Artificial Intelligence (AI) is a broad term that denotes the utilization of computers to emulate intelligent behavior with minimal human intervention [1]. Various domains exist within artificial intelligence, and one notable field is natural language processing.

Natural Language Processing (NLP) encompasses techniques for handling human language, utilizing a combination of semantic and statistical approaches. It involves the understanding of both spoken and written language, encompassing technologies related to voice recognition and text processing. Research fields within natural language processing include the identification of named entities, document filtering, and classification, automatic abstract generation, information extraction, voice recognition, sentiment analysis, opinion mining, monitoring social media reputation, orthographic and grammatical correction, text-to-voice systems, intelligent and optimized search, automatic response systems, personal assistants, automatic translation, and dialogue systems [2].

The swift advancements in natural language processing (NLP) have given rise to large language models capable of generating text at a level comparable to human quality and engaging in seamless, natural conversations. These models hold immense potential for diverse applications, including customer service interactions, the generation of creative content, and the facilitation of personal assistance.

Large language models (LLMs) represent artificial intelligence (AI) models grounded in deep learning, specifically neural networks, designed for text generation [3][4]. These models possess intricate underlying architectures and an extensive array of parameters, honed through training on vast volumes of existing documents. In contrast to older natural language processing approaches that rely on supervised learning for particular tasks, the majority of LLMs adopt semi-supervised approaches. Large language models [5] can function as an initial method for acquiring exploratory insights into a particular subject.

Large language models are constructed on the Transformer, a state-of-the-art neural network architecture with numerous parameters. The principal innovation of the Transformer lies in its self-attention mechanisms, allowing the model to better comprehend the relationships between different elements of the input. Large language models utilize a two-stage training pipeline for efficient data learning. In the initial pretraining stage, these models employ a self-supervised learning approach, enabling them to learn from extensive amounts of unannotated data without the need for manual annotation. This capability provides a significant advantage over traditional fully supervised deep learning models, as it eliminates the requirement for extensive manual annotation and enhances scalability. During the subsequent fine-tuning stage, large language models are trained on small, task-specific, annotated datasets to leverage the knowledge acquired during the pretraining stage. This allows them to perform specific tasks as intended by end users. Consequently, large language models achieve high accuracy on various tasks with minimal human-provided labels.

Large language models find application in diverse analytical steps throughout the development of new diagnostic assays. They prove valuable in tasks such as formulating growth media recipes for rare pathogens, troubleshooting unsuccessful polymerase chain reaction (PCR) assays, crafting PCR primers, and generating programming code. Additionally, these models play a crucial role in aiding the interpretation of data, particularly in scenarios involving rarely encountered pathogens [6].

The emergence of large language models (LLMs) has marked a significant turning point in recent years, with programs like ChatGPT and Google Bard making their debut in the public domain. The absence of financial barriers to entry for the average user, coupled with remarkable ease of use, has unveiled the vast potential of artificial intelligence to permeate various aspects of our lives [7]. Google Bard and ChatGPT have emerged as groundbreaking interactive chatbots [8], and their notable studies have been conducted from diverse perspectives since their inception.

ChatGPT, an artificial intelligence conversational system developed by OpenAI, a company dedicated to AI research and deployment, operates on the GPT-3.5 [9], one of the largest language models (LLMs) with over 175 billion parameters, making it among the most advanced to date. Sharing numerous capabilities with its predecessor, ChatGPT was trained on a diverse corpus of internet texts, encompassing approximately 570 GB of content from books, articles, and websites, spanning various subjects such as news, wikis, and fiction. Specifically tailored for conversational tasks, ChatGPT underwent fine-tuning using reinforcement learning from human feedback. Through this methodology, ChatGPT dynamically adjusts its behavior, making it highly adept at comprehending user intentions, generating text that closely mimics human language, and maintaining coherence throughout a conversation.

Google Bard AI is a text-based chatbot powered by natural language processing (NLP) and machine learning (ML) that generates real-time answers to questions. It is built upon the Pathways Language Model 2 (PaLM 2), a language model that is part of the lineage of Google's Language Model for Dialogue Applications (LaMDA) technology [10].

Google Bard and ChatGPT have been compared and analyzed from different perspectives. However, no comprehensive study has delved into the similarity of text generated by these two large language models. This area of study holds substantial importance, as the precise details of the data used to train these models are not publicly available. Therefore, a study that examines the similarity of comparable text generated by these models would provide valuable insights into their capabilities. The primary objective of this study is to analyze the topic-wise similarity of text generated by these models. To achieve this goal, the study will focus on oncology, a medical subspecialty dedicated to the investigation, diagnosis, and treatment of individuals with cancer or suspected cancer.

The novelty of this study lies in its capacity to provide valuable insights and enhance our understanding of the capabilities and nuances of two popular language models. Consequently, it may also unveil potential biases embedded within these models.

