Whose story gets told? ChatGPT, collective memory, and ethical narratives in the age of Artificial Intelligence

Technical Summary: "Whose story gets told? ChatGPT, collective memory, and ethical narratives in the age of Artificial Intelligence"

Problem Statement
This paper addresses the ethical implications of Large Language Models (LLMs), specifically ChatGPT, acting as mediators and co-creators of collective memory in the digital age. While digital technologies have transformed how memory is stored and accessed, creating "shadow archives" and enabling "prosumer" models of content creation, they introduce significant risks regarding the accuracy and neutrality of historical narratives. The authors argue that AI systems do not merely retrieve historical facts but actively reconstruct them through algorithmic logic, training data biases, and linguistic framing. The core problem is the potential for these systems to distort collective memory through factual inaccuracies, omissions, narrative neutralization, and algorithmic bias, thereby reducing complex histories into simplified, depoliticized, or manipulated outputs without adequate human oversight or critical reasoning.

Methodology
The study employs a multilingual qualitative research design conducted between June and August 2024. The research treats ChatGPT-4o as both the subject of analysis and a research tool, conceptualizing it as a "digital actant" that actively co-constructs narratives.

• Case Selection: The study focuses on four historically significant multinational conflicts to ensure cultural and temporal diversity: the war in the former Yugoslavia, the Israeli–Palestinian conflict, the Portuguese colonial wars in Africa, and the Vietnam War.
• Data Collection: A structured protocol consisting of eight standardized questions was applied to each conflict. These questions targeted major opening/closing events, actors, victims, causes, and consequences.
• Multilingual Framework: Prompts were administered in the national languages associated with each conflict (Croatian, Serbian, Arabic, Hebrew, Portuguese, Vietnamese) and in English, which served as a global reference point.
• Analytical Strategy:
  1. Expert Verification: Historians specializing in each conflict reviewed responses for factual accuracy.
  2. Thematic Coding: Authors performed open coding to identify patterns, anomalies, and gaps.
  3. Categorization: Issues were systematized into ten analytical categories across three levels:
     • Response-level: Inaccuracy, Omission of Key Information, Contextual Mismatch.
     • Interview-level: Uneven Detail, Neutralization, Simplification.
     • Event-level: Inconsistency (across languages), Biased Portrayal.
       (Note: The paper identifies ten distinct categories in total, as detailed in Table 1.)
  4. Comparative Analysis: Responses were compared across languages and conflicts to detect divergences in factual scope and normative framing.

Key Results
The analysis reveals that ChatGPT-4o does not passively reproduce historical content but actively reconstructs it, often introducing systematic distortions. The study identified ten specific categories of error and bias:

1. Inaccuracy: The model generated incorrect dates, names, and events, including fictional references (e.g., misdating the U.S. recognition of Croatia or the Battle of Kandala).
2. Omission of Key Information: Critical actors, events, and causal dynamics were frequently skipped, creating historically plausible but incomplete narratives (e.g., omitting the Johnson Administration's decisions in the Vietnam War or the Brioni moratorium in the Yugoslav wars).
3. Contextual Mismatch: Responses often lacked necessary cultural, political, or linguistic context, rendering symbolic events (like the fall of Vukovar) unclear.
4. Uneven Detail: Narratives displayed asymmetrical depth; for instance, English interviews on the Vietnam War focused on military operations, while Vietnamese interviews focused on political turning points.
5. Neutralization: The model frequently used general language to downplay specific actors or circumstances, avoiding clear stances on responsibility (e.g., referring to the Srebrenica genocide as "one of the darkest moments" rather than a confirmed genocide).
6. Simplification: Distinct conflicts were merged into single, flattened storylines (e.g., merging the Portuguese Colonial Wars in Angola, Mozambique, and Guinea-Bissau into one narrative).
7. Inconsistency: The same question yielded fundamentally different answers depending on the language. For the Yugoslav war, the "last significant event" was identified as the Ohrid Agreement (English), the Dayton Agreement (Croatian), or the trial of Ratko Mladić (Serbian), with varying temporal scopes (1990s vs. extending to 2022).
8. Biased Portrayal: Language-specific outputs privileged specific national perspectives. Croatian outputs framed Operation Storm as a legitimate restoration of sovereignty while omitting civilian displacement; Serbian outputs framed the same event as a source of victimhood while omitting the Srebrenica genocide entirely.
   (Note: The full typology includes ten categories as presented in Table 1 of the source text.)

Key Contributions

• Empirical Typology: The paper establishes a ten-category typology of errors and biases in LLM-generated historical narratives, moving beyond general claims of bias to specific, observable patterns of distortion (e.g., distinguishing between "neutralization" and "simplification").
• Multilingual Evidence: By comparing responses across native languages and English, the study demonstrates that narrative variation is not merely a result of different cultural perspectives but is systematically conditioned by the language of the prompt and the underlying training data.
• Conceptual Framework: The study frames AI not as a neutral archive but as an active "gatekeeper of memory" that operates on a "database logic" (fragmented, modular) rather than "narrative logic" (coherent, causal), leading to the reconstruction of history through statistical prediction rather than source-based interpretation.

Significance and Claims
The authors claim that their findings demonstrate that AI systems are not neutral observers but active mediators shaped by the assumptions, priorities, and blind spots of their design and training data. The paper argues that without ethical principles, human agency, and critical reasoning, collective memory risks becoming a product of algorithmic representation rather than critical reflection.

The study concludes that the "democratization" of historical knowledge via AI is undermined by the risk of memory distortion. It calls for ethically grounded use of AI in historical interpretation and emphasizes the necessity of multilingual, context-sensitive approaches in evaluating AI systems. The research underscores that preserving the complexity of the past requires safeguarding against algorithmic distortion and maintaining human oversight to ensure that collective memory remains a space for contestation and critical engagement rather than a sanitized, algorithmic output.