Introduction

In the evolving discourse on artificial intelligence (AI), the quest for strong AI (Searle, 1980) remains paramount. This chapter seeks to elucidate the foundational assumptions that underpin philosophers’ and computer scientists’ assertions about the prerequisites for achieving strong AI. The current trajectory in AI, notably within computer science, is increasingly influenced by Google Brain's white paper on the Transformer architecture (Vaswani et al. 2017), a groundbreaking deep-learning model that reshaped the scientific landscape of Natural Language Processing (NLP) and Machine Learning (ML). The Transformer architecture forms the backbone behind the development of Large Language Models (LLMs) such as OpenAI's ChatGPT and Google's BERT. The Transformer architecture emphasizes the importance of handling sequential data through self-attention mechanisms, reflecting the human ability to emphasize certain textual elements over others based on context. In this chapter, I refer to these LLMs as ‘Statistical AI’, since Transformer architecture relies heavily on statistical methods when modelling these self-attention mechanisms. LLMs have not only achieved immediate commercial success and cultural impact for the tech industry, but the world's leading computer scientists like Ilya Sutskever also believe that they could be ‘slightly conscious’ (Sutskever 2022).

Contrasting this prevalent view, the chapter turns to the critiques offered by philosophers Noam Chomsky and Robert Brandom. Chomsky's internalist critique underscores the limitations of purely statistical models in capturing authentic human linguistic practices. Statistical models of reasoning have great practical use cases but are irrelevant to science. They are not proper models of reasoning, for human linguistic practices do not require agents to look up probability tables of what word should be used in an utterance (Chomsky et al. 2023). Conversely, Brandom's externalist perspective emphasizes the need for AI to engage in autonomous discursive practices (ADPs). ADPs are practices that are regulated by norms that are implicit within the practice itself and not regulated by external factors (i.e. training data in an AI context). ADPs involve making inferences when deployed within the context of communication or actions with other agents, while also allowing for self-correction when compared to the norms the ADP is grounded on and even correcting norms themselves (Brandom 2006).