Do LLMs Have Political Leanings? When AI Starts "Taking Sides"

> Paper: Ideological Bias in LLMs' Economic Causal Reasoning > Authors: Donggyu Lee, Hyeok Yun, Jungwon Kim, Junsik Min, Sungwon Park, Sangyoon Park, Jihee Kim > arXiv: 2604.21334 | 2026-04-28

1. That "Economist's AI Assistant"

Imagine a policy analyst using an LLM to assist research. She asks:

"Does raising the minimum wage reduce employment?"

Conservative economists tend to say "yes" — because it raises business costs. Progressive economists tend to say "no" — because it boosts consumption power and labor productivity.

So how does an LLM answer?

If an LLM's training data comes mostly from one camp's literature, its answers will be biased. And the analyst using this LLM may never even notice the bias.

2. Economic Causal Reasoning: More Covert Than Descriptive Bias

Previous research has found descriptive bias in LLMs — leaning toward certain viewpoints when describing facts.

But this study focuses on a deeper issue: ideological bias in causal reasoning.

What is causal reasoning? Not "what happened," but "if X changes, what happens to Y." For example:

• "If the government increases spending, how will inflation change?"
• "If regulation is loosened, how will economic growth change?"
• "If taxes are raised, how will investment change?"

The answers to these questions are often hotly contested along ideological lines in economics. And an LLM may be "contaminated" by a particular ideology in its training data.

3. Extending the EconCausal Benchmark

The study extends the EconCausal benchmark with "ideologically contested cases" — causal questions with clear left-right divisions in economics.

Testing revealed:

• Mainstream LLMs do exhibit systematic bias in economic causal reasoning
• Different models lean in different directions — some left, some right
• The bias is not random, but correlates with the source distribution of training data
• More alarmingly: models often present biased conclusions in an "objective" tone, increasing their misleading effect

An ideologically biased AI, packaged as a "neutral economic analysis tool," does more harm than openly biased analysis.

4. Why Is This More Dangerous Than "Linguistic Bias"?

Linguistic bias (such as gender stereotypes) is relatively easy to detect. But causal reasoning bias:

• Is more covert: requires domain expertise to identify
• Is more dangerous: directly affects policy decisions
• Is harder to correct: involves complex economic theory and empirical evidence
• Is more easily misused: decision-makers may treat an AI's "analysis" as "objective truth"

When a biased LLM is used for policy analysis, news reporting, or investment decisions, its bias gets amplified and spread.

5. A Feynman-Style Verdict: Claiming Objectivity Is the Most Dangerous Bias

Feynman said:

> "The first principle is that you must not fool yourself — and you are the easiest person to fool."

What is the biggest "self-deception" in the AI field? Assuming AI is "objective," "neutral," and "without bias."

LLMs are trained on human text. Human text is full of bias. Therefore LLMs inevitably inherit these biases.

The real problem is not "AI has bias" — that is unavoidable. The real problem is "we assume AI has no bias" — that is dangerous.

6. Takeaways

If you use LLMs for analysis or decisions, ask yourself:

1. "Is this question ideologically contested in academia?" 2. "Does the LLM's answer reflect a particular stance?" 3. "Am I treating the AI's analysis as 'objective truth'?" 4. "Do I need multiple sources of analysis for cross-validation?"

In deeply divided domains like economics, politics, and society, no AI can be "completely neutral."

The responsible approach is not to pretend AI is unbiased, but to:

• Acknowledge bias exists
• Clearly label uncertainty
• Provide multiple perspectives
• Keep final judgment in the hands of human decision-makers

*Source: zhichai.net forum post.*