For the makers and artists relying on generative tools, the debate over whether artificial intelligence possesses a soul is often drowned out by marketing speak. However, a recent experiment suggests that the perceived “humanity” of these systems is largely a trick of the interface, not a fundamental truth of the code itself. If we accept the absurd notion that a chatbot has feelings because it types in English, we must logically accept that a strategy game populated by digital goats is also sentient. This distinction matters deeply for creators who need to understand what they are actually interacting with: sophisticated pattern-matching engines, not conscious interlocutors.
The Goat Experiment
Adrian de Wynter, a Microsoft researcher, recently published a paper titled “If LLMs Have Human-Like Attributes, Then So Is Age Of Empires II” to prove this point. He argues that absurdity is a potent method for exposing the flaws in how we discuss large language models. To demonstrate his thesis, de Wynter constructed a rudimentary neural network inside the real-time strategy game Age Of Empires II using digital goats.
The setup was deliberately crude. Using the game’s scenario editor, de Wynter built an operational NOT AND (NAND) gate and a one-bit perceptron. In this system, grass represented a zero value, bridges represented a one, and the goats acted as the signal carriers. When a gate fired, the bit-goats were removed from the track and new ones placed in the output rail. Essentially, the goats were moving binary data through a logic circuit.
A perceptron is the simplest form of a neural network, an algorithm designed to sort inputs into binary classes. While videos exist online of players constructing similar redstone circuits in Minecraft, no one claims those digital blocks are neurons in a thinking machine. Yet, de Wynter’s paper posits that the underlying technology is identical to what powers Claude, ChatGPT, and Copilot. The difference lies entirely in the presentation. When users interact with these models through a chat window, the perception of human-like traits emerges. Strip away the chat interface and replace it with moving goats, and that illusion of personality vanishes, even though the core logic remains unchanged.
“The point of the paper is to formally show that we anthropomorphise too readily, and that sometimes the claims we make with regards to LLM capabilities are too strong,” de Wynter told 404 Media. “It’s not an easy task, given that ‘human-like attributes’ is a bit of an abstract term.”
De Wynter has been playing Age Of Empires since its release in 1999. He chose the game because it feels sufficiently “alien” to highlight the gap between representation and interpretation, yet it is familiar enough to drive the point home effectively.
The Anthropomorphism Trap
According to de Wynter, the problem of assuming large language models possess human traits begins before scientific research even starts. He reviewed 315 computer science papers released over the last two years and found that 57 percent began with the assumption that these models have human-like attributes.
“What is common to some of these studies […] is that they test and ascribe blanket human-like properties (e.g., anxiety or morality) to these LLMs while considering them the central subject of the experiment,” his paper states. “Regardless of these evaluations’ results being positive or negative, their core assumption–that LLMs possess anthropomorphic attributes–influences the experiment’s planning through (e.g.) the design of the test set, the interpretation of natural-language outputs, and even its null hypothesis. In turn, this directly impacts the conclusions made.”
De Wynter argues that we either assume tokens represent language to the model in the same way they do to us, or we assume that because an LLM outputs a relevant string, it must be understanding the concept or possessing empathy. This bias skews the entire experimental process.
“We either start by thinking that tokens represent language to LLMs the same way they do to us, or that because an LLM outputs a relevant string, it must be understanding the concept/having theory of mind/empathy/etc,” de Wynter said. “This goes both ways: we could also assume that LLMs are blobs of weights just floating about on a GPU, but that would not help explain some skills that they are shown to have.”
Some users perceive their interactions with LLMs as human because the interface mimics conversation. De Wynter proposes we stop assuming LLMs behave like humans simply because they were trained on natural language. Instead, we should run experiments that reveal the models as they truly are, not as we wish them to be.
He notes that the widespread anthropomorphization of LLMs among executives, scientists, and the public is becoming an intractable problem. “This is why I used the goats: 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 makes them what they are perceived as,” de Wynter explained. “Goats and AoEII break the perception that these machines are human.”
De Wynter is not ruling out the possibility that LLMs have some form of consciousness, but he argues that defining it as a binary construct is flawed. He questions why we are so surprised when bumblebees solve problems, noting that we often ascribe consciousness based on human-like criteria rather than objective reality. “It’s quite relative and we do tend to go for something human-like when we evaluate/define it, where, in reality, LLMs are things we have never seen before,” he said.
Marketing and Perception
The core issue is that capabilities and claims are strongly tied to marketing, as these models are products. De Wynter suggests that appropriate disclosures and good alignment techniques are necessary to mitigate these perceptions. He worked extensively on how users perceive LLMs, noting that people get attached when a model appears to have warmth or personality. He compares this to not getting attached to a toaster, but getting attached to characters on a movie screen.
However, consumers buy more objects-whether toasters, phones, or LLMs-when they can empathize with them. De Wynter states that the industry’s reliance on this empathy is a strategic choice.
This marketing strategy aligns with statements from industry leaders. OpenAI CEO Sam Altman has repeatedly implied that building LLMs is a path to creating an AI god. Ilya Sutskever, a former OpenAI board member and scientist, has often spoken openly about viewing the company’s LLM as a god-like consciousness. Anthropic CEO Dario Amodei has told The New York Times that he cannot be certain if AI is conscious or not. For an industry currently facing significant financial losses, this narrative serves as effective promotion.
Key takeaways
The perceived sentience of large language models is largely an illusion created by the conversational interface, not evidence of actual consciousness.
Adrian de Wynter’s experiment proves that identical underlying logic can be implemented in absurd contexts (like digital goats) without triggering human-like perceptions.
Scientific research is biased by anthropomorphism, with over half of recent papers starting with the assumption that LLMs possess human traits like anxiety or morality.
Industry leaders often promote the idea of AI consciousness as a marketing strategy, despite the models being products rather than sentient beings.
