What is it like to be a bot? The world according to GPT-4

Abstract

The recent explosion of Large Language Models (LLMs) has provoked lively debate about “emergent” properties of the models, including intelligence, insight, creativity, and meaning. These debates are rocky for two main reasons: The emergent properties sought are not well-defined; and the grounds for their dismissal often rest on a fallacious appeal to extraneous factors, like the LLM training regime, or fallacious assumptions about processes within the model. The latter issue is a particular roadblock for LLMs because their internal processes are largely unknown – they are colossal black boxes. In this paper, I try to cut through these problems by, first, identifying one salient feature shared by systems we regard as intelligent/conscious/sentient/etc., namely, their responsiveness to environmental conditions that may not be near in space and time. They engage with subjective worlds (“s-worlds”) which may or may not conform to the actual environment. Observers can infer s-worlds from behavior alone, enabling hypotheses about perception and cognition that do not require evidence from the internal operations of the systems in question. The reconstruction of s-worlds offers a framework for comparing cognition across species, affording new leverage on the possible sentience of LLMs. Here, we examine one prominent LLM, OpenAI’s GPT-4. Inquiry into the emergence of a complex subjective world is facilitated with philosophical phenomenology and cognitive ethology, examining the pattern of errors made by GPT-4 and proposing their origin in the absence of an analogue of the human subjective awareness of time. This deficit suggests that GPT-4 ultimately lacks a capacity to construct a stable perceptual world; the temporal vacuum undermines any capacity for GPT-4 to construct a consistent, continuously updated, model of its environment. Accordingly, none of GPT-4’s statements are epistemically secure. Because the anthropomorphic illusion is so strong, I conclude by suggesting that GPT-4 works with its users to construct improvised works of fiction.