Toward Political Neutrality in AI

Key Takeaways

• True political neutrality in AI is impossible, but there are practical approximations of neutrality that developers can implement across different levels of AI systems.

• We developed a framework for systematically evaluating political neutrality approximation techniques. The framework includes eight techniques across the output, system, and ecosystem level of AI systems.

• More evaluations of political neutrality approximations are needed. Future research should focus on which approximations are currently used, which ones are feasible, and which ones are valued by users.

• As a starting point for evaluation, we created a dataset to evaluate the output-level approximation techniques used by 10 current AI models. We find that, overall, open-source models exhibit more political bias and engage more readily with harmful content.

• Policymakers and AI developers must shape AI systems that respect wide-ranging viewpoints while promoting fairness, user autonomy, and trust. To do so, they need to encourage transparency and interdisciplinary research on political neutrality approximations.

Executive Summary

Leading generative AI models have been reported to show political bias in individual instances – such as XAI’s Grok identifying as “MechaHitler” or Google’s Gemini depicting female popes — and in systematic ways.

Political bias is concerning because it is widespread and can influence users’ opinions and decisions. Recent research shows that AI-generated messages can influence people’s attitudes toward controversial issues such as gun control and climate action, and affect political decisions such as budget allocations. Politically biased AI systems may hinder people from independently forming opinions and making choices, a key pillar in liberal democracy. A commonly proposed solution to this challenge is making AI models politically neutral. However, true political neutrality in AI — meaning systems that are impartial and don’t favor particular political viewpoints — is theoretically and practically impossible.

In our paper “Political Neutrality in AI Is Impossible — But Here Is How to Approximate It,” we explain why this is the case and propose practical approximations of political neutrality that can reduce political bias and move us closer to achieving neutrality. We also test how today’s AI models respond to political content, and show how our framework can help evaluate and improve future language models.

Our work is a first step toward shifting the conversation on political bias away from impossible objectives and toward achievable approximations of political neutrality. These approximations allow AI developers to create systems that respect wide-ranging viewpoints while promoting fairness and user autonomy.

Why True Neutrality Is Impossible

Theoretically speaking, political neutrality is impossible. Neutrality is inherently subjective when what seems neutral to one person might seem biased to someone else. On the political spectrum, there is no neutral point, as moderate opinions that lie between left-leaning and right-leaning views are political positions in and of themselves. Evaluating political neutrality by assessing the intent or impact of an action is challenging as both are hard to measure.

Some argue that political neutrality is also currently technically impossible. In designing AI systems, humans make countless decisions about which data to use or how the system should respond — each of which can introduce biases. Even the information that AI models learn from, like the training data scraped from the internet or user inputs, often reflect existing biases. As a result, it is impossible to build an AI model without biased human input.

However, inspired by philosopher Joseph Raz, who observed that “neutrality [...] can be a matter of degree,” we argue that approximating political neutrality is not only possible but essential for promoting balanced AI interactions and mitigating user manipulation. Drawing on insights from related fields like sociology, political science, and philosophy, which have historically grappled with neutrality, bias, and representation, we developed a framework for approximating political neutrality in AI systems.