3.1 Data collection and cleaning

Social media platforms such as Twitter, Facebook, Instagram, and Reddit allow users to express their emotions, interests, hobbies, and opinions in real-time within an online community. Reddit (https://Reddit.com) has approximately 57 million daily active users until 2023, one of the largest social media outlets in terms of users [11]. Reddit users can share text, links, images, or videos in various sub-communities (called subreddits and dedicated to specific topics) [48]. Everyone has access to the public subforum (called subreddits on Reddit), and users can comment and vote on posts and comments for free and anonymously. Reddit has more than 100,000 active subreddits as of 2023. A large number of existing studies use data from the platform to examine public perceptions, attitudes, and opinions. These studies cover a broad range of topics, including disasters [49], vaccines [50], advertisement [51], tobacco [52], vehicles [53], climate change [54], digital governance [55], political psychology [56] and collective identity [57]. The Application Programming Interface (API) is a set of tools that defines how a software application interacts with other components, services, or platforms [58]. APIs allow for communication and data exchange between different software systems, enabling them to connect and interoperate with each other.

In terms of ethical considerations in the present research, the official API of Reddit is freely and publicly available to third parties [12, 59]. In accordance with Reddit’s privacy policy, developers are permitted to write programs or applications, such as Apify used in this study, that interact with the Reddit platform through specific requests and commands. These actions include retrieving specific information, posting content, or performing other tasks [60]. Reddit enables researchers to extract the subreddits, threads, comments, and associated metadata through various programming languages [61]. Therefore, data collection and analysis in this study comply with the terms and conditions of the data source. Following the principle of data minimization, it only collects data relevant to the purpose of the study. In addition, to adhere to the principles of privacy and untraceability, personal user information is anonymized because Reddit posts and comments are user-generated.

This study utilized Apify (https://apify.com) for collecting Reddit posts and comments as the primary data source. Apify is a platform for data collection for web scraping, data extraction, and automation. It offers tools and services that assist developers in extracting data from web pages, executing automated tasks, and building web crawlers. Personal information has been anonymized. The data consists of posts and comments. The sample data is shown in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. An example of a sample post regarding GPT-3.

https://doi.org/10.1371/journal.pone.0302502.t001

It sets GPT3, GPT3.5, and GPT4 as keywords. These keywords aim to examine whether people’s perceptions of ChatGPT have evolved, particularly since the ChatGPT update. Apify collected 11,730, 10,109, and 12,046 relevant entries (posts and comments) for the respective keywords. This study performed random sampling to ensure sample uniformity, resulting in 10,000 entries preserved for each GPT version. The dataset consisted of 30,000 samples from June 2020 to August 15, 2023.

In terms of data cleaning, emoticons, digits, punctuation, links, unnecessary words, non-ASCII characters, and stopwords were removed from the textual data suggested by Yadav et al [61, 62]. The NLTK library provides the English stopwords list. It consists of common English words with no semantic or informational value [63, 64]. Stopwords are typically filtered out in natural language processing to enhance the efficiency and accuracy of text analysis. Additionally, all uppercase letters have been converted to lowercase. Furthermore, considering a slight overlap in the data sources, this study also conducted duplicate text filtering. In the end, 23,773 entries were retained, forming the database Processed_GPT_total.json, displayed in supporting information.

To improve the reproducibility of our research results, the data and research design have been stored on Protocols.io. Protocols.io has assigned a protocol identifier (DOI) to our protocols, which is DOI: dx.doi.org/10.17504/protocols.io.bp2l6xee1lqe/v1. Open Access license is freely available for anyone. Readers can find relevant details by visiting the following URL: https://www.protocols.io/private/899FA95EAEFD11EEAB870A58A9FEAC02.

3.2 Topic modeling

All data analyses were processed using Python (version 3.10). For topic modeling, prior research proposes three techniques: Latent Dirichlet Allocation (LDA), which adeptly discerns latent topics through probabilistic approaches; Non-Negative Matrix Factorization (NMF), which underscores the intricate relationship between documents and their inherent topics; and Transformer-based models, renowned for their proficiency in grasping the intricate semantics embedded within textual data [65].

