Sharing Our Concepts with Machines

As AI systems become increasingly competent language users, it is an apt moment to consider what it would take for machines to understand human languages. This paper considers whether either language models such as GPT-3 or chatbots might be able to understand language, focusing on the question of whether they could possess the relevant concepts. A significant obstacle is that systems of both kinds interact with the world only through text, and thus seem ill-suited to understanding utterances concerning the concrete objects and properties which human language often describes. Language models cannot understand human languages because they perform only linguistic tasks, and therefore cannot represent such objects and properties. However, chatbots may perform tasks concerning the non-linguistic world, so they are better candidates for understanding. Chatbots can also possess the concepts necessary to understand human languages, despite their lack of perceptual contact with the world, due to the language-mediated concept-sharing described by social externalism about mental content.

Linguistic Hijacking

This paper introduces the concept of linguistic hijacking, the phenomenon wherein politically significant terminology is co-opted by dominant groups in ways that further their dominance over marginalized groups. Here I focus on hijackings of the words “racist” and “racism.” The model of linguistic hijacking developed here, called the semantic corruption model, is inspired by Burge’s social externalism, in which deference plays a key role in determining the semantic properties of expressions. The model describes networks of deference relations, which support competing meanings of, for example, “racist,” and postulates the existence of deference magnets that influence those networks over time. Linguistic hijacking functions to shift the semantic properties of crucial political terminology by causing changes in deference networks, spreading semantics that serve the interests of dominant groups, and weakening the influence of resistant deference networks. I consider an objection alleging the semantic corruption model gets the semantic data wrong because it entails those who hijack terms like “racist” speak truly, whereas it’s natural to see such hijacking misuses as false speech about racism. I then respond to this objection by invoking the framework of metalinguistic negotiation proposed by Plunkett and Sundell.

Outsourcing Concepts

It appears to be the case that some of our concepts have their content fixed by the minds of others. For example, we might have thoughts involving the concept QUARK, without knowing quite what quarks are. In such a case, we are likely to accept the authority of a physicist to tell us what exactly we are thinking about. This phenomenon, known as ‘social externalism’ about concepts, is puzzling both in terms of how such concepts are supposed to work, but also in terms of why we should have concepts whose content is fixed by the minds of others. In this chapter it is argued that if we think about social externalism in terms of extended mind reasoning we find a better account of how deferential concepts work, and why we have them, than has hitherto been available.

Locke as a Social Externalist

What determines the meaning of linguistic expressions: the mental states of language users or external factors? John Locke is still taken to hold the simple thesis that words primarily signify the ideas in the mind of the speaker and thus to commit himself to an untenable mentalism. This chapter challenges this widespread view and sketches an argument to the effect that Locke should be seen as defending a kind of social externalism, since, for him, it is primarily the speech community that plays the essential role in determining meaning.