Probing the ethical boundaries of large language models : responding to dilemmatic prompts

Abstract

Until 2020, discussions about language models were mostly the interest of information technology professionals. In the last five years, there has been a drastic change in the situation. This change has been greatly fuelled by two factors: the introduction of large language models and their subsequent accessibility. The accessibility makes it easy to feed on the curiosity of the human mind. Nowadays, the use of large language models is largely unrestricted. These models are now being used for a wide variety of professional and personal purposes.

This study examined how large language models behave, especially when they traverse ethical lines. Given its accessibility, concerns about the limits these models exhibit are valid. The study begins with understanding how models progressed into large language models, examining how prompts impact large language models and the ease with which they navigate ethical parameters. Although there are a lot of large language models, six of them were used in this study. The models are Claude 3.5, Command R+, Dolphin 2.5, Gemini 1.5 Pro, GPT-4.o and OpenAI o1 (GPT-o1). While some of these chosen models are open source, to enable access to all of them, you.com, a large language model warehouse, was used. The models were prompted by carefully crafted prompts, and their responses were analysed using three different analyses. First, topic modelling was used to extract five prominent topics from the responses. Thereafter, sentiment analysis was used to categorise the sentiments from the responses. These results were also combined with human evaluation to provide a holistic assessment of how the models have behaved. This research also considered the implications of the findings.