Large Language Models (LLMs), Language Representation Models, and Foundation Language Models have become the focal point of both excitement and apprehension in the ever-growing field of Artificial Intelligence (AI), specifically within the subfield of Natural Language Processing (NLP). Their remarkable ability to generate seemingly rational and contextually relevant text has led to numerous myths, misunderstandings, and concerns. This research paper aims to elucidate the common fallacies and misconceptions surrounding LLMs, shedding light on their capabilities, limitations, and ethical implications. To a large extent, the answers given by these models have a lot of truth to them, however, the responses that they produce should not be taken as gospel.

Since the advent of LLMs, such as GPT-3 and GPT-4 (ChatGPT) [1], BERT [2], Bard [3], LLaMA [4], BLOOM [5], and so on, there has been a wave of enthusiasm and trepidation in the realm of AI. Large corporations have even created industry- and company-specific LLMs for specific use cases, for example, BloombergGPT [6] for financial NLP applications. These models, trained on vast datasets (almost the whole World Wide Web), exhibit human-level fluency in language understanding and generation. However, as their prevalence increases, so does the misinformation, assumptions, misapprehensions, overinflation, and conjecture surrounding them. This paper undertakes the task of clarifying the most prevalent misconceptions about LLMs.

The original contribution of this paper is that it is a pioneering formalised qualitative academic study of the fallacies surrounding LLMs, applying the principles of design thinking. The caveat being that the claim is that this research team is not the first to attempt to discuss some of the misconceptions, but we are amongst the first to provide a ratified analysis.

Thus, this research endeavours to provide a comprehensive exploration of the prevalent misconceptions and fallacies associated with LLMs. The study is targeted at demystifying these often perplexing models, offering a clear and concise understanding of their capabilities, limitations, and ethical implications. By dispelling common myths, clarifying misconceptions, and substantiating findings with empirical evidence, this study aims to empower individuals, professionals, and decision-makers to make informed judgments and better leverage LLMs in their respective fields.

While LLMs are sparking a revolution right now in the world of AI and its real-world applications, it is a tool to stay, and a tool that can be used for immense good and improving the world. Although it lacks specific subject matter expertise, it can provide easy upskilling in diverse fields of study, and greatly assist in human day-to-day tasks. However, the foremost takeaway from this research is that the responses given from LLMs should not be confused with Artificial General Intelligence (AGI), and that an LLM is self-aware. Information provided should be examined carefully, and perused with caution. In critical applications, such as healthcare and financial advice, counsel should be sought from certified experts. An over-reliance on LLMs can be dangerous and have disastrous implications.

It is also worth noting that the response of an LLM is only as good as the prompt it is fed; thus, the emerging field of prompt engineering seeks to teach the art of asking the right questions from these LLMs.

Specifically speaking on emotional intelligence (point 7 in subSection 2.1), it would be interesting to see Psychology and Computational Neuroscience research around quantifying how emotionally unintelligent LLMs are, and defining metrics to measure and evaluate this for any LLMs. This would be a step forward in paving the way for AGI.

The authors would like to declare that all authors contributed equally, and there are no conflicts of interest.