In this study, the Latent Dirichlet Allocation (LDA) model will be used to conduct the topic model. It is a hierarchical Bayesian probability model consisting of three levels: documents, topics, and words. The fundamental concept behind the LDA model is that each document can be represented as a mixture distribution of latent topics, where each topic is a probability distribution over all words in the vocabulary [66]. The LDA model introduces latent random variables, specifically the topic mixture weights θ, which are T-dimensional parameters representing the topics. These topic mixture weights are endowed with a Dirichlet prior distribution [67]. (1) The parameters α₁,…, α_T in the Dirichlet distribution are regarded as hyperparameters for the multinomial θ = P(θ₁,…, θ_T). Each hyperparameter αⱼ can be interpreted as a prior distribution on the frequency of sampling topic j from a document. To facilitate computation and due to the assumption that different topics are used similarly across documents, LDA assumes the exchangeability of topics within a document. For computational convenience, the model assumes symmetric Dirichlet distributions with identical hyperparameters α, i.e., α₁ = α₂ =… = α_T = α. It is worth noting that Dirichlet distributions and multinomial distributions are conjugate distributions. Assuming a T-dimensional random vector θ follows a Dirichlet distribution, the T components θ₁, θ₂,…, θ_T are continuous and non-negative, satisfying θ₁ + θ₂ +… + θ_T = 1. In the initial LDA literature, Dirichlet priors were mainly introduced on the document-topic distribution θ, without assuming priors on the topic-word distribution φ. However, for leveraging the properties of conjugate distributions and facilitating inference and computation, subsequent research introduced a Dirichlet before the topic-word distribution φ, assuming the exchangeability of words within a topic. The hyperparameter of this Dirichlet prior is denoted a s β. Consequently, the LDA model incorporates two sets of priors: one for the document-topic distribution from a symmetric Dirichlet distribution (α) and another for the topic-word distribution from a symmetric Dirichlet distribution (β) [68].

Perplexity is a common metric used to evaluate the performance of topic models and is particularly applicable to a selected number of LDA topics to model. It measures how well the model fits the documents in a given dataset, with a lower perplexity score indicating better model performance on that dataset.

The perplexity is calculated based on the likelihood probability of the model on the training set by computing the average log-likelihood probability on the word sequence of each document on the training set. Perplexity is calculated through the Equation in LDA: (2) Where N_d is the number of words, and M is the size of the document d. N is the size of the words set, and K is the size of the topic set. p(z_p|d_j) is the probability that the topic z_p appears in document d_j. P(w_i|z_p) is the probability that the word w_i occurs in z_p. It is clear that perplexity is mainly influenced by M, N and K according to equation [69]. By comparing the perplexity scores for different numbers of topics, the optimal number of LDA topics to model can be determined to obtain the best fit on a given dataset, while also being mindful of overfitting issues [70].

3.3 Sentiment analysis

Existing research typically uses three approaches for sentiment analysis: rule-based approaches, machine learning, and deep learning [71]. Rule-based methods, often referred to as rule-based sentiment analysis, utilize predefined sentiment lexicons and rules to evaluate sentiment [72]. Machine learning methods, on the other hand, learn sentiment classification models from data through supervised or unsupervised learning, offering good generalization capability [73]. Deep learning methods such as RNNs, CNNs, and Transformers automatically learn features from text, adept at capturing long-term dependencies, albeit requiring substantial data and computational resources [74].

The present study employs a rule-based approach to classify sentiments in posts and comments. The sentiment categorization utilizes a weighted approach that combines VADER with TextBlob. By combining the strengths of TextBlob and VADER, two distinct sentiment analysis tools that employ different algorithms and semantic processing approaches, and weighting their results, the overall performance of sentiment analysis can be effectively enhanced. These tools exhibit varied performances in different contexts, and through weighting their outcomes, one can leverage their respective strengths to improve the overall performance of sentiment analysis. Simultaneously, this integrated approach helps mitigate individual biases of different sentiment analysis tools, enabling the system to better adapt to diverse text samples, especially those of specific types. Furthermore, when confronted with complex and diverse language expressions, a single sentiment analysis method may exhibit instability. Integrating multiple methods enhances the model’s robustness, allowing it to adapt to various types and styles of text flexibly. Such a comprehensive approach demonstrates significant advantages in improving the accuracy and adaptability of sentiment analysis.

