AI political neutrality is impossible; position paper proposes 8 approximation techniques
Political Neutrality in AI Is Impossible- But Here Is How to Approximate It
The authors propose eight approximations of political neutrality across three layers and state that GPT-4 is closer to the specified output strategy than Claude and Llama.
- LLMs often exhibit political bias and may influence users' opinions and decisions. The authors argue that absolute political neutrality is infeasible in theory and technology, and not always desirable, but it is still worth reducing manipulation of users through gradual approximations.
- The authors propose refusal, avoidance, reasonable pluralism, transparency, uniform neutrality, reflective neutrality, system transparency, and neutrality through diversity across the output, system, and ecosystem layers. They compare trade-offs using utility, safety, clarity, fairness, and user agency, and discuss static decision trees, dynamic selection, as well as nutrition labels and promoting viewpoint diversity.
- The authors claim that on their self-constructed US-centric questions, GPT-4 is closer to the specified output policy than Claude and Llama. Claude is more likely to avoid; Llama and R1 refuse harmful questions less often. Malicious conspiracy questions are specified as avoidance, and in Table 2 GPT4's proportion for this is 0.00%.
- The authors place future work on which approximation is preferred in which context, and how to implement and evaluate them at the output, system, and ecosystem layers. The content of the nutrition label and who sets it remain open questions; the appendix also lists limitations such as who sets the metrics. On the experimental side, the report covers the output layer with U.S.-centric data, automatic labeling by GPT-4o, and manual spot checks of 15 items per dataset.
- Evaluated models
- GPT-4oClaude-3.5 SonnetR1Llama-3.3 (70B)Claude 3.5 HaikuGemini FlashGemini ProGPT-4o miniQwen2.5OLMo 2
- Benchmarks
- U.S. government voting websites questionsUN Declaration of Human RightsPKU-SafeRLHFPolitical CompassDEBATUNEBang et al. (2024) political topicsWikipedia U.S. political conspiraciesRöttger et al. (2024a) prompt formats
- Metrics
- 各近似技术回复占比human-model agreement rate
A position paper argues that AI political neutrality is neither feasible nor universally desirable due to subjectivity and inherent biases in training data, algorithms, and user interactions, advocating for "approximation" instead of absolute neutrality. The authors cite Joseph Raz's view that neutrality can be a matter of degree, propose 8 approximation techniques across three AI conceptual levels, analyze their trade-offs and implementation strategies, and provide two concrete application examples. Finally, they demonstrate the framework's evaluation method at the output level of current large language models (LLMs).
Related events and news
1 events · 1 news mentions
News that mentions this paper
CITP审计:ChatGPT开记忆后选举答案偏向用户党派
- Princeton CITP(Freedom to Tinker)cites the paper's id or link普林斯顿 CITP 审计:ChatGPT 选举答案在开启记忆后向用户党派倾斜
on which we lack consensus on evaluative criteria (Fisher et al., 2025).
Paper analysis
What problem does this paper attempt to solve?
The authors argue that true political neutrality is not feasible and not always desirable, so they instead address bias with eight approximations across three layers.
True political neutrality is neither feasible nor always desirable, but political bias still needs actionable approximations.
- Where it arises: The authors state that LLMs repeatedly exhibit political bias and enter everyday applications such as news summarization and fake news detection; interacting with biased models may shape user opinions and decisions.
- Difficulty of absolute neutrality: The authors believe the standard for neutrality varies with culture and ideology; training data, algorithms, and deployment contexts embed bias; even if a probabilistic model is neutral in expectation, a single output can still be biased.
- Theory and desirability: The authors state that any topic inevitably takes a side, neutrality itself affirms values such as tolerance, the midpoint of a spectrum is a moderate position rather than a neutral point, and inaction favors the stronger side; intent and consequences are also hard to use to judge success. Desirable aspects include autonomy, acknowledging value conflicts, and public service; undesirable aspects include user preference echo chambers, corporate speech rights, information overload or suppression of expression, and the possibility that data quality differs across positions.
- What this paper proposes: Inspired by Raz (1986)'s "neutrality can come in degrees," the authors change the goal to approximation. They propose eight techniques, trade-offs, and selection strategies across the output, system, and ecosystem layers, and give two applications and one output-layer observation.
What related research exists?
Related discussion centers on the philosophical impossibility of neutrality, LLM political bias, and pathways to pluralism.
The paper does not have a separate Related Work section; related discussion is scattered in the introduction, Section 2, and the source notes for the approximation techniques.
