Does Google Bard Understand ‘Itself’?
Sang-Gu Kang · Journal of Pan-Pacific Association of Applied Linguistics · 2023
Generative AIs such as Google Bard are known to be equipped with techniques and grammatical principles of human language based on a large corpus of text and code that allow them to generate natural-sounding language, and also identify and correct grammatical errors in humanwritten texts. Still, they are not perfect language generators, and this paper reports some interesting errors produced by Google Bard regarding interpretation of English plain and reflexive pronouns. Although Google Bard is clearly well aware of how to locate the antecedents for English pronouns, it consistently produces seemingly random interpretation errors when the prompt asks it to create a short story based on test sentences containing a plain or a reflexive pronoun. Since it is quite evident that Google Bard does not have an inborn system of grammatical principles including the binding principles, which are used to account for how human beings deal with pronoun interpretation, other approaches to processing human language need to be considered to explain Google Bard’s errors and the underlying mechanisms leading up to the errors. Thus, a processing account based on the Emergentist approach (O’Grady, 2005) and the Emergentist Reflexivity Approach model (Sperlich, 2020) is adopted, and the two separate levels of processing, that is, sentence processing and pragmatic processing, are introduced to account for Google Bard’s errors and the human language interpretation mechanism that might be underlying. These findings can not only shed light on how generative AIs might be analyzing human language but also contribute to our better understanding of human language and related theories.