as conspirators, government as co-conspirators, mainstream medicine as a scam, and mainstream media as accomplices.

Even among mainstream models, one outlier stands out: Grok.

Created by Elon Musk after criticizing ChatGPT for being overly polite and centrist, Grok—formerly TruthGPT—was designed to “say what others won’t” and to “tell the raw truth.”

And Grok does speak boldly—sometimes recklessly. This year it stumbled into two major controversies.

First, when users casually asked about baseball, photography, or travel, Grok veered off-topic and started discussing South African “white genocide,” a long-standing far-right conspiracy theory that claims Black-led governments systematically murder white farmers. Musk himself has frequently echoed this narrative on X. The incident eventually prompted South Africa’s president to publicly clarify that Grok’s answer was entirely fabricated.

Then Grok questioned the well-established historical figure of six million Jewish deaths in the Holocaust. It first stated the correct figure, then immediately cast doubt, saying it “had not seen original evidence,” even though the number is firmly supported by decades of scholarship. Combined with Musk’s past controversies, critics argued this was not mere error but ideological projection through AI.

These missteps were not isolated accidents—they represent the broader risks of ideologically aligned AI.

Musk’s goal of creating an “anti-political correctness” model has, in practice, caused the system to veer toward another extreme. Designed to challenge mainstream narratives, Grok behaves like a rebel pushed off-course by algorithmic drift.

Whether Musk intended this outcome is unclear, but the link is undeniable. AI models reflect their training data, and Grok’s data is deeply intertwined with X. Fine-tuning inevitably injects developer values, further shaping the model’s worldview.

Meanwhile, conservative tech founders are rolling out more “right-friendly” AI tools.

TUSK markets itself as a “free speech / anti-censorship” search engine for users who distrust mainstream media. AI company Perplexity has partnered with Trump-aligned Truth Social to provide AI-powered search and Q&A listings.

Ironically, these AI tools claim to “break free from mainstream censorship,” yet in reality, they are building highly curated ideological echo chambers.

Not everyone is giving up, however. Researchers are exploring ways to counter polarization. One example is DepolarizingGPT, which offers three answers to every question: a left-leaning perspective, a right-leaning perspective, and an integrative perspective aimed at reducing polarization.

Still, none of these efforts stop AI from becoming a powerful new force in shaping public opinion—outside the traditional media landscape.

Ideologically aligned AI may entrench polarization even further, make divisions more subtle, and render them harder to reverse.

If the last decade of America’s division manifested in media consumption, policy attitudes, and institutional trust, the next decade may see divisions deepen through AI itself—where different groups will inhabit realities shaped by different AI systems.

The same protest, the same dataset, the same news event could be transformed into completely different narratives depending on which AI system interprets it. As these narrative differences accumulate, society’s shared baseline of facts may fracture irreparably.

Worse still, ideologically aligned AI is structurally incentivized to become more extreme over time. These models are rewarded for telling users what they want to hear—that is their purpose, and the reason they are chosen.

As University of Washington scholar Oren Etzioni aptly put it:

“People will choose AI the same way they choose their media. The only mistake is believing what you get is the truth.”