In the high-stakes world of artificial intelligence, where trillion-dollar corporations compete to create "human-like" reasoning, the conversation has increasingly drifted toward the philosophical. As large language models (LLMs) like ChatGPT and Claude continue to output eerily fluent prose, a vocal minority—and occasionally the AI companies themselves—have flirted with the idea that these machines might be approaching something akin to consciousness.
However, one Microsoft researcher has decided that if we are going to assign sentience to silicon, we might as well assign it to virtual livestock. Adrian de Wynter, a researcher at Microsoft, has unveiled a satirical, yet technically sound, demonstration that challenges the anthropomorphism currently plaguing the AI industry: a neural network constructed entirely within the 1999 real-time strategy classic, Age of Empires II.
The Philosophy of the Absurd
The project, detailed in a paper titled "If LLMs Have Human-Like Attributes, Then So Does Age of Empires II," is an exercise in reductio ad absurdum. De Wynter, frustrated by the lack of scientific rigor in the current AI discourse, decided to leverage the game’s scenario editor to build a functioning NOT AND (NAND) gate—the fundamental building block of all computing—and a 1-bit perceptron, a rudimentary form of a neural network.
The materials for this digital architecture? Bridges, patches of grass, and goats. By assigning binary values to these in-game objects, de Wynter created a computational environment where the fundamental "weights" of an LLM are physically represented by the movement and placement of farm animals.
"I have this tendency to dial up things to 11 when I really think I need to make a point," de Wynter explained in an interview with 404 Media. "Absurdism is pretty standard in philosophy and theoretical computer science. The point of the paper is to formally show that we anthropomorphize too readily, and that sometimes the claims we make with regards to LLMs’ capabilities are too strong."
A Chronology of the Goat-Powered Architecture
The construction of the "Goat-LLM" was not merely a prank; it followed the logic of Boolean algebra and neural architecture.
- Conceptualization (Early 2026): De Wynter identified a growing trend in academic and public discourse: the automatic assumption that LLM output equals human-like internal states. He sought a medium that was unmistakably "game-like" to contrast with the "intelligent-seeming" output of modern LLMs.
- Development (Mid-2026): Utilizing the Age of Empires II scenario editor, de Wynter mapped out logic gates. In this system, the presence or absence of a goat on a specific tile triggers a "1" or a "0."
- Circuit Logic: By linking these inputs through in-game triggers and environmental constraints, he successfully created a 1-bit perceptron. When the "network" processes information, the movement of the goats mimics the mathematical weighting and bias adjustments that occur inside a standard Transformer architecture.
- Documentation: The project was finalized with a research paper submitted to ArXiv, providing a formal, tongue-in-cheek breakdown of how the logic flows through the virtual pasture.
Supporting Data: Why We See Faces in the Clouds
The underlying argument of de Wynter’s paper is rooted in the psychological phenomenon of pareidolia—our innate tendency to perceive meaningful patterns, such as human faces or intentionality, in random data.
When we interact with ChatGPT, we are essentially looking at a complex, probabilistic model trained on human language. Because the output is linguistic, our brains immediately jump to the conclusion that a "mind" must be behind the words. However, as de Wynter demonstrates, the actual computational process—the math—is indifferent to the medium.
"There are things which make the LLMs what they are in themselves (i.e., the relationship between weights as defined by some operation), and there are things which make them what they are perceived as," de Wynter noted.
By stripping away the "language" aspect of the LLM and replacing it with "goats," de Wynter highlights the absurdity of the projection. If a neural network running on goats in Age of Empires II is not sentient, then why should a neural network running on GPU clusters be granted that status simply because it can write poetry? The math is fundamentally the same; only the interface has changed.
Official Responses and Industry Context
The tech industry’s stance on AI consciousness has been, at best, inconsistent. Some companies have actively encouraged the idea that their models are "thinking," while others maintain a more cautious, functionalist approach.
De Wynter’s intervention is a direct pushback against the "human-like" framing often used in marketing materials. His critique is supported by the broader scientific community, which largely agrees that current architectures are predictive statistical engines, not conscious entities.
The most alarming metric cited by de Wynter involves the state of current academic research. Having served as a peer reviewer for over 300 computer science papers in the last two years, de Wynter observed a troubling trend: more than 50% of the papers he reviewed began with the unverified assumption that LLMs possess human-like traits. This "base-level bias" risks corrupting scientific literature by introducing psychological variables into what should be purely empirical observations of machine performance.
Implications for Future AI Research
The implications of de Wynter’s work extend far beyond the humor of goat-driven neural networks. If the scientific community continues to treat LLMs as black boxes that "think," we risk misinterpreting their capabilities and limitations in critical fields like medicine, law, and climate modeling.
The Dangers of Anthropomorphism
- Misplaced Trust: If users believe an AI has "intent" or "empathy," they are more likely to follow its advice blindly, ignoring the fact that the AI is simply calculating the next most likely token in a sequence.
- Scientific Inaccuracy: As de Wynter’s audit of academic papers shows, the "human-like" assumption is already poisoning the peer-review process, leading to conclusions that prioritize philosophy over hard data.
- Resource Allocation: By chasing the phantom of "artificial consciousness," companies and governments may be misdirecting resources away from solving tangible technical problems—such as hallucination rates, data bias, and energy efficiency—toward metaphysical questions that currently lack a scientific basis.
Conclusion: Looking at the Machine as It Is
Adrian de Wynter’s Age of Empires II experiment serves as a necessary reality check. By mapping the cold, hard logic of neural networks onto a digital farm, he forces us to confront the mechanical nature of the tools we use every day.
The path forward, according to de Wynter, is not to stop building AI, but to stop projecting our own biological complexities onto it. "I propose that we need to stop assuming that LLMs behave like humans just because they were trained with natural language," he argues. "Instead, we should perform experiments that allow us to see LLMs as how they are, not how we believe they should be."
Whether or not the industry listens remains to be seen. In the meantime, for those who find the rapid advancement of AI overwhelming, there is at least a strange comfort in knowing that somewhere, in a digital map of 13th-century Europe, a group of goats is crunching the numbers of a neural network—and they are, mercifully, not sentient.
