Microsoft: A Unique Experience Questions AI Consciousness

Since the emergence of ChatGPT, discussions around the consciousness of artificial intelligences have intensified. Can we really consider these systems as conscious, or are they merely sophisticated text prediction tools? Opinions among experts vary. Some see these models as the beginnings of human-like intelligence, while others regard them simply as advanced machines without true understanding.

In this context, Adrian de Wynter, a researcher at Microsoft, decided to make an original contribution to the debate. Rather than engaging in complex technical discourse, he chose a more accessible and provocative approach.

A Provocative Demonstration

Adrian de Wynter published an article with an evocative title: “If LLMs Have Human Attributes, Then Age of Empires II Does Too.” Through this demonstration, he seeks to highlight how some claims about AI consciousness can be misleading.

How to Determine if an AI is Conscious?

For de Wynter, there is no universal method to prove consciousness in an AI. Discussions on this topic often rely on interpretations and hypotheses that are difficult to verify. To illustrate this complexity, he conducted an original experiment with the virtual goats from the game Age of Empires II, released in 1999.

Using the game's scenario editor, de Wynter transformed these goats into computational elements, creating simple logic gates such as NAND, XNOR, and AND. These gates are fundamental to modern computer systems. By using these mechanisms, he was able to build a perceptron, a basic form of artificial neural network, thus demonstrating that the goats could simulate certain principles of modern AI.

The Results of the Experiment

The virtual goats were used to create a complex logical system. De Wynter dubbed them the “bit-goats,” a contraction of the words “bit” and “goat.”

What Exactly Does the Researcher Want to Prove?

The experiment highlights the concept of “substrate,” meaning the medium on which an AI is built. Today, we often imagine massive computing infrastructures for models like ChatGPT. However, de Wynter shows that a similar system could theoretically be implemented on very different substrates, such as virtual goats.

If the same logical mechanisms can be reproduced in a video game, then the substrate itself is not what defines a system's capabilities. Just because an AI operates on sophisticated servers does not automatically mean it possesses particular properties. Conversely, a system based on digital goats does not necessarily change its behavior.

So, Does AI Have Human Qualities or Not?

According to de Wynter, when interacting with ChatGPT, the experience is smooth and convincing, giving the impression of an intelligent entity. Language is often associated with consciousness. When a machine expresses itself convincingly, we tend to attribute human qualities such as understanding or emotions to it.

But if ChatGPT's responses came from a system of virtual goats, our perception would be different, even though the model remains unchanged. The questions and answers would be identical; only the presentation would change. This shows that our expectations influence our perception of AIs.

The Consciousness of AIs: Perhaps a Misguided Question

De Wynter's experiment does not seek to prove or refute the consciousness of LLMs but to show that our positions are often influenced by prior assumptions. Believing that a model is conscious before studying it can bias the analysis, just as the opposite can occur.

The current debate suffers from this difficulty, as experiments to prove or refute human qualities in AIs often rely on vague concepts. The image of the goats from Age of Empires II acts as a mirror revealing how much our perception of an AI depends on the context in which it is presented.