A Microsoft AI researcher built a functioning neural network inside a 1990s video game using goats, grass, and bridges to make a provocative point: if goats can power the basic building blocks of artificial intelligence, then treating chatbots as conscious entities makes no sense. The absurdist experiment highlights a troubling trend in AI research where scientists increasingly attribute human-like qualities to language models without sufficient evidence.

Why Are Researchers Anthropomorphizing AI Chatbots?

Adrian de Wynter, a Microsoft AI researcher based at the University of York, published a paper titled "If LLMs have human-like attributes, then so does Age of Empires II" to challenge what he sees as a widespread bias in artificial intelligence (AI) research. Large language models, or LLMs, are AI systems trained on vast amounts of text data to generate human-like responses. De Wynter's concern is that researchers, companies, and the public are projecting consciousness and emotional intelligence onto these systems far too readily.

The problem runs deeper than casual conversation. De Wynter analyzed 337 academic papers on LLMs published over the last two years and found that 57% of them assumed that these systems could possess human-like traits. This foundational assumption, he argues, colors how researchers design experiments, interpret results, and draw conclusions about what their AI systems can actually do.

People seem drawn to seeing themselves in technology. Some users report having "relationships" with chatbots like ChatGPT, Claude, and Gemini. Companies behind these AI systems have not discouraged this perception. In fact, they may benefit from it. Research shows that people are more likely to buy products and services when they can empathize with them, and that includes AI subscriptions.

How Did a Researcher Turn a Video Game Into an AI Experiment?

De Wynter's approach was deliberately absurdist. Rather than build a full language model, he used Age of Empires II's scenario editor to construct a NAND gate, which is one of the fundamental building blocks of modern computing and neural networks. In his setup, goats acted as individual bits of data, while grass and bridges served as the circuit infrastructure.

A perceptron is the simplest possible neural network, containing just one computational unit. De Wynter demonstrated that even this rudimentary component could be created within the game's mechanics. If something as absurd as Age of Empires II goats can embody the basic architecture of neural networks, then attributing consciousness to commercial chatbots becomes equally absurd, according to his reasoning.

De Wynter explained his methodology to 404 Media, noting that he deliberately "turns absurdism up to 11" to make his point. The goal was not to suggest that goats are intelligent, but rather to expose the logical flaw in assuming that any system capable of performing mathematical operations must therefore possess human-like cognition.

De Wynter

What Does This Reveal About How We Interact With AI?

De Wynter's research identifies a critical distinction between the underlying technology and how people experience it. The goat-based neural network in Age of Empires II and a commercial chatbot like Claude are functionally similar in their basic computational architecture. The key difference lies in the interface and interaction design.

Chatbots are trained using natural language, the everyday language humans speak and write. They employ techniques specifically designed to mimic the patterns, rhythm, and tone of natural conversation. This design choice makes it easy and almost inevitable for users to project personality, emotion, and consciousness onto the system. The interface is "conversation friendly," which encourages anthropomorphization.

De Wynter also highlights the role of confirmation bias. When people are primed to look for human-like traits in technology, they tend to find them, even when simpler explanations exist. A chatbot's ability to generate contextually relevant text can feel like understanding, but it reflects statistical patterns learned during training, not genuine comprehension.

How to Evaluate AI Claims More Critically

• Distinguish between capability and consciousness: An AI system can be excellent at mimicking human conversation without possessing awareness, emotions, or understanding. Recognize that sophisticated output does not equal sentience.
• Question anthropomorphic language in research: When reading about AI systems, pay attention to whether researchers use human-centric terms like "understands," "believes," or "wants." These may reflect bias rather than actual system properties.
• Consider the interface effect: Remember that conversational design is intentional. Companies craft interfaces to feel natural and engaging, which naturally encourages users to treat systems as if they were conscious agents.
• Apply Occam's Razor: If a simpler explanation exists for AI behavior, such as pattern matching or statistical prediction, prefer it over more complex claims about consciousness or understanding.

De Wynter's work invokes a well-established principle in philosophy and cognitive science: "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." In other words, do not assume complexity where simplicity suffices.

The implications extend beyond academic debate. As AI systems become more integrated into daily life, the tendency to anthropomorphize them could lead to misplaced trust, unrealistic expectations, and poor decision-making. Users might rely on chatbots for advice on matters requiring genuine human judgment or emotional intelligence, only to discover that the system was performing statistical pattern matching rather than reasoning.

De Wynter's goat experiment serves as a reminder that the most impressive AI systems remain fundamentally different from human intelligence. They excel at specific, narrow tasks within their training domain. They do not possess consciousness, desires, or understanding in any meaningful sense. The fact that a video game's livestock can instantiate the same basic computational principles as ChatGPT should give us pause before we treat either as conscious entities.