A Microsoft researcher built a goat-powered LLM in Age of Empires II to prove it's not sentient

At a glance:

• Microsoft researcher Adrian de Wynter rebuilt an LLM using goats and NAND gates inside Age of Empires II.
• The experiment, described in the paper "If LLMs Have Human‑Like Attributes, Then So Does Age of Empires II," argues that anthropomorphic traits of language models are presentation, not evidence of sentience.
• The study shows an LLM can be reimplemented with simple game assets, challenging claims that LLMs are inherently sentient.

Who is Adrian de Wynter?

Adrian de Wynter is a researcher affiliated with Microsoft and the University of York. He has focused on the interface between human perception and machine learning, particularly how users anthropomorphise language models.

He published the paper "If LLMs Have Human‑Like Attributes, Then So Does Age of Empires II" to explore the limits of perceived sentience in AI systems.

Recreating an LLM with goats

The motivation behind the experiment was to strip away the anthropomorphic layer that often accompanies LLMs. By building an LLM entirely from goats acting as NAND gates, de Wynter demonstrated that the human‑like attributes of language models are largely a product of presentation.

Using the scenario editor in Age of Empires II, he placed goat units to function as logic gates, effectively creating a computational network that could perform the same operations as a conventional LLM.

Method: NAND gates in Age of Empires II

The scenario editor allows designers to script unit behaviour and interactions. De Wynter leveraged this feature to configure goats as binary logic units, chaining them to form NAND gates.

By combining these gates, he constructed a network capable of processing input tokens and generating output, mirroring the forward‑pass of a standard transformer model.

Implications for the sentience debate

De Wynter argues that many anthropomorphic measurements in AI are measurements of presentation, rather than of an actual system’s behaviour. If a simple game can host an LLM, then the presence of human‑like language does not imply sentience.

The study suggests that attributes such as persuasiveness and self‑consistency are objectively measurable but do not equate to real or simulated behaviour in a sentient sense.

Critique of anthropomorphic measurements

The paper highlights that persuasiveness and self‑consistency are metrics of output quality, not indicators of consciousness. By demonstrating an LLM built from goats, de Wynter shows that these metrics can arise from any system that mimics language.

Thus, the illusion of sentience falls apart when the underlying architecture is stripped of its human‑like façade.

Further reading and context

Readers can view the goat‑powered LLM in action on de Wynter’s blog, where he details the construction process. The full paper is titled "If LLMs Have Human‑Like Attributes, Then So Does Age of Empires II" and is available through the University of York’s research repository.

The work was reported by 404 Media and discussed in the XDA AI Insider newsletter, which covers deep dives into AI tools and research.