The VADER sentiment analysis hinges on a lexicon that links linguistic components to emotional strengths, referred to as sentiment scores [75]. Calculating the sentiment score for a specific text is a straightforward process of summing up the intensity values for each word in the text. Notably, each word’s intensity falls within a range of -4 to 4, and this lexicon is derived from human evaluations. The sentiment score for a sentence is determined by adding up the sentiment scores of its sentiment-carrying words [76]. Nevertheless, it employs the Hutto normalization process on the final sum to bring it within a range of -1 to 1. The compound score is formulated as follows: (3) Where x is the sum of the sentiment scores of the constituent words within the sentence, and alpha is the normalization parameter. These scores aid in determining the polarity of a given sentence. It offers versatility in handling intricate textual data analysis tasks. Upon analyzing a sentence, it provides two outputs: polarity and subjectivity. Polarity yields a value within the range of [–1, 1], where -1 indicates a negative sentiment, and +1 signifies a positive sentiment. Conversely, subjectivity produces an output between [0, 1] and reflects personal opinions and judgments.

Rule-based sentiment analysis in TextBlob relies on natural language processing techniques, utilizing predefined rules and syntactic structures to identify the emotional polarity within the text [77]. Initially, the text is decomposed into words and phrases, and part-of-speech tagging is conducted to comprehend the grammatical roles of each word in the sentence. Subsequently, based on a pre-defined sentiment lexicon, each word is assigned a sentiment polarity score, such as positive, negative, or neutral. Rules may also take into account relationships between words, where the presence of negation words, for instance, could alter the emotional polarity. By weighting or averaging sentiment scores for all words in the text, the overall emotional polarity of the text can be determined. The advantage of this method lies in its simplicity and ease of implementation, while its accuracy can be enhanced by continuously updating and expanding the sentiment lexicon [78].

Two sentiment analysis tools are used in this study: TextBlob and VADER. Each tool possesses analytical strengths, making them invaluable for nuanced sentiment extraction [79]. TextBlob, with its linguistic expertise, provides a polarity score, while VADER, adept at discerning both context and sentiment intensity, yields a compound score. Given the distinct virtues of each tool, relying solely on one would be an oversight. Therefore, a composite sentiment score is meticulously derived from both [78]. Let S_t represent the sentiment score from TextBlob and S_v the score from VADER. If this study designates and as their respective weights such that a + b = 1, the weighted sentiment score, S_w, is defined as (4) In this study, the weights are set at a = 0.4 for TextBlob and b = 0.6 for VADER, reflecting a slight inclination towards the latter’s comprehensive capabilities.

3.4 Sentiment trend analysis

Sentiment trend analysis, a burgeoning field, has become vital for comprehending public perception of specific issues [80]. In this study, sentiment trend analysis integrates the strengths of the aforementioned two methods, utilizing the results with a weighted approach. Initially, a DataFrame (df) is created to organize the data, encompassing columns for date, weighted sentiment scores, and sentiment labels (’Positive’ or ’Negative’). The date column undergoes conversion to the datetime type for accurate time series analysis. Subsequently, the ’Sentiment’ column values are determined based on the weighted scores, with ’Positive’ assigned if the score exceeds 0, and ’Negative’ otherwise. Finally, the DataFrame is grouped by date and sentiment, daily sentiment counts are computed, and a line plot is generated using matplotlib to illustrate the daily counts of positive and negative sentiments over time.

A lucid visualization is paramount to fully capturing the chronology of sentiment dynamics [81]. As such, daily sentiment metrics are illustrated, clearly depicting the populace’s emotional ebbs and flows. This graphical elucidation not only bestows a daily sentiment snapshot but also illuminates prevailing trends, proving indispensable for decision-makers, ranging from corporate strategists to policymakers, who anchor their choices on the pulse of public sentiment.