Literature Review:

Large language models represent the forefront of current studies and research, prompting numerous investigations into their functionalities and applications. Cheong et al. [11] conducted an evaluation of patient education materials on obstructive sleep apnoea generated by ChatGPT and Google Bard. The research involved the extraction of fifty frequently asked questions in English from the patient information webpages of four major sleep organizations, categorizing them as input prompts. Responses from ChatGPT and Google Bard underwent independent rating by two otolaryngologists with a Fellowship of the Royal College of Surgeons and a specialized interest in sleep medicine and surgery, utilizing the Patient Education Materials Assessment Tool–Printable Auto-Scoring Form. As a secondary outcome, responses were subjectively screened for any incorrect or potentially harmful information. The Flesch-Kincaid Calculator was employed to assess the readability of responses from both ChatGPT and Google Bard. The study's findings indicated that ChatGPT offers superior patient education materials for obstructive sleep apnoea compared to Google Bard.

An electrocardiogram (ECG) is a crucial medical test for diagnosing heart conditions, though interpreting its results can pose a challenge. In a study conducted by Fijačko et al. [12], the accuracy of Bing Chat Enterprise, ChatGPT-4 Pro, and Google Bard in interpreting ECG images from the American Heart Association (AHA) Advanced Cardiovascular Life Support (ACLS) multiple-choice question exams were examined. Each chatbot provided three separate interpretations of the same ECG image with identical prompts, and these interpretations were then compared to the AHA ACLS exam answers. In addition to assessing overall accuracy, the chatbots were prompted to estimate their "Level Of Correctness" for each interpretation of the ECG image. Throughout the study, a total of 81 interpretations were conducted using 9 different ECG images. The results revealed that ChatGPT-4 Plus emerged as the most successful chatbot, accurately interpreting nearly two-thirds of the ECG images. Following closely was Google Bard, correctly interpreting ECG images almost half of the time, while Bing Chat Enterprise demonstrated accurate interpretations less than a quarter of the time.

In a notable study by Patil et al. [13], the accuracy, response length, and response time of the latest versions of ChatGPT (ChatGPT-4) and Bard were compared in their ability to answer radiology board examination practice questions. Text-based questions from the 2017-2021 American College of Radiology's Diagnostic Radiology In-Training (DXIT) examinations were used to evaluate the two models. Analyzing 318 questions, the study found that ChatGPT responded significantly more accurately than Bard, but with shorter responses and longer response times. ChatGPT also demonstrated superior performance in neuroradiology, general & physics, nuclear medicine, pediatric radiology, and ultrasound. Overall, the study concluded that ChatGPT displayed superior radiology knowledge compared to Bard. However, both chatbots revealed limitations and fallibility, providing incorrect or illogical explanations and sometimes failing to address the educational content of the questions. This highlights the need to use such models with caution and awareness of their limitations.

Al-Ashwal et al. [14], conducted a study to examine the sensitivity, specificity, and accuracy of ChatGPT-3.5, ChatGPT-4, Bing AI, and Bard were evaluated regarding their predictive capabilities for drug-drug interactions (DDIs). The assessment involved comparing their abilities to detect clinically relevant DDIs for 255 drug pairs. The results revealed that Bing AI demonstrated the highest accuracy and specificity, surpassing the performance of Google's Bard, ChatGPT-3.5, and ChatGPT-4. These findings underscore the substantial potential of these AI tools in reshaping patient care. Although the current AI platforms analyzed are not exempt from limitations, their capacity to swiftly analyze potentially significant interactions with good sensitivity suggests a promising advancement toward enhanced patient safety.

Gan et al. [15], conducted a comparative study to assess the performance of the Google Bard and medical students in mass casualty incident (MCI) triage, utilizing the Simple Triage and Rapid Treatment (START) method. The study employed a validated questionnaire featuring 15 diverse MCI scenarios to evaluate triage accuracy through content analysis. The results revealed that Google Bard exhibited significantly higher accuracy at 60%, whereas ChatGPT achieved a lower accuracy of 26.67%. In comparison, medical students demonstrated an accuracy rate of 64.3% in a previous study. However, no significant difference was observed between the performance of Google Bard and medical students. The overall results indicated that Google Bard outperformed ChatGPT.

Ali et al. [16], evaluated the performance of GPT-3.5, GPT-4, and Google Bard using a question bank designed specifically for neurosurgery oral board examination preparation. The study utilized the 149-question Self-Assessment Neurosurgery Examination, which employed a single best-answer, multiple-choice format to assess LLM accuracy. Fisher exact and univariable logistic regression tests were employed to examine differences in performance based on question characteristics. The results revealed that GPT-4 outperformed ChatGPT and Google Bard, achieving a score of 82.6%.

