Microsoft researcher Adrián de Wynter built a neural network inside the Age of Empires II map editor to critique anthropomorphism in LLM evaluation. The architecture uses in-game objects: goats on grass represent 0, goats on a bridge represent 1, with logic wired through built-in scripts. The model successfully learned the AND logical operation.

After analyzing over 300 AI research papers, de Wynter found that more than 50% of authors attribute human traits to algorithms, leading to methodological errors and circular experimental logic. He argues that the illusion of consciousness in language models stems from their text interface—moving the same computation to virtual goat behaviour produces identical results without triggering empathy.

The researcher urges colleagues to evaluate models strictly by input and output rather than attributing cognitive functions to code. The map script is published on GitHub, with full details on adewynter.github.io.