4.1 Word frequency

In this part, word clouds and frequency graphs provide initial insights into the diverse perspectives of the general public on the topics (Question 1) and attitudes (Question 2) toward GPT. The analyzed entries exceeded a total count of 23,773 entries. Fig 1 displays the top 20 most common words from the entries. As Fig 1 shows, the public’s positive attitude toward ChatGPT is evident through the words "like" and "good", indicating their appreciation and approval of the model. However, the discussions also unveil contemplation about the practical applications of ChatGPT, encompassing terms such as "use," "using," "way," "make," and "need," highlighting the discourse on how to harness ChatGPT’s capabilities fully. In addition, words like "would," "think," "know," "could," and "even," express doubts and uncertainties, reflecting concerns about its potential limitations and abilities. Technical aspects of the discourse include terms like "model," "bot," "prompt," "data," "code," and "models," revealing the audience’s attention to ChatGPT’s internal working model, data processing, and technological implementation.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Top 20 most common words.

https://doi.org/10.1371/journal.pone.0302502.g001

4.2 Topic modeling

This part presents the emergent topics and themes identified through topic modeling. It aims to address research question 1: What are the emerging topics related to ChatGPT? This study combines qualitative and quantitative content analysis to uncover and discover latent topics and themes of the public’s discussion, which are believed to hold significant potential for research in the field of social media [82]. As a commonly used quantitative method for topic classification, the LDA model aids in determining the most optimal number of topics for classification. In Fig 2, the perplexity-topic number curve is plotted.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Perplexity vs. number of topics.

https://doi.org/10.1371/journal.pone.0302502.g002

Typically, the optimal number of topics is determined based on lower perplexity levels [83]. When the number of topics is set at 8, the perplexity is at its lowest. However, the lowest perplexity may not always signify the best model performance. With a high number of topics, models often overfit, resulting in excessive and non-convergent topic counts. An excessive number of topics may lead to high redundancy, resulting in low distinctiveness and uniqueness between topics [71, 83]. Hence, many studies rely on human judges to determine the optimal number of topics. This method also adheres to certain principles: (1) high coherence between words and topics; (2) the quality of topic, ensuring non-repetition, non-conflict, and coverage of primary content [84]. This study tests the topic categorization and high-frequency words of each topic when the number of topics was set at 8 (Table 2). However, it demonstrates poor coherence and topic quality. Across the topics, there is a lack of coherent themes, with words appearing disjointed and unrelated within each category. The representative words fail to form distinct and meaningful topics, undermining the effectiveness of the model in capturing the underlying structure of the data.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Top words of eight topics.

https://doi.org/10.1371/journal.pone.0302502.t002

Therefore, this study also tested the number of topics corresponding to the point of significant decrease in perplexity, i.e., the number of topics (7) near the inflection point of the curve. When the number of topics is set at 7, the distribution of word frequencies in relevant topics is shown in Table 3. The top words in each topic exhibit good coherence and topic quality. Table 3 shows that the top 10 words in each topic are categorized into seven topics, which are then assigned to three themes. The analysis of seven topics demonstrates the wide range of discussions regarding ChatGPT on the Reddit community. These discussions cover technical inquiries, philosophical pondering, impacts on society, creative applications, and entertainment. The topics reflect the multifaceted nature of ChatGPT and highlight the diverse perspectives and interests of the public when using it.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Top words of seven topics and three themes.

https://doi.org/10.1371/journal.pone.0302502.t003

The first topic concerns people’s general impressions of ChatGPT. The keywords such as "like," "think," and "good" indicate that individuals are generally favorable towards ChatGPT. This topic focuses on how people perceive ChatGPT’s potential benefits, usability, functionality, and positive impact on them. The second topic appears to focus on technical inquiries for assistance. Terms such as "bot," "prompt," and "link" indicate that users are seeking information on how to use ChatGPT for various tasks. The terms "questions," "message," and "action" suggest a desire to optimize ChatGPT’s functionality for specific purposes. The third topic delves into philosophical discussions, examining consciousness, AGI (Artificial General Intelligence), and human reasoning of ChatGPT. The keywords like "consciousness," "agi," "humans," and "belief" imply that users are exploring ChatGPT’s human consciousness, intelligence, and spiritual characteristics. The fourth topic explores the technical details of ChatGPT, focusing on coding and textual manipulation. The keywords such as "code," "prompt," "text," and "language" suggest discussions on how to utilize ChatGPT for code generation or text creation efficiently. This topic covers its capabilities in software development, content creation, and language-oriented tasks. The fifth topic explores ChatGPT’s impact on diverse fields of life, including the arts, healthcare, and quantum phenomena. The words "music," "art," "health," and "quantum" indicate discussions on how ChatGPT brings the revolution and advances to the artistic and scientific domains.

