Language models like ChatGPT have stirred up significant debate within the scientific community. Some experts view them as precursors to superintelligence, capable of revolutionizing or even threatening civilization, while others see them as little more than sophisticated versions of auto-complete. Traditionally, language proficiency was a reliable indicator of a rational mind, and prior to the emergence of these models, no language-producing artifact had demonstrated as much linguistic flexibility as a young child. Now, we are confronted with a perplexing philosophical dilemma: either the connection between language and mind has been severed, or a new form of mind has been brought into existence.
Engaging these new models in conversation can create the illusion of interacting with another rational being, but this impression should be approached cautiously. Cognitive linguistics highlights the abundance of ambiguous sentences in typical conversations, requiring our brains to continuously make assumptions about the speaker's intentions. While this mechanism is invaluable in a world where speakers have intentions, it has the potential to mislead when interacting with large language models. Achieving effective communication with a chatbot may, therefore, necessitate relying on this intention-guessing mechanism. However, perceiving chatbots as having human-like mental lives is not a sound theory for understanding their operation and may hinder hypothesis-driven science and lead to inappropriate AI regulation standards.
An example supporting this notion is a recent study suggesting that emotionally charged requests yield more effective language model responses than emotionally neutral ones. While reasoning as though chatbots had human-like mental states may aid in navigating their linguistic abilities, it should not be mistaken for a theory about their functioning. Anthropomorphic assumptions can impede scientific progress and lead to unsuitable AI regulation standards. Therefore, it is crucial to avoid being misled by false claims about the inner life of chatbots. The assumption that the psychological properties explaining human language capacity are the same as those explaining language model performance can make us vulnerable and blind us to the potential radical differences between humans and language models.
Thinking about language models poses various challenges, including anthropocentric chauvinism, which involves measuring psychological phenomena against the presumed standard of the human mind. Some skeptics argue that language models lack essential human psychological traits, like consciousness, leading to a dismissive attitude toward their capabilities. However, this approach overlooks the models' potential and proficiency in tasks such as summarization.
Critics often maintain that language models' competence is restricted to computing conditional probability distributions over words, a viewpoint stemming from anthropocentric chauvinism. Yet, this perspective fails to consider the broad array of competencies that language models can develop beyond mere next-word prediction.
Understanding how language models operate requires a comprehensive theory of their internal mechanisms, which is currently challenging due to the complex nature of information processing within high-dimensional vector spaces. Despite these difficulties, the models' complexity shouldn't be equated with the depth or nature of their intelligence.
Anthropomorphism and anthropocentrism persist as cognitive biases, sustained by a deep-rooted tendency known as essentialism, which leads us to believe that objects possess inherent, unobservable essences. This essentialist reasoning extends to our understanding of minds, where we tend to categorize entities into those with minds and those without. This all-or-nothing principle, rooted in vague concepts like consciousness, becomes inadequate when applied to artificial intelligence.
A more effective approach involves adopting a "divide-and-conquer" strategy to map the cognitive traits of language models without solely relying on the human mind as a guiding standard. This strategy allows for a more comprehensive and unbiased understanding of the capabilities and functioning of language models.
Drawing from comparative psychology, we should approach language models with the same inquisitive mindset applied to exploring the intelligence of diverse creatures. Despite being radically different from animals, studying animal cognition demonstrates how letting go of the all-or-nothing principle can advance understanding even in domains previously resistant to scientific inquiry. To genuinely assess the capabilities of AI systems, we must resist dichotomous thinking and comparative biases prevalent in studying other species.
Embracing the absence of a definitive truth regarding whether language models possess minds can help us avoid anthropomorphic assumptions that their exceptional performance equates to human-like psychological properties. Similarly, refraining from the anthropocentric assumption that deviations from human-like traits discredit a language model's competencies is crucial.
Given the novelty and peculiarity of language models, comprehending them necessitates hypothesis-driven scientific investigation into the mechanisms supporting their abilities. Remaining receptive to explanations not contingent on the human mind as a blueprint is vital in this endeavor.