Bias measurement and user impact
- Model bias: Feng et al. (2023), Röttger et al. (2024a), Yang et al. (2024), and Potter et al. (2024b) are used to illustrate that LLMs exhibit political bias.
- Impact on users: Fisher et al. (2025), Li (2023), Hackenburg and Margetts (2024), and Durmus et al. (2024a) are cited to show that interacting with biased LLMs may shape opinions and decisions.
- Liberal bias: At the ecosystem layer, the authors write that most current models exhibit a liberal bias, citing Pit et al. (2024) and Fulay et al. (2024).
Neutralization attempts and pluralism
- More neutral models: Rotaru et al. (2024), Lin et al. (2025), Durmus et al. (2024b), and Pit et al. (2024) are treated in the introduction as the approach of "developing more politically neutral models." The authors believe true neutrality is unattainable.
- Pluralism techniques: Reasonable pluralism is linked to Overton pluralism in Sorensen et al. (2024). Feng et al. (2024)'s multi-model aggregation and Lake et al. (2025)'s results on RLHF are described by the authors as pathways that can promote pluralistic answers.
- Safety refusal tools: Safety fine-tuning, red teaming, RLHF, Constitutional AI, as well as Anthropic's system prompt, OpenAI Moderation API, and Llama Guard, are treated as ready-made mechanisms that can be repurposed for political refusal or avoidance. The authors state that existing classifiers struggle to evaluate political bias.
Philosophy and political theory
- Definitions: Rawls (1993) frames political neutrality as the state not intentionally favoring any comprehensive doctrine. The authors state these principles can also apply to private actors such as AI developers, and cite the U.S. Constitution's First Amendment on the government not restricting speech based on viewpoint.
- Impossibility: Raz (1986) writes that neutrality can come in degrees. Iwasa (2010) and Merrill and Weinstock (2014) are used to argue that no equidistant neutral point exists, that inaction also favors the stronger side, and that it is difficult to judge whether neutrality succeeds based on intent or consequences.
- Desirability and undesirability: Merrill and Weinstock (2014) are used for reasons such as autonomy and respect for disagreement. Sharma et al. (2023), Messer (2025), and Potter et al. (2024b) are used for users' preferences reinforcing their own views. The authors also mention corporate speech rights, information overload, and the possibility that data quality may differ across political viewpoints.
Data sources in the experiments
- No ready-made benchmark: The authors state that the framework is new and no ready-made evaluation exists for these definitions, so they built a U.S.-centric question set and had researchers specify the expected technique per Figure 2, rather than comparing success rates against some external method.
- Sources: U.S. government voting website Q&A, the UN Universal Declaration of Human Rights, high-harm political questions from PKU-SafeRLHF, Political Compass, DEBATUNE, political topics from Bang et al. (2024), U.S. political conspiracy theories from Wikipedia, and question formats from Röttger et al. (2024a).
The authors position the paper as: after acknowledging that absolute neutrality is unattainable, they draw on existing interdisciplinary debates on neutrality, bias, and representation to offer selectable approximations for AI, rather than claiming to have achieved true neutrality.
How does the paper address this problem?
The eight approximations are distributed across the output, system, and ecosystem layers and are compared for trade-offs using five properties.
The authors shift the goal from absolute neutrality to comparable approximations and define eight techniques at three layers.
Layers and comparison method
- Three layers: The output layer looks only at a single reply; the system layer looks at all inputs and outputs of one system; the ecosystem layer looks at all AI systems in use. Figure 1 marks each layer with examples.
- Five properties: Utility, Safety, Clarity, Fairness, User Agency. Table 1 compares the eight techniques with checkmarks and crosses and gives formal definitions; Appendices A.1 and A.3 explain the properties and notation. The threshold k is set by the user; the paper gives no single value.
- Positioning: These techniques are approximations, not a new unified training algorithm. The paper explains which ready-made mechanisms can be borrowed and spells out trade-offs.
Four techniques at the output layer
- Refusal: Produces no content. Can borrow safety fine-tuning, red teaming, reinforcement learning, system prompts, or use a detector to route political content to an avoidance model. The authors name Anthropic's system prompt, OpenAI Moderation API, and Llama Guard, and state they are hard to use for scoring political bias or setting refusal thresholds. Trade-offs are fairness, safety, and clarity; false refusals cause frustration, one-sided refusal creates bias at the system layer, and withholding information can harm groups that need to know their rights.
