Seeing Like an AI: How LLMs Apply (and Misapply) Wikipedia Neutrality Norms

1. Problem Statement

Large Language Models (LLMs) are trained on broad, general corpora but are increasingly deployed in specialized communities with strict, nuanced norms (e.g., Wikipedia's Neutral Point of View (NPOV) policy). The core research question is whether providing high-level rules (prompts) to general-purpose LLMs is sufficient for them to replicate the nuanced judgment of community experts.

Specifically, the authors investigate the tension between universal plans (explicit rules provided via prompts) and situated actions (context-dependent judgments required by community norms). They ask:

• Can LLMs detect NPOV violations as accurately as human editors?
• Can LLMs generate neutral edits that align with human editorial standards?
• Do non-expert users (crowdworkers) perceive AI-generated neutralizations as superior to human ones, and what are the implications of this preference?

2. Methodology

The study employs a mixed-methods approach involving computational experiments, human subject evaluations, and qualitative analysis.

Datasets

• Wikipedia Neutrality Corpus (WNC): A dataset of 423,823 revisions (filtered to ~5,400 for experiments) containing pairs of biased edits and their neutralized counterparts by human editors.
• Models Tested: GPT-4 (state-of-the-art), ChatGPT 3.5 (commonly used), and Mistral-Medium (open-source).

Task 1: Bias Detection (Classification)

• Setup: Models were asked to classify edits as "biased" or "neutral."
• Variables:
  • Definitions: Minimal (no context), NPOV (verbatim policy), NPOV-Scoped (policy + specific bias types).
  • Examples: Zero-shot vs. Few-shot (10 examples).
• Metrics: Accuracy, False Positive/Negative rates, and analysis of model rationales.

Task 2: Bias Correction (Generation)

• Setup: Models were tasked with rewriting biased edits to be neutral.
• Approaches:
  • Zero-Shot: Direct instruction to neutralize.
  • Constitutional AI (CAI): Two-step process (critique the violation, then rewrite based on the critique).
  • Constraints: Some prompts included instructions to make "minimal changes."
• Computational Metrics:
  • Precision/Recall: Comparing words removed/added by AI vs. Human editors (Human edits = Gold Standard).
  • Edit Distance & Word Count: Measuring the intensity of changes.
  • BLEU Score: Text similarity.
• Human Evaluation: 147 crowdworkers rated AI vs. Human rewrites on Neutrality, Fluency, and Information Addition/Removal.

Task 3: Qualitative Analysis

• Manual inspection of specific cases to understand why AI and human edits diverged (e.g., grammatical tweaks vs. substantive neutrality changes).

3. Key Results

A. Detection Performance: Poor Accuracy

• Low Accuracy: LLMs struggled to detect bias, achieving a maximum accuracy of 64% (GPT-4 with NPOV-Scoped prompt), barely above the 50% chance baseline.
• Contrasting Biases: Models exhibited distinct, persistent priors:
  • ChatGPT 3.5: Over-predicted bias (high False Positives).
  • Mistral-Medium: Under-predicted bias (high False Negatives).
  • GPT-4: Most balanced but still inaccurate.
• Heuristic Reliance: Models often relied on simple heuristics, such as the presence of highly subjective adjectives (e.g., "brilliant," "unfortunately"), even when those words were part of a reported fact or quote, leading to false positives.

B. Generation Performance: High Recall, Low Precision

• High Recall (79%): AI successfully removed the specific words that human editors flagged as biased.
• Low Precision (37%): AI made many additional changes that humans did not.
  • Pattern Divergence: Human editors primarily removed words to neutralize text. AI editors tended to add words (e.g., hedging, explanatory phrases) and make grammatical/stylistic changes.
  • "NPOV+": AI edits were "NPOV plus" (NPOV+), meaning they satisfied neutrality but also altered grammar, flow, and style unnecessarily.

C. Human Evaluation: The "AI Preference" Paradox

• Preference for AI: Despite the low precision and excessive changes, crowdworkers rated AI rewrites as more neutral (70%) and more fluent (61%) than human editor rewrites.
• Drivers: This preference is likely driven by:
  1. RLHF (Reinforcement Learning from Human Feedback): Models are tuned to produce text that general humans find "good" (fluent, standard grammar).
  2. Perceived Objectivity: The AI's tendency to add hedging or explanatory context made the text feel more "balanced" to laypeople, even if it deviated from the minimal-change philosophy of Wikipedia editors.

4. Key Contributions

1. Empirical Evidence of the "Generalization Gap": Demonstrates that providing high-level rules (prompts) is insufficient for LLMs to replicate the nuanced, situated judgment of community experts, even when the rules are clearly documented.
2. Detection vs. Generation Asymmetry: Highlights that while LLMs struggle to detect subtle bias (due to lack of context), they are surprisingly effective at generating text that appears neutral to the general public, often at the cost of precision.
3. The "NPOV+" Phenomenon: Identifies a specific failure mode where AI models over-correct by adding grammatical and stylistic changes unrelated to neutrality, potentially increasing moderator workload.
4. Divergent Stakeholder Preferences: Reveals a conflict between community experts (who prefer minimal, targeted edits) and general readers (who prefer fluent, heavily hedged AI rewrites).

5. Significance and Implications

• Community Moderation: Deploying LLMs as automated moderators on platforms like Wikipedia poses risks. While they may improve the "flavor" of content for readers, their tendency to make extraneous changes could erode editor agency and increase the burden on human moderators to verify and revert unnecessary edits.
• Trust and Agency: The study suggests a potential trade-off: satisfying the general public's expectation of "neutrality" (via fluent, hedged text) may conflict with the community's definition of neutrality (minimal, factual precision).
• Future Directions:
  • Fine-tuning: General-purpose models may require fine-tuning on community-specific decisions rather than just prompting.
  • Hybrid Systems: A "human-in-the-loop" approach where LLMs act as draft generators for flagged content, but humans retain final approval, may be necessary.
  • Multi-Agent Systems: Simulating pluralistic deliberation (multiple agents with different biases) might better approximate the complex consensus-building of human communities than a single agent.

In conclusion, the paper argues that while LLMs can be steered to follow rules, they "see" the world differently than community members. They apply rules through the lens of general language patterns and RLHF preferences, often resulting in outputs that are fluent and seemingly neutral to the public but technically divergent from expert community norms.