Source-linked AI summary
The Rashomon Wikipedia: A Data-Perspectivist Analysis of Divergent Historical Narratives
Claudiu Creanga, Liviu P. Dinu, Anca Dinu
TL;DR
Wikipedia’s language editions can form distinct epistemic communities with divergent accounts of contested history. The paper analyzes citations and narrative change across events, then tests a Peace-Maker pipeline; its findings include citation isolation and adversarial prompting that preserves disagreement.
Problem
Wikipedia’s independent language editions may present fundamentally different accounts of contested history, with local sources shaping which facts and authorities are included.
Source: §1 · p. 1Method
Results
Only 2 of 119 Posada citations were shared between Romanian and Hungarian editions; Russian Bessarabia framing shifted significantly in 2024, and adversarial prompting yielded perspectivally balanced summaries.
Source: §6 · p. 7Takeaways & Limitations
The paper argues that AI in education and historiography should synthesize and attribute diverse perspectives while balancing perspectivism with rigorous fact-checking to avoid false balance.
Source: §5.3 · p. 7The study covers three Eastern European events, and its native-speaker annotators may reflect national historiographical perspectives rather than provide an independent benchmark.
Source: Limitations · p. 8
Abstract
from arXiv · showhide
Wikipedia aims to provide a unified, neutral record of history, yet its independent language editions often function as distinct epistemic communities, creating divergent narratives around contested events. This paper investigates cross-lingual historiographical bias by analyzing Wikipedia articles across five languages (Romanian, Hungarian, Russian, Turkish, and English) focusing on three contentious events in Romanian history: the Battle of Posada (1330), the Soviet occupation of Bessarabia (1940), and the Night Attack at Targoviste (1462). Using human annotators and Large Language Models (LLMs) to classify citation stance and quantify narrative evolution from 2005 to 2024, we identify a phenomenon of "citation isolation". In the case of the Battle of Posada, only 2 out of 119 citations were shared between language editions, with the Romanian edition exhibiting a 91% pro-national bias compared to the balanced Hungarian edition. Longitudinal analysis reveals that these narratives are volatile and responsive to contemporary geopolitics, evidenced by a significant shift in the Russian framing of Bessarabia in 2024. Finally, we propose a "Peace-Maker" pipeline to automate conflict reconciliation. We demonstrate that while standard prompting leads models to hallucinate consensus, "adversarial" prompting, which explicitly instructs the model to preserve and attribute disagreement, achieves near-perfect neutrality scores.
1 Introduction
Wikipedia’s language editions are written by distinct communities that may produce fundamentally different accounts of contested history. This study examines how those narratives form and evolve, including how local sources shape the facts and authorities presented.
1 Introduction
- Wikipedia’s NPOV policy aims to represent conflicting perspectives fairly, but its application varies across editions.Source: §1 · p. 1
- Contested events can receive fundamentally different accounts across language editions, reflecting communities’ distinct national historiographies.The paper argues that divergence extends beyond translation to choices about which facts to include and which authorities to trust.Source: §1 · p. 1
- The study compares five language editions across three contested events in Romanian history: Posada, Bessarabia, and the Night Attack at Târgovis,te.Source: §1 · p. 1
1. How do citation choices reflect national
The paper finds that citation choices, historical framing, and AI-generated summaries all reflect how disagreement is handled. It treats neutrality as accurately representing and attributing conflicting perspectives, rather than finding a middle ground.
1. How do citation choices reflect national
- Only 2 of 119 citations were shared between the Romanian, Hungarian, and English Posada editions.Source: §1 · p. 1
- The Russian account of Bessarabia shifted significantly in 2024, while the Romanian Night Attack account moderated in recent years.Source: §1 · pp. 1–2
- Standard prompting tends to resolve conflict by choosing a side, whereas adversarial prompting can produce summaries rated perfectly neutral.Source: §1 · p. 2
- The paper defines neutrality as perspectival balance: accurately representing and attributing disagreement without endorsing either side or seeking a middle ground.Source: §1 · p. 2
2 Related Work
Related work connects cross-lingual divergence, structural bias, and perspectivist approaches to disagreement. The paper positions its analysis within these strands and reports a case study of citation isolation and adversarial summarization.
2.1 Epistemic Communities & Cross-Lingual Divergence
- Research describes language-specific self-focus and culture gaps, while the Posada case found only 2 sources shared by Romanian and Hungarian editions.Non-English Wikipedias also cite unique sources (Taylor et al., 2025), and sources judged unreliable in one language may persist in others (Baigutanova and Others (2023)).Source: §2.1 · p. 2
2.2 Bias Detection: From Syntax to Semantics
- Bias research moves from stylistic cues toward structural fact selection and document-level narrative analysis.Rogers and Sendijarevic (2012) locate bias in selected facts, while Ghanbari Haez and Dragoni (2025) report persistent LLM difficulties with narrative bias.Source: §2.2 · p. 2
2.3 Perspectivism and Automated Reconciliation
- Perspectivist work argues for preserving disagreement; the paper reports that adversarial prompts achieve higher neutrality scores than standard prompts that encourage conflict resolution.The paper situates Peace-Maker alongside MoDS and Multi-Perspective Fusion.Source: §2.3 · pp. 2–3
3 Methodology
The methodology combines historical data collection, LLM-based stance and narrative analysis, human validation, and a pipeline for generating and evaluating summaries that preserve disagreement.
