In a thought-provoking essay, science fiction writer Ted Chiang challenges the notion that Large Language Models (LLMs) possess consciousness. He draws an intriguing parallel between LLMs and Microsoft Word, suggesting that the idea of consciousness in AI is absurd. This concept is further explored by Microsoft AI researcher Adrian de Wynter, who takes the absurdity a step further by building a basic neural network within the classic real-time strategy game, Age of Empires II (AoEII).
De Wynter's paper, titled 'If LLMs Have Human-Like Attributes, Then So Does Age of Empires II', presents a unique approach to understanding the anthropomorphism of AI. He constructs an operational NOT AND (NAND) gate and a 1-bit perceptron within the game, using goats as the 'bits'. This creative experiment highlights the tendency to anthropomorphize AI, even when the underlying technology is the same.
The author argues that the problem of anthropomorphizing LLMs begins before scientific research. A review of 315 computer science papers revealed that 57% started with the assumption that LLMs possess human-like traits. This assumption influences the experiment's planning, interpretation, and conclusions, leading to a biased understanding of AI capabilities.
De Wynter emphasizes the importance of focusing beyond the interface we use to interact with LLMs. He believes that the perception of human-like traits in AI is a result of our interface with it, rather than the technology itself. The author uses the example of goats in AoEII to illustrate how removing the chat panel can break the perception of human-like intelligence.
The essay also delves into the marketing aspect of AI, where executives, scientists, and the public anthropomorphize LLMs, leading to an intractable problem. De Wynter suggests that appropriate disclosures and alignment techniques are necessary to mitigate this issue. He argues that people's attachment to AI is similar to their attachment to movie characters, and that marketing plays a significant role in shaping these perceptions.
In conclusion, the essay encourages a critical examination of our assumptions about AI consciousness and the anthropomorphism of LLMs. By exploring the absurdity of AI consciousness and the impact of our interactions with AI, we can gain a deeper understanding of the technology and its limitations.