The sixth topic focuses on the broad social and economic influences of ChatGPT. Keywords like "market," "jobs," and "impact" suggest discussions on ChatGPT’s impact on the job market and the global economy. The debates could probably revolve around potential job loss resulting from ChatGPT and the ethical concerns about AI. The seventh topic concerns the correlation between ChatGPT and politics and entertainment. The keywords such as "trump" and "president" relate to ChatGPT’s function in political discussions. The terms "spider," "gif," and "gypsy" indicate the potential utilization of ChatGPT within cultural and entertainment contexts.

Through the qualitative content analysis, seven topics were systematically coded and categorized into three themes. Theme 1 covers Topics 1 and 3, which focus on users’ positive views of ChatGPT and its potential advantages and positive influence on various aspects of life. Theme 2 encompasses Topics 2 and 4, focusing on the technical methods of ChatGPT, including queries, assistance, coding, and practical applications. Discussions cover topics such as the application of ChatGPT for specific tasks, seeking guidance, and sharing experiences regarding coding and language generation. Theme 3 comprises Topics 5, 6, and 7, focusing on the broader social impact of ChatGPT on art, music, health, politics, market, employment prospects, scientific progress, and the entertainment industry.

4.3 Robustness of topic modeling

To verify the robustness of the model, it randomly selects 10,000 samples as a subset from the original dataset for testing [85]. To examine the robustness, the same parameters are used for two models [86].

Fig 3, compares the perplexity scores of the two models when the number of topics ranges from 1 to 20, finding that the results of the two runs are highly similar. The solid line with circles represents the perplexity changes of the original dataset when LDA topic modeling is performed, while the dotted line with triangles represents the perplexity changes of a subset of the original dataset under the same modeling process. It can be seen that the variation range of the perplexity scores of the two curves is between -8.5 and -11.1, and the overall trend shows a slow increase at first and then a decrease. Within the topics range of 1 to 8, the difference between the two curves is minimal, showing a high degree of similarity. The largest gap is when the number of topics is 20, and it is only 0.23 at this time, indicating that our model is more robust.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Robustness test-perplexity similarity test.

https://doi.org/10.1371/journal.pone.0302502.g003

When the topic modeling of the subset is performed and the number of topics is 7, the high-frequency words corresponding to each topic are shown in Table 4. Three Themes were summarized from the 7 Topics generated by LDA topic modeling based on a subset of the original dataset, which is consistent with our previous LDA modeling results using the original dataset. This also confirms the posts and comments about GPT. The three widely discussed themes are User Perception, Technical Methods, and Impacts on Society.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Top words of seven topics and themes based on a subset.

https://doi.org/10.1371/journal.pone.0302502.t004

To illustrate the robustness, the theme generated by the two models has a high degree of similarity with the top words (Fig 4). Regardless of whether it is applied to the original dataset or its subset, it can produce highly similar theme results. It has summarized the seven topics derived from the original dataset into three themes, and it also depicts the high-frequency words associated with different themes generated by two topic modeling processes. Among these, the intersection of each pair of themes represents the high-frequency words that yield identical results from both topic modeling techniques.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Theme similarity with the high-frequency word map.

https://doi.org/10.1371/journal.pone.0302502.g004

To further corroborate the reliability of this result, it repeatedly experimented with and sampled the original dataset five times [87]. Since comparing the results when the number of topics is 1 is not meaningful, it utilizes the perplexity scores from the second to the eighth topics of the original dataset as a baseline. These baseline scores are then compared against the perplexity scores generated by the subsets sampled on the other five occasions. A similarity test is subsequently conducted to assess the comparability shown in Fig 5. The horizontal axis of the heat map above represents the number of topics using LDA for topic modeling, ranging from 2 to 8, and the vertical axis represents the number of subsets we randomly extracted from the original dataset [88]. Different squares represent the difference in perplexity between the topic modeling results of 5 randomly selected subsets of the original data set and the topic modeling of the original data set.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Perplexity difference heat map.

https://doi.org/10.1371/journal.pone.0302502.g005

In Fig 5, the difference between the five tests and the baseline (perplexity of the original dataset) is small. When the number of topics is less than or equal to 6, the difference in confusion does not exceed 0.10. The relative maximum difference is when the number of topics in the second extracted subset is 8, at which time the difference in perplexity is only about 0.14. After several times of randomly extracting the subset and re-performing the number of topics-perplexity experiment, the perplexity difference is very small for five experiments in the range of the number of topics is 8, which further confirms the high robustness of our modeling. Rre-performing topic modeling on randomly selected subsets and topic modeling performed on the original dataset show extremely high similarity in results, and the results of multiple repetitions of the modeling show very little difference in perplexity. Therefore, based on the comparatively low perplexity differences between the subsets and the original data, the robustness of the model can be supported.

