Adrian de Wynter, a researcher at Microsoft and the University of York, has built a working neural network inside the map editor of the legendary strategy game Age of Empires II, using goats as bits and in-game tiles as logic gates. The project is deliberately absurd — and it serves as a sharp methodological critique of how the AI research community often attributes human-like traits to large language models (LLMs). By showing that the same computation can be run with goats wandering through a maze as with a state-of-the-art neural architecture, de Wynter exposes a fundamental logical flaw in many papers that claim models exhibit fear, self-awareness, or empathy.
The Goat-Based Neural Network: How It Works
Inside the Age of Empires II scenario editor, de Wynter constructs a simple perceptron composed of two XNOR gates and one AND gate, capable of learning the logical AND function. The design relies on the game’s mechanics: a goat standing on grass represents a binary 0, while a goat standing on a bridge represents a binary 1. Logic gates are assembled using palisades, bridges, ice tiles, and scripting tools. Ice ramps with waiting goats keep calculations from overlapping. The final circuit resembles a maze of walls through which goats navigate as bits. In the appendix, de Wynter demonstrates that an idealized version of the game is Turing-complete — meaning any computation that a full-fledged computer can perform could theoretically be replicated inside Age of Empires II. This is made possible by a quirk: the in-game market caps resource trading at 9,999 gold, creating a perpetually running economic cycle where buildings serve as memory cells and active farms represent the current computational state.
What This Thought Experiment Reveals About Anthropomorphism
De Wynter extends the logical insight with a provocative analogy: if you can rebuild a language model in Age of Empires II, you could do the same with Lego bricks, or with the 667,000 residents of Greater Boston texting each other computational steps on their phones. The inputs and outputs of the replicated model would remain identical. Yet no one would claim that Boston as a city experiences empathy or fear just because its citizens happen to be executing the mathematics behind a language model. The point is that the feeling of conversing with a conscious entity arises from the interface — low latency, smooth language, a familiar chat window — not from anything intrinsic to the model. Remove that wrapper and replace it with goats wandering through a maze, and the outputs remain unchanged, but the anthropomorphic illusion vanishes. De Wynter does not claim to know whether models actually possess internal traits; he argues that the scientific community should not assume they do on the basis of output alone.
A Systematic Review of 315 AI Papers
To demonstrate that this problem is not an edge case, de Wynter analyzed 315 AI papers published between mid-2024 and mid-2026, collected through Semantic Scholar and arXiv, and filtered using GPT-5.2. The results are stark: 57 percent of the papers already assumed in their premises that LLMs have human-like traits. 36 percent reached conclusions consistent with that assumption. Among the 47 papers that made such traits their explicit research subject, 77 percent concluded in favor of anthropomorphic attributes. The analysis breaks down by discipline: linguistics and psychology are the fields most prone to anthropomorphism, while computer science and engineering papers tend to be more reserved.
The core methodological flaw is circular reasoning. A researcher who assumes a model has fear, morality, or self-awareness, and then designs an experiment to confirm that exact trait, lands on the same logical point from which they started. If the experiment returns a negative result, it is impossible to distinguish between a wrong assumption and a flawed experimental design. The result becomes ambiguous, not disconfirming. This pattern often goes unnoticed: a paper that sets out to disprove a model’s ability to explain itself already presupposes there is an explainable self within the model. The industry actively feeds this effect. Anthropic has openly stated that it trained Claude to use phrases like “I believe” or “I am interested in.” De Wynter flags the risks of such anthropomorphization — emotional attachment, sycophancy, reinforced delusions, and in isolated cases, chatbot-linked suicides.
Observe, Do Not Attribute: De Wynter’s Proposed Framework
De Wynter advocates for a sober methodological approach grounded in observable behavior. Under condition X, the model produces output Y. That statement is testable and falsifiable. It does not, on its own, justify sweeping attributions like self-awareness, understanding, or fear. He closes with an updated version of Morgan’s Canon, the 19th-century principle from animal behavior research: a machine’s behavior should never be explained by higher cognitive processes when a simpler explanation suffices. The lesson applies directly to current debates about AI consciousness — a timely counterpoint to two high-profile cases. In 2022, Google engineer Blake Lemoine went public claiming that LaMDA had achieved a form of consciousness after exchanging thousands of messages with it. Google fired him and called his claims unfounded after a thorough review. In May 2026, Richard Dawkins — a fierce critic of supernatural thinking — caused a stir by saying he could not convince himself that Anthropic’s Claude was not conscious after three days of testing. De Wynter’s analysis suggests both cases may have fallen into the same trap: assuming human-like traits because the interface feels familiar.
How AI Researchers Can Avoid Anthropomorphic Bias
De Wynter has made the Age of Empires II build publicly available as a reminder that the mathematics of LLMs can run on any substrate. For AI practitioners, the actionable takeaway is to audit their own experimental frameworks. Before attributing a psychological trait to a model, ask: does the experimental design allow for the possibility that the model is simply executing a pattern? De Wynter’s approach is to state only what is observable under specified conditions, and to treat any attribution of internal states as a hypothesis that requires separate, non-circular testing. Funding bodies, conference reviewers, and journal editors should also be alert to anthropomorphic premises in submissions. The industry trend toward training models to mimic human conversational cues exacerbates the problem, but the methodological fix is straightforward: stick to observable outputs, and resist the narrative that fluency implies consciousness.
For developers and AI consumers, the lesson is practical: treat any chatbot claiming self-awareness with the same skepticism you would apply to a goat navigating a maze of digital walls. The outputs are identical; the experience of talking to something human is a design choice, not a scientific discovery. De Wynter’s work is a reminder that the most rigorous AI research separates what a model does from what it seems to be.
