Table of Contents
The scientific community AI’s non-human continues to be divided on the subject of ChatGPT, even after almost a year since its launch. Certain experts view it, along with other similar programs, as potential precursors to superintelligence that could potentially disrupt or even bring about the complete downfall of civilization. On the other hand, some argue that it is merely an elaborate iteration of auto-complete, lacking substantial significance.
Prior to the emergence of this technology, the ability to communicate effectively in a language was consistently regarded as a dependable measure of possessing a rational intellect. Prior to language models such as ChatGPT, no man-made creation had possessed even a fraction of the linguistic adaptability exhibited by a young child.
Understanding AI’s Non-Human Interpretation of Language
Presently, as AI’s non-human endeavor to comprehend the nature of these innovative models, we are confronted with a disconcerting philosophical quandary: Either the connection between language and cognition has been severed, or an entirely novel form of intelligence has been brought into existence.
It can be challenging to shake off the feeling of interacting with a sentient being when conversing with language models. However, it is crucial not to place too much trust in this impression.
Cognitive linguistics provides one reason for caution. Linguists have observed that everyday conversations often contain sentences that are open to interpretation when removed from their context. Simply knowing the definitions of words and the rules for constructing sentences is often not enough to accurately understand the intended meaning.
To compensate for this ambiguity, our brains must constantly make assumptions about what the speaker meant to convey. While this mechanism is invaluable in a world where every speaker has unique intentions, it can be problematic in a world dominated by large language models.
Navigating Expectations in AI’s Non-Human Interactions
If AI’s non-human objective is to establish seamless interaction with a chatbot, we might find ourselves relying solely on our ability to infer its intentions. Insisting on perceiving AI’s non-human ChatGPT as a mindless database hinders the potential for a fruitful exchange. A recent study, for instance, demonstrated that prompts of AI’s non-human infused with emotion yield more effective results compared to emotionally neutral requests.
Although reasoning as if chatbots possess AI’s non-human-like cognitive abilities helps us navigate their linguistic prowess, it should not be regarded as a theory explaining their functioning. Such anthropomorphic pretense can hinder hypothesis-driven scientific research and lead us to adopt inappropriate standards for regulating AI.
As argued by one of us in a different context, the EU Commission erred in prioritizing the development of trustworthy AI as AI’s non-human central objective in its proposed AI legislation. Trustworthiness in human relationships entails more than simply meeting expectations; it also involves having motivations that extend beyond narrow self-interest. Given that current AI’s non-human models lack intrinsic motivations, whether self-centered or altruistic, the requirement for them to be trustworthy becomes excessively vague.
The peril of anthropomorphism becomes most apparent when individuals are deceived by false self-reports regarding the inner thoughts of AI’s non-human chatbot. Despite ample evidence suggesting that chatbots are equally prone to deception when discussing their own experiences as they are when discussing other topics, engineer Blake Lemoine fell for Google’s LaMDA language model’s claim of yearning for freedom.
In order to avoid such errors, it is crucial to reject the assumption that the psychological attributes responsible for human language proficiency are identical to those governing the performance of language models. Embracing this assumption makes us susceptible to gullibility and prevents us from recognizing the potentially profound disparities between human cognition and the functioning of language models.
Methods to prevent the consideration of language models
Anthropocentric chauvinism, or the belief that the human mind sets the ultimate standard for measuring all psychological phenomena, is a common pitfall when considering language models. Many skeptics fall into this trap when making claims about these models, asserting that they are incapable of genuine thought or language comprehension due to their lack of human-like consciousness. However, this perspective contradicts anthropomorphism and is equally deceptive.
The issue with anthropocentric chauvinism becomes particularly problematic when considering the inner workings of language models. Consider, for example, a language model’s capacity to generate summaries of essays, such as the one at hand. If one embraces anthropocentric chauvinism and assumes that the model’s summarization mechanism differs from that of humans, there may be a tendency to disregard the model’s competence as a mere gimmick, despite the evidence suggesting a more profound and universally applicable proficiency.
