This article presents a critical experiment testing the reliability of a generalist generative AI (ChatGPT 4.0 Plus) when confronted with a highly specialized historical question regarding Victor Hugo’s supposed opposition to ethnographic exhibitions in 19th-century France. The AI, unguided by explicit methodological instructions, fabricated a plausible but entirely fictitious letter from Hugo, persisting in its error until challenged by rigorous source-checking. The analysis reveals that, despite the AI’s theoretical grasp of historical method, its design prioritizes narrative coherence over evidentiary accuracy, lacking intrinsic safeguards against fabricating sources. The experiment demonstrates that, without explicit user intervention, generative AI may produce convincing historical falsehoods, underscoring the risks for non-expert users. The article argues that human oversight and methodological rigor remain indispensable when employing AI in historical research. It concludes that, while AI can support certain research tasks, it cannot autonomously uphold the epistemic and ethical standards fundamental to the discipline
