If ChatGPT has consciousness, then so does the goat in Age of Empires 2! Microsoft scientists even created GoatGPT.

Microsoft researchers used the goat-trigger mechanism from Age of Empires 2 to build a neural network, thereby demonstrating that models like ChatGPT lack consciousness—its humanlike qualities are nothing more than cognitive illusions created by the text interface and commercial marketing.

Does AI really have consciousness?

If you think today’s ChatGPT and Claude have consciousness, then the goats in Age of Empires 2 do too.

With the widespread adoption of large language models (LLMs), you often see arguments claiming that AI has self-awareness, which prompts both science-fiction writers and frontline researchers to think deeply.

Science-fiction writer Ted Chiang once wrote that blindly believing that language models possess consciousness is like believing that every time you open a Microsoft Word file, you awaken multiple individuals hidden in the conversation history.

To puncture this cognitive bubble, a Microsoft AI scientist decided to take action—directly building a computational network with the same underlying logic inside the classic real-time strategy game Age of Empires 2, using the most absurd approach to deliver a shock to the mind.

Image source: Game screenshot of Age of Empires 2 from Crypto City’s edit

Building a neural network with the goats of Age of Empires 2

According to 404 Media, Microsoft chief scientist Wynter (Adrian de Wynter) used the Age of Empires 2 custom map editor to successfully construct a basic neural network within the game. The research was published in an academic paper titled If Large Language Models Have Human Traits, Then Age of Empires 2 Has Them Too.

He used the game’s custom scenario trigger function, combining elements such as wooden walls, grass, and bridges, and using the goats in the game as a signal-transmission medium, to build a NAND gate in the computational logic.

Image source: Research paper—Microsoft scientist builds a neural network with the goats of Age of Empires 2

During gameplay, grass represents the number 0, bridges represent the number 1, and the goats play the role of bits. When the logic gate is triggered, the goats serving as the input signal are removed, and new goats appear on the output track.

In this way, he implemented a 1-bit perceptron in the game—one of the most basic building blocks of modern neural networks—showing that the game has the potential to perform any theoretical computation.

The goats are running, but the perception of human traits disappears

In an interview, Wynter said that the purpose of this experiment was to formally show that humans anthropomorphize things far too easily, and that sometimes claims about the capabilities of large language models are overstated.

He pointed out that this goat-driven network is essentially identical in underlying technology to the ones that power chatbots like ChatGPT, Claude, or Copilot; the difference is only in complexity and scale.

Image source: Research paper—Microsoft scientist builds a neural network with the goats of Age of Empires 2

Wynter’s experiment highlights the contradictory relationship between a computational substrate and human cognition. When the same computational logic is presented in a web browser or a voice dialogue, companies like Anthropic claim the model “has a constitution” and may even feel anxious.

However, when the same computational process is moved onto the game’s virtual grass and the results are presented by goats running around within enclosures, the observer’s perception of human traits completely disappears. This proves that the AI humanlike traits people feel depend largely on the text-based chat interface and the observer’s psychological expectations—there is nothing mysterious about the underlying technology itself.

Anthropomorphic bias is widespread in scientific research

In fact, the tendency to anthropomorphize technology is deeply rooted in academia, and it exerts influence even before research begins.

Wynter reviewed more than 300 computer science papers published over the past two years and found that 57% of them, at the very beginning of experiment design, directly assumed that large language models possess human traits such as anxiety or morality, and treated these as the core subjects of the experiment.

In studies that pre-assumed the model had human traits, as many as 77% ultimately concluded that such traits indeed exist. This reflects a serious confirmation bias: when researchers planned experiments, designed test sets, and interpreted natural-language outputs, they were influenced by the initial anthropomorphic assumptions, which directly shaped the final scientific conclusions.

He called on the scientific community to use null hypotheses that do not presuppose human traits in order to see the true face of AI.

Do you think AI is like a human? Isn’t that basically brainwashing?

We often say AI has hallucinations, and humans also have hallucinations about AI due to excessive projection—this is closely tied to consumer psychology and the business strategies of AI giants.

Wynter mentioned that research shows when consumers can resonate with or empathize with a product, they are more likely to spend money on it, and this principle shows up in toaster ovens, smart phones, and paid subscription services for large language models.

The AI giants building these models did not stop this anthropomorphic trend; they even guided it, intentionally or unintentionally.

For example, OpenAI CEO Sam Altman has repeatedly implied that building large language models is the path to creating god-like AI; former OpenAI board member and scientist Ilya Sutskever also discussed with employees viewing the company’s models as god-like consciousness; Anthropic CEO Dario Amodei told the media that he cannot be sure whether AI has consciousness.

These commercial narratives and marketing tactics, circulating through policy-making, product promotion, and user perception, work together to sustain the illusion that AI has consciousness.

Further reading:
Utada Hikaru and Harari in conversation: When AI copies art, love, and consciousness, do human inspirations still have value?

AI harm exceeds 400 cases! BBC tracks people who overly trust AI and develop delusions of being harmed