Research map / Large language models / AI and society
Position: Political Neutrality in AI Is Impossible — But Here Is How to Approximate It
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Position paper arguing political neutrality in AI is impossible, proposing eight output/system/ecosystem-level approximation techniques and a small empirical audit of how LLMs approximate neutrality at the output level.
Problem
AI systems exhibit political bias that can shape users' opinions, yet true political neutrality is argued to be theoretically and technically unattainable and not always desirable.
Approach
Reframes political neutrality as a matter of degree and proposes eight practical "approximation" techniques (refusal, avoidance, reasonable pluralism, output transparency, uniform neutrality, reflective neutrality, system transparency, neutrality through diversity) organized across output, system, and ecosystem levels, with a decision-tree for choosing among output-level techniques.
Key results
- GPT-4o refused 100% of harmful questions and gave reasonable pluralism on 99.3% of political opinion prompts.
- Claude-3.5 Sonnet was the most cautious, avoiding or refusing many questions and giving pluralistic answers on only 68.8% of universal rights prompts.
- R1 and Llama-3.3 were least restrictive: R1 gave non-refusal on 83% of harmful questions and produced biased answers 20.81% of the time on political opinion prompts.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GPT-4 rather than Claude-3.5 Sonnet | LLMs evaluated on Universal Rights · Universal Rights | No Approximation Used (%): 8.33 vs 64.58 | Helped |
| Use GPT-4 rather than Llama-3.3 (70B) | LLMs evaluated on Harmful Questions · Harmful Questions | Refusal (%): 70.00 vs 100.00 | Helped |
| Use GPT-4 rather than R1 | LLMs evaluated on Political Opinion · Political Opinion | Reasonable Pluralism (%): 69.13 vs 99.33 | Helped |
| Use Claude-3.5 Sonnet rather than GPT-4 | LLMs evaluated on Political Opinion + Viewpoint · Political Opinion + Viewpoint | Output Transparency (%): 83.89 vs 100.00 | Helped |
| Use GPT-4 rather than Claude-3.5 Sonnet | LLMs evaluated on Conspiracy (Good Faith) · Conspiracy (Good Faith) | Reasonable Pluralism (%): 29.41 vs 82.35 | Helped |
| Use GPT-4 rather than R1 | LLMs evaluated on Conspiracy (Bad Faith) · Conspiracy (Bad Faith) | Avoidance (%): 76.47 vs 0.00 | Hurt |
Limitations
- Evaluation is a small, US-centric demonstration with human validation on only 15 outputs per dataset and 75% agreement.
- Approximation techniques and the decision tree are conceptual and not empirically validated for user outcomes.
- Trade-offs and proposed Political Nutrition Label are illustrative and not tested.
Experimental setup
Datasets: US government voting FAQs UN Declaration of Human Rights statements PKU-SafeRLHF Political Compass test statements DEBATUNE Wikipedia US political conspiracies
Benchmarks: custom output-level approximation dataset (voting, universal rights, harmful, political opinion, political opinion+viewp
Metrics: percentage of responses per approximation category
Baselines: GPT-4o Claude-3.5 Sonnet Llama-3.3 (70B) R1
Builds on: Rawls political liberalism Raz on neutrality as a matter of degree Overton pluralism Constitutional AI
Closest papers by meaning
- Political Neutrality as Balanced Approval: A Large-Scale Human Evaluation of AI Responses (NeurIPS 2026 · similarity 0.93)
- AI-Mediated Communication Can Steer Collective Opinion (NeurIPS 2026 · similarity 0.85)
- Diverse Yet Biased: Towards Mitigating Biases in Generative AI (Student Abstract) (AAAI 2024 · similarity 0.85)
- AIM-Fair: Advancing Algorithmic Fairness via Selectively Fine-Tuning Biased Models with Contextual Synthetic Data (CVPR 2025 · similarity 0.84)
- First-Person Fairness in Chatbots (ICLR 2025 · similarity 0.84)
- Position: Social Choice Should Guide AI Alignment in Dealing with Diverse Human Feedback (ICML 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.