To what degree is Artificial Intelligence capable of being (inter)culturally competent?
Résumé
The massive adoption of Chat GPT in 2023 has propelled into the spotlight the capacity of chatbots based on Large Language Models (LLMs) to produce seemingly authentic utterances in natural lan-guage and to adapt to contextual conversational cues, in a way reminiscent of human dialogue. At the same time, LLMs have markedly improved translation software, bringing viable real-time technological solutions to allow “bots” to “chat” in various languages: the ecosystem of intercultural communication is changing rapidly. However, human communication is shaped by many factors, including cultural knowledge associated with different social groups, and the way individuals seek to play out various identities in their face-to-face interactions (Goffman, 1959). Corpus-trained AI logically reflects the “culture”(s) of its training data, including associated biases (Shrestha & Das, 2022). Early studies report that training AI to differentiate between national corpora can indeed help optimise algorithmic performance (Messner, 2022). But LLMs trained on massive corpora are exposed to a huge mixture of cultural elements, which humans would typically associate with different groups. Does this make the AI culture-neutral, culture-omniscient, or simply culture-blind? In this fast-moving area, this paper aims to provide participants with up-to-date knowledge, based on a review of the relevant literature in the fields of intercultural communication and computer science, not only of the current limits of AI in intercultural communication, but also the degree to which it might be possible for chatbots to become more “culturally competent”, along with the potential ethical implications of this.