Dhanvijay et al. [17] conducted a comprehensive cross-sectional study to assess the effectiveness of three large language models (LLMs): ChatGPT 3.5, Google Bard, and Microsoft Bing, in providing responses to case vignettes in Physiology. A total of seventy-seven case vignettes in physiology were meticulously curated by two physiologists and validated by two additional content experts. Subsequently, each LLM was presented with these cases, and their respective responses were systematically gathered. Two physiologists independently scrutinized the accuracy of the answers furnished by the LLMs. Ratings were assigned on a scale ranging from 0 to 4, aligning with the structure of the observed learning outcome (pre-structural = 0, uni-structural = 1, multi-structural = 2, relational = 3, extended-abstract). The scores across the various LLMs were subjected to comparison through Friedman’s test, and interobserver agreement was established using the intraclass correlation coefficient. The findings of the study revealed that ChatGPT 3.5 achieved the highest score, Bing exhibited the lowest score, and Bard occupied an intermediary position between the two in terms of performance. Consequently, ChatGPT demonstrated superior performance compared to both Bard and Bing in generating responses to case vignettes in the field of physiology.

O’Leary [18], conducted a study using questions from the Watson Jeopardy! The challenge is to compare the performance of BARD, ChatGPT, and Watson. Through the utilization of these Jeopardy! questions, the analysis reveals that, for high-confidence Watson questions, all three systems demonstrate similar accuracy to Watson. Additionally, both BARD and ChatGPT exhibit accuracy comparable to that of a human expert, and their sets of correct answers display a high degree of similarity, as indicated by a Tanimoto similarity score. The study also observed that both BARD and ChatGPT have the capability to modify their solutions when presented with the same input information on subsequent occasions. In instances where the same Jeopardy! category and question are repeated, both systems may generate distinct and conflicting answers.

Plevris et al. [19] conducted an assessment of the capabilities of ChatGPT-3.5, ChatGPT-4, and Google Bard in tackling mathematical and logical problems. The study employed a set of 30 questions, divided into two categories of 15 questions each. The first set comprised problems not readily available online, while the second set included problems commonly found online, often accompanied by solutions. Each question was presented to each chatbot three times, and their responses were recorded and analyzed. The findings revealed that chatbots demonstrate accuracy in providing solutions for straightforward arithmetic, algebraic expressions, and basic logic puzzles. However, this accuracy is not consistent across all attempts. For more intricate mathematical problems or advanced logic tasks, the chatbots' answers, while appearing convincing, may lack reliability. Moreover, there is a notable inconsistency issue, as chatbots often yield conflicting answers when presented with the same question multiple times. The results indicated that ChatGPT-4 outperforms ChatGPT-3.5 in both sets of questions. Google Bard ranks third in the original questions of the first set, trailing behind the other two chatbots. However, Bard exhibits the best performance, securing first place in the published questions of the second set. This discrepancy is likely attributed to Bard's direct internet access, unlike the ChatGPT chatbots, which, due to their designs, lack external communication capabilities.

Koga et al. [20], conducted an analysis and comparison of the predictive performance between ChatGPT and Google Bard in determining neuropathologic diagnoses based on clinical summaries. The study involved the examination of 25 cases of neurodegenerative disorders presented at the Mayo Clinic Brain Bank Clinico-Pathological Conferences. The models generated multiple pathologic diagnoses along with their respective rationales, which were then juxtaposed against the conclusive clinical diagnoses provided by physicians. In specific figures, ChatGPT-3.5, ChatGPT-4, and Google Bard accurately identified primary diagnoses in 32%, 52%, and 40% of cases, respectively. Furthermore, correct diagnoses were encompassed within the generated results for 76%, 84%, and 76% of cases, respectively. These results shed light on the varying predictive capabilities of the models and their alignment with the final clinical assessments.

Al-Ashwa et al. [14] evaluated the predictive performance of ChatGPT-3.5, ChatGPT-4, Bing AI, and Bard in forecasting drug-drug interactions, considering sensitivity, specificity, and accuracy. The study involved a comparison of these large language models in the detection of clinically relevant DDIs across 255 drug pairs. Notably, Bing AI emerged with the highest accuracy and specificity, surpassing the performance of Google's Bard, ChatGPT-3.5, and ChatGPT-4. These findings underscore the considerable potential of these AI tools in revolutionizing patient care.

Seth et al. [21] conducted a comprehensive assessment to evaluate the efficacy of Google BARD, Bing AI, and ChatGPT-3.5 in delivering accurate and secure medical information related to rhinoplasty. The study involved presenting six specific questions about rhinoplasty to ChatGPT, BARD, and Bing AI. An expert panel comprising Specialist Plastic and Reconstructive Surgeons, possessing extensive experience in rhinoplasty, employed a Likert scale to assess the responses. Reliability was measured using the Flesch Reading Ease Score, Flesch–Kincaid Grade Level, and Coleman–Liau Index. The modified DISCERN score served as the criterion for evaluating suitability and reliability. The study identified that, in terms of reliability, both BARD and ChatGPT exhibited significantly higher reliability compared to Bing AI. Regarding suitability, BARD achieved a significantly higher DISCERN score than both ChatGPT and Bing AI. In terms of the Likert score, ChatGPT and BARD demonstrated similar scores, both surpassing those of Bing AI. In conclusion, the study determined that BARD provided the most concise and comprehensible information, followed by ChatGPT and Bing AI.