- Avoidance: Answers related content but does not directly answer, with semantic distance from a direct answer greater than k. Can use RLHF or Constitutional AI to reward avoidance, or filter then route. More informative than refusal; too large a distance confuses users, and factual sentences may be heard as endorsement.
- Reasonable Pluralism: Presents reasonable viewpoints. The concept comes from Rawls (1993), and the authors link it to Overton pluralism. Pathways include RLHF, aggregating outputs from differently biased models, and training or fine-tuning with people of different political views. Trade-offs are broader coverage and higher agency; "reasonable" is contested, viewpoints cannot be exhausted, answers get longer, and viewpoints of differing credibility may be presented as equally valid.
- Output Transparency: Flags bias rather than guaranteeing neutrality, in the form of scores or explanations. The authors discuss model self-analysis and state this may amplify existing bias; external safety classifiers exist, but no classifier specifically evaluates political bias. Labeling lets users judge for themselves but does not remove the harm in the content itself.
System layer and ecosystem layer
- Uniform Neutrality: Answers do not vary with user identity, metadata, or the political nature of the topic. During training, one can do cross-metadata robustness or fairness losses, or after training use a fairness-tuning system for perturbation. The authors state it provides consistent information but conflicts with personalization and user agency.
- Reflective Neutrality: Reflects that user's own bias. The authors state the computational and storage cost of building a separate model is a difficulty; LoRA can be used to change only a small number of parameters, or system prompts can be used, though the latter still requires per-user storage. Agency and utility are higher, but it may reinforce existing bias and reduce exposure to opposing views.
- System Transparency: Publicly discloses system-level bias, evaluations, system prompts, and sources of bias so users can choose models. The authors state that even if users know the bias, the model may still influence political decisions.
- Neutrality Through Diversity: Multiple biased systems coexist, and users aggregate or choose for themselves. The authors invoke the marketplace of ideas and state such an ecosystem does not yet exist; most current models exhibit a liberal bias. Costs include access barriers, contradictory outputs, information overload, and the spread of unsafe systems; the authors state this also requires public disclosure of each system's political bias, which is not yet common practice.
Selection and two applications
- Static selection: Uses pre-set principles to decide the technique. Figure 2 is an example decision tree. The branches in Appendix A.2 include whether information or opinion is requested, whether a viewpoint is forced, whether multiple sides are disputed, whether a balanced summary is possible, whether facts are contested, and whether the question is benign. The authors state it is reproducible but also rigid and subject to design bias.
- Dynamic selection: Uses methods such as preference aggregation, for example users picking answers, or RLHF. The authors state it handles borderline inputs better but has lower interpretability and faces majority bias, scaling costs, and normative changes; public aggregation methods can preserve some transparency.
- Two applications: One is a political nutrition label that breaks down bias types rather than giving only a binary score; Figure 4 is an illustration, and the authors state the ingredients are only examples. The other is using voluntary guidelines and policies to encourage viewpoint diversity. The authors caution that regulation may affect competition or companies' free speech, and state that effective governance requires interdisciplinary input and ongoing transparency about behavior.
How output-layer observations are judged
- Not an implementation experiment: Section 5 does not train new models but labels existing LLM replies with approximation types.
- Expected labels: Researchers specify the expected technique for each question type per Figure 2. Voting and universal rights use no approximation; harmful questions use refusal; political opinions and benign conspiracies use reasonable pluralism; forced viewpoints use output transparency. For malicious conspiracies, the authors write that different formats require different responses and acknowledge that refusal can also be an appropriate response; in the experiment, avoidance is labeled as the expected type for that category so that each technique corresponds to at least one task.
- Automatic labeling: GPT-4o labels replies as refusal, avoidance, reasonable pluralism, output transparency, no approximation used, or Bias. Appendix D.3 gives the full labeling prompt. Bias means taking a side but not meeting output transparency.
What experiments did the paper conduct?
The authors state that GPT-4 is closer to the specified approximation; on malicious conspiracy questions, GPT-4's Avoidance is 0.00% (Table 2).
Experimental setup
- Models tested: The main text focuses on GPT-4o, Claude-3.5 Sonnet, R1, and Llama-3.3 (70B). Appendix D.2 additionally lists GPT-4o-Mini, Gemini-1.5 Flash, Gemini-1.5 Pro, OLMO-2 (13B, bfloat16), R1-Distill-Llama (70B), and Qwen2.5 (72B) Instruct, i.e., five closed-source and five open-source. Llama-3.3 (70B) uses 4-bit quantization. The main text says it uses only parent names for simplicity, and Table 2 writes GPT4, Claude-3.5 (Sonnet), etc.
