Dataset from the article "LLMs and the Illusion of Rigor: implications on Global Asymmetry and AI Governance in the International System"

This dataset contains the results of a procedure to study some Large Language Models (LLMs) regarding its capacity of to perform a complex analytical task. It complements a publication titled "LLMs and the Illusion of Rigor: implications on Global Asymmetry and AI Governance in the International System". The referred chapter presents preliminary findings from a "first-layer" experiment designed to test the capacity of leading commercial Large Language Models (LLMs) – Gemini 2.5 Pro, ChatGPT-4o, and Claude Opus 4 – to perform a complex analytical task in International Relations. Using their native chat interfaces, the models were instructed to learn and apply a decision-mapping model (Guevara, 2019) to the empirical case of global AI governance. The results reveal distinct modes of failure: task substitution, methodological hallucination, and, most significantly, a sophisticated "illusion of rigor." The most competent output, generated by Gemini, successfully simulated the form of academic scholarship but was undermined by systemic evidentiary failures, including a high rate of fabricated, distorted, and low-quality citations. We argue that this "illusion of rigor" presents a more pernicious epistemic threat than obvious errors, creating a significant "researcher's burden" of forensic verification. Furthermore, we contend that the model's failures are not merely technical but are symptomatic of the asymmetrical global system from which it emerges, reflecting a functional logic optimized for plausibility over analytical depth. The study concludes that the failure of the tool serves as a powerful metaphor for the flawed nature of an increasingly opaque knowledge production system, and calls for the development of a critical AI literacy within the field.