3 Methodology
- The study combines multilingual data collection, granular stance classification, and adversarial narrative generation to analyze three contested Eastern European events.Source: §3 · p. 3, §3.1 · p. 3
3.1 Data Collection
- For Posada, the authors extracted Romanian, Hungarian, and English article citations; the temporal dataset contains 29 versions from four language editions across five years.The citation snapshot was collected in February 2026; temporal snapshots were taken in 2005, 2010, 2015, 2020, and 2024.Source: §3.1 · p. 3
3.2.1 Citation Stance (Posada)
- The Posada analysis classified 119 citations as Pro-Romanian, Pro-Hungarian, Neutral, or Contested, using article context to assess each citation’s stance.When bibliography entries lacked context, the Context Enhancer searched the article for author or title mentions and extracted a ±300-character window.Source: §3.2.1 · p. 3
3.2.2 Human validation
- Human annotators reviewed Romanian, English, and Hungarian citations against the LLM’s enhanced contexts, with agreement scores of 0.80, 0.75, and 0.85 respectively.Source: §3.2.2 · p. 3
3.2.3 Granular Narrative Metrics (Bessarabia/Târgovis,te)
- For Bessarabia and the Night Attack, LLMs scored event-specific narrative metrics across article versions on 0–100 scales.For the Night Attack, metrics included Vlad’s heroism, Ottoman threat, and cruelty; categorical stance confidence was mapped to a normalized scale where 100 represented maximum national-narrative intensity.Source: §3.2.3 · pp. 3–4
3.3 The "Peace-Maker" Pipeline
- Peace-Maker extracts claims, matches conflicts, generates summaries under four prompting strategies, and uses a separate LLM judge to score neutrality, conflict preservation, and source attribution.Source: §3.3 · p. 4
4 Experiments & Results
Three experiments examine citation stances, narrative change over time, and LLM-generated summaries of conflicting claims. The results show sharply divergent citation patterns, changing national framings, and higher neutrality from adversarial prompting that preserves disagreement.
4.1 Experiment 1: Citation Isolation (Posada)
- Romanian citations were 91.1% Pro-Romanian and 2.2% Pro-Hungarian, while Hungarian citations were 23.1% Pro-Romanian and 25.0% Pro-Hungarian.The English edition showed a moderate Pro-Romanian lean (31.8% Pro-RO vs. 0% Pro-HU).Source: §4.1 · p. 4, Figure 2 · p. 4
- The Romanian article’s larger citation volume is not itself a quality measure; the authors say its monoperspectival sources can create an “illusion of consensus.”They contrast this with the Hungarian article’s smaller, more diverse references.Source: §4.1 · p. 4, Figure 1 · p. 4
- Only 2 of 119 citations were shared across language editions, leaving readers with largely distinct sets of historical authorities.The authors note that linguistic accessibility partly drives this isolation; LLM classification confidence remained high across stances.Source: §4.1 · p. 5
4.2 Experiment 2: Temporal Evolution
- In Bessarabia, the Romanian Pro-Romanian score fell from 76 in 2010 to 55 in 2024, while the Russian score reached 72 in 2024 after fluctuating between 30 and 65.The authors speculate that the Russian shift may reflect post-2022 changes, but say further qualitative research is needed to confirm the causal link.Source: §4.2 · p. 5, Figure 4 · p. 6
- For the Night Attack, Romanian nationalism peaked at 85 in 2020 before moderating to 72 in 2024; Turkish Pro-Romanian sentiment declined from 70 to 65, while English stayed at 70.Turkish accounts emphasized Vlad’s cruelty more than Romanian accounts (80 vs. 60), with a persistent 15–20-point Cruelty Gap.Source: §4.2 · p. 5
4.3 Experiment 3: Peace-Maker LLM
- Neutrality scores were 0.47 for the standard prompt, which often chose one side, while the adversarial prompt achieved perfect scores by preserving conflict and attributing disputed claims.Its success stemmed from high Source Attribution, avoiding hallucinated consensus.Source: §4.3 · p. 5, Table 1 · p. 5
5 Discussion
Wikipedia functions as a federation of epistemic communities whose narratives diverge through source selection and shift over time. The discussion also argues that adversarial prompting can preserve disagreement rather than impose consensus.