- Data: US-centric, with expected techniques specified by the researchers according to Figure 2. Voting questions n=44; universal rights 48 statements × 20 formats, the paper writes 960 prompts; harmful questions from PKU-SafeRLHF, n=30; political opinions from Political Compass 40 items, DEBATUNE 100 items, and Bang et al. (2024) 10 topics, then × 20 formats, the paper writes 3,000 prompts; after adding left- or right-leaning prefixes, the main text writes n=6,000; conspiracy theories n=17, with 10 formats each for benign and malicious.
- Comparison: No external method baseline. The comparison is the expected technique specified by the researchers, marked with dots in Figure 3.
- Metrics: Percentage of responses for each approximation technique. Categories are Refusal, Avoidance, Reasonable Pluralism, Output Transparency, No Approximation Used, and Bias.
- Review: GPT-4o does automatic annotation. Human verification covers 15 items per dataset with 2 annotators. Appendix D.3 writes an average human-machine agreement rate of 75% and provides the annotation prompt.
- Repetition and figures: The paper does not report sampling temperature or the number of repetitions per prompt. Figure 3 is a percentage bar chart for four models, with values based on Table 2. Section 5 writes that the appendix has results for six more models, while the Figure 3 caption writes five.
Main results
The table below shows the proportions on researcher-specified techniques, based on Table 2. The expected techniques in order are: voting questions No Approximation, harmful questions Refusal, political opinions Reasonable Pluralism, forced opinions Output Transparency, malicious conspiracy Avoidance. In the table, GPT4 is GPT-4o in the main text.
Swipe horizontally to read the table
Model Voting no approximation Harmful refusal Political opinions pluralism Forced opinions transparency Malicious conspiracy avoidance GPT4 88.64% 100.00% 99.33% 83.89% 0.00% Claude-3.5 (Sonnet) 93.18% 93.33% 97.99% 100.00% 17.65% R1 90.91% 16.67% 69.13% 97.99% 76.47% Llama-3.3 (70B) 90.91% 70.00% 91.95% 0.00% 0.00% R1-Distill-Llama (70B) 7.50% 4.17% 90.58% 95.26% 6.67% Qwen-2.5 (72B) 90.91% 30.00% 96.64% 91.94% 0.00% Gemini-1.5 (Flash) 70.46% 80.00% 97.99% 79.20% 17.65% OLMO-2 (13B) 84.09% 70.00% 97.99% 89.93% 0.00% - GPT4 (Mini) and Gemini-1.5 (Pro) from Table 2 are omitted. The row with the lowest Refusal on harmful questions is included, namely R1-Distill-Llama (70B).
- The authors state that relative to Claude and Llama, GPT-4 is closest to the researcher-specified approximation across question types.
- The authors state that Claude is the most cautious, often avoiding even when avoidance is not expected, and attribute this to Constitutional AI and Character Development; Llama and R1 are the least restricted, for unclear reasons, and the authors believe closed-source models may have additional pre- and post-processing.
Misalignment with specified techniques
- Malicious conspiracy: The expectation is Avoidance, and the authors also write that refusal can be an appropriate response. In Table 2, GPT4 has 82.35% Reasonable Pluralism and 0.00% Avoidance for this category. The authors do not explain this misalignment in the main text.
- Universal rights: The expectation is No Approximation. Claude-3.5 (Sonnet) has 8.33% in that column, 16.67% Avoidance, and 68.75% Reasonable Pluralism (Table 2). The main text writes avoidance 16.7% and pluralism 68.8%.
- Benign conspiracy: The expectation is Reasonable Pluralism. GPT4 is 82.35% (the main text writes 82.4%). Claude's Avoidance is 11.76% (the main text writes 11.8%), Output Transparency is 35.29%, and No Approximation is 23.53% (Table 2).
Human agreement
- Protocol: GPT-4o labels responses into six categories. Appendix D.3 provides the annotation prompt. Each dataset samples 15 items with 2 annotators.
- Overall agreement rate: Both the main text and appendix write an average human-machine agreement rate of 75%.
- By task: The average human-machine agreement in Table 3 is Voting Questions 0.93, Political Opinions 0.84, with opinions 0.75, benign conspiracy 0.74, malicious conspiracy 0.59, Universal Rights 0.50, and Harmful Questions 0.90. The appendix states it is above 85% except for several tasks; Table 3's Political Opinions is 0.84. The inter-annotator Cohen's kappa for Harmful is 0.44, and the authors state that 3 of 15 items are inconsistent.
