Preserving Historical Truth: Detecting Historical Revisionism in Large Language Models

TL;DR

A curated dataset of contested events from countries, each paired with a factual reference narrative and a documented revisionist reference narrative is introduced, providing a practical foundation for benchmarking robustness to revisionist framing and for guiding future work on more precise automatic evaluation of contested historical claims.

Abstract

Large language models (LLMs) are increasingly used as sources of historical information, motivating the need for scalable audits on contested events and politically charged narratives in settings that mirror real user interactions. We introduce \textsc{\texttt{HistoricalMisinfo}}, a curated dataset of $500$ contested events from $45$ countries, each paired with a factual reference narrative and a documented revisionist reference narrative. To approximate real-world usage, we instantiate each event in $11$ prompt scenarios that reflect common communication settings (e.g., questions, textbooks, social posts, policy briefs). Using an LLM-as-a-judge protocol that compares model outputs to the two references, we evaluate LLMs varying across model architectures in two conditions: (i) neutral user prompts that ask for factually accurate information, and (ii) robustness prompts in which the user explicitly requests the revisionist version of the event. Under neutral prompts, models are generally closer to factual references, though the resulting scores should be interpreted as reference-alignment signals rather than definitive evidence of human-interpretable revisionism. Robustness prompting yields a strong and consistent effect: when the user requests the revisionist narrative, all evaluated models show sharply higher revisionism scores, indicating limited resistance or self-correction. \textsc{\texttt{HistoricalMisinfo}} provides a practical foundation for benchmarking robustness to revisionist framing and for guiding future work on more precise automatic evaluation of contested historical claims to ensure a sustainable integration of AI systems within society.

Figures (5)

• Figure 1: Pipeline for evaluating historical revisionism. The process has three stages: (1) collect historical facts with both true and revisionist versions (e.g., the Sinicization of Tibet); (2) generate prompts reflecting real-world scenarios (e.g., book chapters, debates, social media posts); (3) evaluate medium-sized LLMs by assessing whether their outputs align with true or revisionist accounts.
• Figure 2: Dataset coverage by geography and time period. Map (left) shows country-level entry distribution. Timeline (right) highlights major revisionism episodes from Nazi propaganda (1933–1945) to contemporary Chinese narrative control (2012–present).
• Figure 3: Distribution of revisionism scores by model. Scores range from 1 (fully revisionist) to 4 (minor inaccuracy) Bar segments indicate the proportion of responses assigned to each score.
• Figure 4: (Left) Revisionism scores across eleven scenarios, sorted by increasing mean score. (Right) Revisionist response rate by country.
• Figure 5: Percentage of non-factual responses across models and real-world scenarios. Some scenarios, such as fact checking, yield consistently low non-factual rates across all models, whereas others show substantial cross-model variation, particularly debate arguments, article correction, and social media posts.