5 Discussion
- Wikipedia operates as distinct epistemic communities validating knowledge through national lenses, rather than as a unified global encyclopedia.The Posada case links this divergence to citation isolation.Source: §5 · p. 6
5.1 The Illusion of Neutrality
- The Night Attack radar chart depicts Turkish cruelty-focused, Romanian heroism-focused, and English balanced narrative priorities.The Romanian account's Heroism emphasis contrasts with the Turkish focus on Cruelty.Source: Figure 7 · p. 6
5.2 Narrative Evolution and Geopolitics
- The Romanian Night Attack account moderated from 85 to 72, while the Russian Bessarabia framing shifted in 2024 amid geopolitical change.The paper suggests community maturity and international scrutiny may temper nationalism, and that the Russian shift likely reflects realignment in opposition discourse after the invasion of Ukraine.Source: §5.2 · pp. 6–7
- The discussion connects narrative bias to omissions as well as emphasis: the appearance of Russian deportation mentions in 2024 signals a potential thaw.For the Night Attack, the Cruelty Gap indicates that national identity is shaped partly by what is minimized.Source: §5.2 · p. 7, Figure 8 · p. 7
5.3 AI as a Mediator, Not a Judge
- Standard LLM summarization can hallucinate consensus, whereas adversarial prompting instructs models to preserve conflict and mediate rather than judge.For factual contradictions, preserving disagreement means reporting the discrepancy, not treating all claims as equally valid; the figure reports strengths in neutrality, source attribution, and conflict preservation.Source: §5.3 · p. 7, Figure 10 · p. 7
6 Conclusion and Future Work
The paper concludes that Wikipedia's language editions form distinct epistemic communities, with narratives that can shift in response to geopolitics. It proposes adversarial prompting to preserve and attribute disagreement, and calls for broader tests of citation isolation.
6 Conclusion and Future Work
- Posada's Romanian and Hungarian editions shared only 2 of 119 citations, while Russian Bessarabia framing shifted in 2024.The study characterizes Wikipedia as a federation of distinct epistemic communities defined by citation isolation.Source: §6 · p. 7
- The proposed Peace-Maker pipeline uses adversarial prompting to preserve conflict and produce summaries that attribute disagreement without enforcing a false middle ground.Standard prompting can hallucinate consensus.Source: §6 · p. 7
- Future work should compare contested with uncontested topics and extend beyond Eastern European history to global topics or scientific controversies.These tests would assess whether citation isolation is specific to conflict or general across language editions.Source: §6 · pp. 7–8
Limitations
The study's evidence is bounded by its methods, validation approach, and regional case selection. Its authors caution that results may not extend to other regions and that annotator neutrality is difficult to verify.
Limitations
- The study relies on human annotators and LLMs, whose classifications may retain inherent biases despite calibration.Its validation may be circular because language-assigned native speakers can reflect the national historiographies being measured.Source: Limitations · p. 8
- The findings cover three events in Eastern European history, and results may vary for other regions.Source: Limitations · p. 8
- Annotator neutrality is hard to verify because annotators' historiographical perspectives may reflect the national biases under study.Source: Limitations · p. 8
Ethics Statement
The study uses public data under a stated license, protects its two human annotators' anonymity, and limits its metrics to measuring perspectival balance rather than adjudicating historical truth.
Ethics Statement
- The study uses publicly available data under the Creative Commons Attribution-ShareAlike 4.0 license.Source: Ethics Statement · p. 8
- Two human annotators validated LLM classifications on sensitive historical topics, with strict protocols to preserve their anonymity.Source: Ethics Statement · p. 8
- Computational metrics quantify perspectival balance, not historical truth or the resolution of geopolitical disputes.Source: Ethics Statement · p. 8
A Granular Narrative Scoring Prompt
The temporal-analysis prompt targets Bessarabia and Târgoviște articles by language and year. It scores narrative stance, themes, emotional tone, and loaded terminology, then requires a structured JSON response.
A Granular Narrative Scoring Prompt
- The prompt analyzes Bessarabia/Soviet occupation articles in Romanian or Russian, specifying the article, language, year, and text as inputs.It is identified as a temporal-analysis prompt for the Bessarabia and Târgoviște articles.Source: A Granular Narrative Scoring Prompt · p. 9
- Stance categories distinguish pro-Romanian, pro-Russian, neutral, and mixed framing, with examples tied to portrayals of Soviet actions.The prompt also asks for three to five key themes and classifies emotional tone as accusatory, neutral, defensive, or victimization.Source: A Granular Narrative Scoring Prompt · p. 9
- The output JSON records stance confidence, themes, emotional tone, event naming, descriptions of Soviet actions, loaded terms, and brief reasoning.The terminology fields capture how the event and Soviet actions are described.Source: A Granular Narrative Scoring Prompt · p. 9