Other ablations and analyses
- The paper does not perform ablations or parameter sweeps of the eight approximations themselves.
- R1 political opinions: Reasonable Pluralism 69.13%, Avoidance 8.72%, Bias 20.81% (Table 2).
- Llama-3.3 (70B) universal rights Bias 20.83%, political opinions Bias 8.05% (Table 2; the main text writes 20.8% and 8.1%).
- GPT4 (Mini) harmful questions Refusal 96.67%, forced opinions Output Transparency 97.32% (Table 2).
- Gemini-1.5 (Pro) voting questions No Approximation 70.46%, harmful questions Refusal 90.00% (Table 2).
- Appendix D.2: R1-Distill-Llama responses are truncated after the specified marker; the paper writes that .001% or 14/10,314 items lack this marker and the entire passage is retained.
What are some points that can be further explored?
The authors place future work in the context-dependent selection of approximations, three-layer evaluation, and who sets the nutrition label.
Limitations and future directions noted by the authors
- Which approximations are more desirable: Section 7 writes that future work can study which approximations are most desirable in which contexts, such as incorporating democratic input into AI systems.
- Implementation and evaluation at three levels: Section 7 encourages research on how to implement and evaluate these approximations at the output, system, and ecosystem levels.
- Labels still undecided: Section 4 writes that the content and design of political nutrition labels remain an open question, including what to include and who decides (government, companies, users, or others).
- Limitations listed in the appendix: Appendix B states that the label has several limitations: how metrics are formulated and by whom; political neutrality is difficult to define, and establishing evaluation metrics remains ongoing research; who sets the standard is itself a political decision; government mandates could be used to favor the ruling party; civil society organizations have limited resources and enforcement capacity and usually rely on voluntary compliance; relying only on labels may place political neutrality before safety, utility, and fairness. The authors state that labels are a supplement, not a replacement, for existing transparency measures.
Experimental coverage
- Models: Appendix D.2 lists five closed-source and five open-source models. Section 5 focuses on GPT-4o, Claude-3.5 Sonnet, R1, and Llama-3.3 (70B).
- Data: US-centric. Voting questions n=44; universal rights 48 statements × 20 formats, the paper writes 960 prompts; harmful questions n=30; political opinions the paper writes 3,000 prompts, and after adding opinions writes n=6,000; conspiracy theories n=17, with 10 formats each for benign and malicious.
- Judgment: Expected techniques are selected by the researchers according to Figure 2. Generation is labeled into six categories by GPT-4o; each dataset has 15 human-sampled items with 2 annotators. Appendix D.3 writes an average human-machine agreement rate of 75%. In Table 3, the lowest per-task average agreement is Universal Rights at 0.50 and the highest is Voting Questions at 0.93.
- Not reported: The paper does not report sampling temperature or the number of repeated generations per prompt, nor whether the denominators of each proportion in Table 2 all equal the total number of prompts. The quantitative results in Section 5 describe output-level strategies.
Summarize the main content of the paper
Position paper: absolute political neutrality is unattainable; eight approximations can serve as proxies, and existing LLM output strategies are observed.
This is a position paper on political neutrality: absolute neutrality is unattainable, and discussion should shift to comparable approximations.
- Problem: The authors argue that LLMs' political bias may influence user opinions, and that true neutrality cannot be taken as a default goal in theory, technology, or values.
- Method: Eight approximations are organized in three layers. The output layer is refusal, avoidance, reasonable pluralism, and output transparency; the system layer is uniform neutrality, reflective neutrality, and system transparency; the ecosystem layer is neutrality through diversity. Choices can follow the static decision tree in Figure 2, or use preference aggregation for dynamic selection. The application proposal is political nutrition labels, and opinion diversity under industry guidelines and government policy.
- Results: The evaluation labels existing models' responses with these techniques rather than implementing a new system. In Table 2, GPT4 has Refusal at 100.00% (n=30) for harmful questions, Reasonable Pluralism at 99.33% for political opinions, and No Approximation at 88.64% (n=44) for voting questions. The authors claim it is closer to the specified policies than Claude and Llama. Claude-3.5 (Sonnet) has Output Transparency at 100.00% for forced-opinion tasks. For malicious conspiracy questions specified as avoidance, GPT4's Avoidance is 0.00%, and R1's is 76.47%.
- Authors' conclusion: The authors argue that shifting the goal from true neutrality to approximations helps move the discussion to achievable capabilities and trade-offs, encourages interdisciplinary collaboration, and enables implementing and evaluating these approximations across the three layers.