Who Has the Last Word? Lessons from Using ChatGPT to Develop an AI-based Spanish Writing Assistant

TY - JOUR

T1 - Who Has the Last Word?

T2 - Lessons from Using ChatGPT to Develop an AI-based Spanish Writing Assistant

AU - Tarp, Sven

AU - Nomdedeu-Rull, Antoni

PY - 2024/2

Y1 - 2024/2

N2 - This article deals with the complex relationship between human intelligence and so-called artificial intelligence in the context of an ongoing project to develop a writing assistant for Spanish learners, both native and non-native. The authors have used ChatGPT to generate validation data to assess the performance of the language model on different parameters before testing it with real users. The article describes how they approached the generation of validation data, what they learned along the way, and what the results were. It first introduces the project and describes its main phases. It then explains the criteria the authors used to determine the types of problems to be covered by the validation data, and how they instructed the chatbot to generate this data. Finally, it summarises the main lessons they learnt from working with the chatbot and some of the challenges they faced in getting it to work properly. The description is accompanied by numerous examples. By engaging with the chatbot in a critical and constructive way, and by establishing close interdisciplinary collaboration with IT specialists, the authors conclude that the key challenge is to demonstrate in practice that humans, not the chatbot, are the masters. In this context, they argue that generative AI language models are not here to replace us, but to help us produce faster and with higher quality to meet our growing and increasingly diverse demands for a better life.

AB - This article deals with the complex relationship between human intelligence and so-called artificial intelligence in the context of an ongoing project to develop a writing assistant for Spanish learners, both native and non-native. The authors have used ChatGPT to generate validation data to assess the performance of the language model on different parameters before testing it with real users. The article describes how they approached the generation of validation data, what they learned along the way, and what the results were. It first introduces the project and describes its main phases. It then explains the criteria the authors used to determine the types of problems to be covered by the validation data, and how they instructed the chatbot to generate this data. Finally, it summarises the main lessons they learnt from working with the chatbot and some of the challenges they faced in getting it to work properly. The description is accompanied by numerous examples. By engaging with the chatbot in a critical and constructive way, and by establishing close interdisciplinary collaboration with IT specialists, the authors conclude that the key challenge is to demonstrate in practice that humans, not the chatbot, are the masters. In this context, they argue that generative AI language models are not here to replace us, but to help us produce faster and with higher quality to meet our growing and increasingly diverse demands for a better life.

KW - Spanish writing assistants

KW - language didactics

KW - generative AI chatbots

KW - training of language model

KW - human-assisted intelligence

UR - http://www.scopus.com/inward/record.url?scp=85180488580&partnerID=8YFLogxK

U2 - 10.5209/clac.91985

DO - 10.5209/clac.91985

M3 - Journal article

SN - 1576-4737

VL - 97

SP - 309

EP - 321

JO - Círculo de Lingüística Aplicada a la Comunicación

JF - Círculo de Lingüística Aplicada a la Comunicación

ER -