Researchers Adrian de Wynter and team released an arXiv paper challenging the tendency to attribute human‑like qualities to large language models. By training a modest neural net on the real‑time strategy game Age of Empires II, they demonstrate that any sufficiently complex substrate can exhibit similar behaviors, calling into question claims of LLM exclusivity.

The study notes that interpreting such behaviors depends on the chosen substrate, whether LEGO constructions or a city‑scale simulation like the Greater Boston Area. Without explicit measurement criteria, observers risk projecting anthropomorphic traits arbitrarily. The authors argue that assuming universal attributes yields circular or empty conclusions, regardless of the experimenter's stance.

To avoid these pitfalls, the paper proposes a null assumption: treat LLMs as non‑unique and design experiments accordingly. It provides concrete examples and briefly surveys related work, even proving that Age of Empires II is functionally Turing‑complete. The authors conclude that rigorous, substrate‑aware methodology is essential for any empirical debate on AI cognition.