Counterfactual History: AI Between Simulation and Historical Truth

Generative artificial intelligence is transforming the question "what would have happened if?" into an interactive and immersive experience, opening new frontie

Generative artificial intelligence possesses the extraordinary ability to bring the classic question "what would have happened if…" to life, offering an interactive immersion into historical crossroads never explored. The educational value of this technology lies in the possibility of comparing the alternative scenario generated by the machine with real sources, the context of the era, and established historiographical interpretations. The risk, however, is that of confusing a textually plausible narrative with a verified historical reconstruction, mistaking a probabilistic hallucination for a parallel truth.

The "What If" as an Interactive Pedagogical Tool

The use of counterfactual reasoning in history teaching makes it possible to overcome passive learning, training students to understand causality, choice, and the contingency of historical events. Chatbots endowed with a historical identity can push students to abandon simple factual questions in favor of reflective, ethical, and hypothetical inquiries, thus fostering the development of historical perspective and critical thinking.

AI offers the potential to simulate alternative historical scenarios in real time, such as the course of World War II or the dynamics of the Cuban Missile Crisis. It is essential, however, to present these uses as exploratory investigations into the educational potential of the technology, and not as definitive empirical evidence about past events.

The Technical Limit: Simulating Does Not Mean Proving

The central methodological problem of using Large Language Models (LLMs) is the absence of an internal mechanism to validate their own responses. Consequently, the models can produce content that is erroneous or entirely devoid of logical sense. This limitation is aggravated precisely in scenarios without historical precedent: the more the narrative moves away from established facts to explore the counterfactual, the more the simulation becomes purely and unreliably speculative.

Hallucinations, structural biases, and security vulnerabilities inevitably compromise the reliability of LLMs in simulations of social and human systems. An AI agent playing the part of an emperor, a general, or a chancellor does not really know what that person would have done in an alternative situation. Although prompting strategies such as chain-of-thought can reduce the incidence of hallucinations in some scenarios, they cannot eliminate them entirely. For this reason, the fundamental rule of thumb for teachers and students requires always asking the agent to make explicit the sources, dates, margins of uncertainty, and any alternative interpretations.

The Historical Method as an Antidote to Hallucination

To transform an AI simulation into a genuine exercise in historical thinking, one must anchor oneself to the core competencies of historiographical inquiry: sourcing, contextualization, and corroboration, that is, the ability to evaluate the origin of a source, place it in its precise context, and compare it with other evidence. At the institutional level, UNESCO recommends an approach to AI in education that is rigorously human-centered, based on ethical and pedagogical validation, privacy protection, adequate training for teachers, and clear age limits for the autonomous use of generative tools.

Operational Box: Wineburg's Heuristics Applied to AI Before believing the alternative timeline proposed by a chatbot, it is necessary to apply the evaluation heuristics developed by historian Sam Wineburg:

• Author and Motivations: Evaluate who produced the original source on which the AI is based, analyzing their motivations and target audience.
• Contextualization: Frame the document or event within its specific historical and cultural moment.
• Independent corroboration: Compare the generated information with other independent primary sources to confirm its validity.

A computational simulation can effectively show how a different choice could have opened alternative paths, but the discipline of history demands that one always distinguish what we know from what we hypothesize or simply imagine. The useful question to ask of the output is never only "what would have happened if?", but rather "what objective evidence makes this answer plausible?". Artificial intelligence has the merit of making the past dynamically interrogable, but it does not have the power to make it predictable. Its best use is not to generate perfect parallel universes, but to provide a contrasting tool that helps us understand more deeply why, in reality, exactly what happened, happened.

Bibliographic References and Sources

• Designing Personified Historical Chatbots to Enhance Critical Thinking — ERIC [2487]
• Towards bad history? A call for the use of counterfactual historical reasoning in history education — Historical Encounters [2497, 2528]
• The Potential of AI to Simulate Real-Time Historical What-If Scenarios — PhilArchive [2493]
• Validation is the central challenge for generative social simulation — Artificial Intelligence Review (2025) [2489]
• Too human to model: the uncanny valley of large language models in simulating human systems — Nature Computational Science (2026) [2488]
• Survey and analysis of hallucinations in large language models — PMC (2025) [2490]
• Teaching historical thinking in practice — History Education Research Journal (2022) [2512]
• Teaching Historical Thinking — ERIC (Sam Wineburg's Heuristics) [2517]
• Guidance for generative AI in education and research — UNESCO (2023) [2504, 2506]

Article by the Editorial Staff of La Bussola dell'IA