Are Chat-bots Truly AI?

Written by Tom Scott

Edited by Alfred Tyrrell

Since the public release of chat GPT in November 2022 the usage and capabilities of so-called AI have exploded. Whilst there is no denying that chat GPT and other related chatbots are a technological revolution, their claim that they are indeed powered by artificial intelligence is a hotly debated topic.

Taking a reasoning-based view of AI, to be considered as an AI a program must be able to solve reasoning-based problems alike to a human mind. In the current state of chatbot development, they cannot be considered to have intelligence.

Chat GPT, along with other chatbots, is a large language model (LLM), a probabilistic model which, using the context of a question or prompt, generates text via a model, predicting which word is most likely to come next. To do this, LLMs are trained on massive online data sets of human written texts, allowing the models to both accurately respond to inputted prompts and to also reel off comprehensible and deceptively fluent sounding sentences to an extent to which human written and chatbot text are mostly indistinguishable. 

This, however, is where the similarity between chatbots and the human mind stops. Whilst these models can accurately format the syntax of human language, they have no knowledge of the actual semantics of the words they are producing. Due to this, Emily Bender calls such chatbots ‘stochastic parrots’ in her highly influential 2021 paper, as whilst these chatbots may appear to have a grasp of the English language, in reality, they are just stitching together words most likely to go together based off of human writings. Due to not comprehending the semantics of the sentences it produces, the AI cannot be thought to perform complex reasoning in the way humans do. 

This can be illustrated further using Searle’s famous thought experiment “the Chinese room” in which there is a room containing a person with a book corresponding Chinese symbols with their responses. People slip letters with Chinese characters underneath the door of the room and after consulting the book the person writes the corresponding symbols onto another piece of paper and slides it back under the door. Due to these people passing their notes under the door and getting coherent responses back, it seems to them that there is a Chinese speaker in the room, but this is evidently not true. In the same way that the person in the room has no understanding of the Chinese language, neither do these chatbots of any language. Whilst it could be argued the matching of one symbol to another constitutes a base-level reasoning, this is not nearly complex enough to ascribe intelligence to these programs. If it were to be considered intelligence why would lower forms of language modelling such as the predictive text on your Nokia brick not be considered AI? 

Undoubtedly, the creation of these chatbots is a huge leap towards the establishment of a truly intelligent artificial system. However, the use of the phrase AI to describe these models at this point in time is a flippant one–a buzzword slapped onto an undeniably innovative model to generate excitement around it.