Microsoft researcher builds goat-powered neural network in Age of Empires II to challenge LLM consciousness

Microsoft Researcher Challenges LLM Consciousness With Absurdist Experiment

Adrian de Wynter, a Microsoft researcher based at the University of York, has built a functioning neural network inside Age of Empires II using goats as AI building blocks to challenge widespread assumptions about LLM consciousness 1

1

. His research paper, titled "If LLMs Have Human-Like Attributes, Then So Does Age of Empires II," makes a pointed argument about anthropomorphism in AI by demonstrating that the same computational principles powering chatbots like ChatGPT, Claude, and Gemini can be replicated using virtual livestock 2

2

. The project deliberately dials absurdism up to 11 to expose flawed assumptions permeating AI research and public perception.

How the Goat-Powered Neural Network Actually Works

Using Age of Empires II's scenario editor, de Wynter constructed working logic gates with in-game elements serving as computational components 4

4

. Grass represents binary 0, bridges represent binary 1, and goats act as bits moving between states. He successfully created NAND gate, XNOR, and AND operations—the fundamental building blocks needed to construct a 1-bit perceptron, one of the simplest forms of artificial intelligence 1

1

. While de Wynter didn't build a complete LLM, the working perceptron serves as proof of concept that neural networks underlying modern chatbots could theoretically be implemented using any sufficiently powerful substrate—whether silicon chips or virtual goats 2

2

.

The Problem With Anthropomorphism in AI Research

De Wynter's investigation revealed a troubling trend in academic literature. From 337 computer science papers he reviewed over the last two years, 57% assumed LLMs could have human-like traits without establishing proper experimental protocols to validate such claims 1

1

. This confirmation bias affects research design, testing methodologies, and ultimately the conclusions drawn about AI sentience. The researcher argues that "in no case is a machine's activity to be interpreted in terms of higher cognitive processes, if it can be fairly interpreted in terms of processes which stand lower in the scale of cognitive evolution and development" 1

1

. The absurdity of LLM sentience becomes apparent when observers watch goats scuttling around performing the same fundamental operations that power commercial chatbots 3

3

.

Why Presentation Matters More Than Substance

The critical insight from de Wynter's work centers on how substrate affects perception of chatbot consciousness. He argues that "many anthropomorphic measurements in AI are measurements of presentation, rather than of an actual system's behaviour" 2

2

. When users interact with ChatGPT through a conversational interface trained on natural language, they readily perceive human-like qualities. Strip away that presentation layer and replace it with goats acting as NAND gate components, and the illusion of consciousness evaporates entirely 2

2

. The same computational processes appear profoundly different depending on whether they're dressed in natural language or represented by virtual livestock, yet the underlying mechanisms remain identical.

Commercial Incentives Behind the Consciousness Narrative

AI companies aren't rushing to discourage anthropomorphism in AI—in fact, they may actively benefit from it. Research indicates that people buy more products when they can empathize with them, including AI and chatbot subscriptions 1

1

. Top executives at AI companies have publicly entertained the possibility that their systems might exhibit signs of consciousness, despite lacking scientific evidence. Chatbots are deliberately trained to mimic the shape and tone of natural conversation, making it effortless for users to project personality, emotion, or even sentience onto them 1

1

. Some users even admit to having relationships with AI chatbots, demonstrating how effectively the presentation layer obscures the mechanical reality beneath.

What This Means for Future AI Research

De Wynter proposes that researchers "need to stop assuming that LLMs behave like humans just because they were trained with natural language" and instead "perform experiments that allow us to see LLMs as how they are, not how we believe they should be" 4

4

. The lack of widely-accepted experimental protocols for evaluating AI sentience means that starting research from either position—assuming consciousness exists or doesn't exist—introduces bias that compromises scientific validity 3

3

. As AI systems become more sophisticated and their outputs more convincing, the need for rigorous, assumption-free testing frameworks becomes increasingly urgent. The goat-powered demonstration serves as a reminder that persuasiveness and self-consistency, while objectively measurable, cannot imply real or simulated consciousness without proper validation methods 2

2

.