Large language models: scaffolded, not extended

Abstract

The article considers the possibility of extended artificial intelligence, which draws on paral­lels between the alleged cases of human cognitive extension according to the extended mind thesis and large language models capable of augmenting their responses on the basis of in­formation from external sources. Since the idea of extended AI draws its inspiration from the hypothesis of A. Clark and D. Chalmers, the article evaluates the arguments for and against the possibility of human cognitive extension. The author highlights the important functional dissimilarities between neurally encoded dispositional beliefs and information stored in external sources. It is argued that the idea of the extension of the mind into the world is highly problematic unless there is tight integration between the brain and exter­nal data stores. It is noted that similar problems can arise for large language models equipped with the technology of retrieval-augmented generation but which lack continuous synchronization with an external data store. In this regard, it is suggested that theorizing about agents’ relations to external information stores may need more moderate hypotheses, ones consistent with the currently dominating methodological assumptions of the cognitive sciences.