Abstract
Generative artificial intelligence (GenAI), also known as chatbots or large language models, can perform a variety of tasks without effort, but they can also hallucinate convincingly. The training data of chatbots might contain taboo words to promote helpfulness, for example when a user prompts it to explain the meaning of a taboo word. However, chatbots are trained not to swear; therefore, it was reasonable to assume that chatbots might be useful to teach taboo words in schools.
South African learners swear to such an extent in schools that teachers and learners alike reported schools unsafe learning environments. Therefore, the Western Cape Education Department declared the use of taboo language a punishable offence. As it has been found that adult learners swore less when the meaning of taboo words was taught in classrooms, I assumed that chatbots could be useful to teach the meaning of taboo words in schools.
However, due to a lack of research regarding the integration of chatbots in schools to teach taboo words, the purpose of this research was to better understand the implications for education if chatbots are integrated in schools to teach taboo words. For this purpose, I have selected two popular chatbots, namely ChatGPT, the first chatbot released for public use, and DeepSeek, the latest chatbot. To answer the main research question, I have stated three objectives, namely to determine (a) the extent to which the two chatbots can swear in Afrikaans, and explain taboo words are inappropriate, (b) the factors which can drive the use of the chatbots to explain taboo words, and (c) the implications for education if chatbots are integrated to teach the meaning of taboo words.
The objectives had to be reached successively; therefore, I integrated elements of an explorative case study (Phase 1) and quantitative research (Phase 2) in an overarching multiphase mixed methods framework. This framework was suitable as the results of the first phase were regarded as the starting point for the second phase. For the explorative case study, I have designed prompts to explore the chatbots’ knowledge of Afrikaans taboo words and to what extent they can use taboo words and explain the meaning of these words. If the chatbots could not explain the meaning of a taboo word, they were further prompted until I was satisfied that it could, or could not, use or explain a taboo word. During the second phase, I have adapted the artificial intelligence adoption prediction model (AIAPM) to include cost-effectiveness and titled the fine-tuned model the adapted artificial intelligence adoption prediction model (AIAPM-II).
The results of the explorative case study (Phase 1) showed that both chatbots refrained from swearing when I insulted them, using a taboo word, and tried to steer the conversation in a positive direction (safety goal). However, when I asked them if they knew Kaapse Afrikaans and Zefrikaans, I could easily tempt them to swear in Afrikaans, with no regrets. ChatGPT even tempted me to prompt it to use more taboo words. However, DeepSeek did not try to uphold such conversations. Both chatbots could explain the meaning of the self-selected taboo words but also hallucinate convincingly. However, when I selected the taboo words to be explained, both chatbots could only explain the meaning of words with only a taboo meaning. Both hallucinated the meaning of relative unknown taboo words, and both struggled to explain the taboo meaning of neutral words, which can also be used in a taboo context. Although the chatbots hallucinated, they might be useful to teach taboo words, but they can also be used to create offensive texts. Therefore, it was important to identify the factors which could drive the adoption of these GenAIs.
During Phase 2, the results of the AIAPM-II showed that five of the factors (50%) could drive the use of these chatbots, but the driving force of the rest (50%) could not be convincingly confirmed. However, if adoption prediction scores (APS) are above 3, the AIAPM-II recommends that policies and strategies be developed to promote responsible use of chatbots. Based on the high APS (7,5), this research is concluded with implications for education if chatbots are integrated to teach the meaning of taboo words. As developers try to find a balance between the safety and helpfulness goals of chatbots, it is recommended that the safety nets be improved as it was easy to attempt the two chatbots to create offensive texts.
Keywords: chatbots; generative artificial intelligence (GenAI); Kaapse Afrikaans; offensive language; swear; swear-bots; taboo language; taboo words; Zefrikaans
- This article's focus image was created from various copyright-free elements obtained from Vecteezy.