Arianna Falbo and Travis LaCroix
In linguistics, code-switching describes the use of two (or more) languages (or language varieties) within a single text, conversation, or utterance. This purely linguistic sense of code-switching has been widely discussed in the context of artificial intelligence (AI)—especially natural language processing (NLP). AI/NLP research typically concentrates on the morphosyntactic features of code-switched linguistic data at the sentential (or sub-sentential) level. However, there is a broader characterisation of code-switching phenomena which is primarily social in nature—namely, cultural code-switching. This broader notion ‘requires that one change how one behaves, talks, and presents oneself as a response to a change in cultural context’ and thus involves ‘a much more profound shift than toggling between languages or dialects does’ (Morton, 2019, p. 76).
Cultural code-switching is ubiquitous: it is used to switch between formal and informal contexts; to wield power and to signal asymmetric authority relations (Popa-Wyatt and Wyatt, 2018; Tirrell, 2012); and, it can help build and reinforce community by signalling in-group affiliation (Dembroff and Saint-Croix, 2019). Since cultural code-switching takes account of the socio-pragmatic aspects of communication, it includes extra-linguistic cultural artefacts (in addition to linguistic ones). Thus, instead of the narrow sense of altering some linguistic features of communication, we take code-switching more broadly to also include non-linguistic communicative signals which convey information—e.g., about oneself and one’s relation to others in a given social context.
As language technologies continue to improve and become more pervasive in society, it will be increasingly essential that they are able to fluidly interact with humans. However, a significant part of this interaction involves non-linguistic communication—i.e., the additional features of communication that are relevant for cultural code-switching. Because engagement in cultural code-switching is especially common for minority communities (e.g., to conform to majority-group norms), this opens the door to exacerbate the extant biases of these technologies, which typically arise from dominant-group conventions (Cave and Dihal, 2020). In order to mitigate these biases it is important for researchers in AI to take cultural code-switching into account in NLP applications. This is particularly true when these systems are deployed in the world or are otherwise integrated into consumer-facing technologies.
The goals of this paper are two-fold. First, we motivate the importance of cultural code-switching in research on the ethics of AI. Despite its ubiquity and, more importantly, the unique value and costs that cultural code-switching has for those from marginalised backgrounds, surprisingly little attention has been given to these phenomena in the philosophical literature on AI. This paper bridges this gap by defending the need to investigate cultural code-switching capacities in AI systems (AIS).
Second, we explore a series of ethical consideration that arise when bringing issues of cultural code-switching to bear on AIS. Drawing upon literature in feminist philosophy of science—particularly on the value-ladenness of scientific inquiry (Longino, 1990; Anderson, 1995, 2004; Longino and Lennon, 1997; Yap, 2016)—and the growing philosophical literature on epistemic oppression (Fricker, 2007; Dotson, 2011, 2014; Pohlhaus, 2012; Medina, 2013), we canvas the potential moral and epistemic risks involved in implementing—or failing to implement—code-switching capacities in AIS. By leaving the socio-dynamic features of cultural code-switching unaddressed, AIS risk negatively impacting already-marginalised social groups by widening opportunity gaps and further entrenching social inequalities.
More generally, reflecting upon the need for cultural code-switching in AIS allows us to better enrich our understanding of these phenomena as influencing not only first-personal identity (Dembroff and Saint-Croix, 2019), but also broader systems of social coordination within and between groups.
Feminism, Social Justice, and AI 