4.4 Sentiment analysis

In this part, two sentiment analysis models, Vader and Textblob, are assigned weights of 0.6 and 0.4 respectively for sentiment classification to explore research question 2. The sentiment analysis categorizes the emotional tone of the entries into three distinct parameters: positive, negative, and neutral. The weighting of positive, neutral, and negative entries is shown in Fig 6 (N = 23,773). The analysis reveals a positive sentiment among Reddit users, with approximately 61.6% of entries conveying affirmative emotional nuances. In contrast, about 20.8% of entries express negative sentiments, while neutral ones account for the smallest segment at 17.6%.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. The weighting of the sentiment pie chart.

https://doi.org/10.1371/journal.pone.0302502.g006

Fig 7 demonstrates the comprehensive sentiment distribution (N = 23,773). There is a noticeable concentration of entries between 0 and 0.6, indicating the prevailing positive emotions. Moreover, it is pertinent to mention that the sentiment analysis identified a significant count exceeding 5100 entries conveying a neutral sentiment. Furthermore, most entries fall within the range of -0.25 to 0.60, suggesting a moderately nuanced sentiment orientation and a notable absence of distinct polarization in the overall sentiment attitudes towards ChatGPT.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Sentiment distribution bar chart.

https://doi.org/10.1371/journal.pone.0302502.g007

Table 5 selectively displays the high-frequency words in posts and comments expressing different sentiments.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. High-frequency words for entries with different sentiments.

https://doi.org/10.1371/journal.pone.0302502.t005

In neutral discussions, individuals mention the usage of ChatGPT, such as "api", and "to access it you have to use the API." The API is an interface that facilitates communication between distinct software systems or services, enabling programs or applications to access the functionalities or data of other systems. Users might engage in conversations regarding the functionalities and limitations of the API, deliberating on the prospect of integrating ChatGPT’s language generation capabilities into their applications or systems using the API [89].

The negative comments with words such as "wrong," "bad," and "problem," reflect their perception of errors, issues, or flaws of ChatGPT. For instance, "It is just a dumb stunt for a dumb application," "The model has been quantized badly, " and "I have doubts about their security claims. " These express suspicion about certain aspects of ChatGPT which generate problematic content at times or unsatisfactory functionality. It suggests skepticism about the accuracy and quality of the content generated by ChatGPT.

It is worth noting that the term "model" appears in both positive, neutral, and negative posts and comments, showcasing varying perspectives on the ChatGPT based on its performance, applications, and potential risks. Positive comments emphasize its impressive capabilities, including generating high-quality text, conducting fluent conversations, and efficiently retrieving information. This technology is acknowledged for its significant advances in natural language processing, benefiting various fields such as intelligent assistants and text creation. On the other hand, negative feedback may suggest that ChatGPT produces incorrect outputs, raises ethical concerns, and has the potential to spread misinformation. These diverse viewpoints reflect the complexity of ChatGPT and its social implications.

4.5 Sentiment trend analysis

This part examines daily sentiment trends by comparing the quantity of positive and negative sentiment posts from January to August 2023 (N = 23,773) to explore research question 3. Based on the GPT-3.5 model, ChatGPT was launched by Open AI on November 30, 2022, gaining a growing user base. On March 15, 2023, OpenAI unveiled the new multimodal model, GPT-4, available for purchase [2]. It aims to ascertain whether version updates have influenced sentiment towards ChatGPT. Fig 8 demonstrates how sentiments changed over time. The graph displays two sentiment classes, denoted by green and red, representing positive and negative sentiments, respectively. The sentiments fluctuate over time during the ChatGPT update.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Daily positive and negative sentiment counts.

https://doi.org/10.1371/journal.pone.0302502.g008

Fig 8 shows that fluctuations in 23,773 entries occur during mid-February, mid-to-late March, May, and mid-July. In most instances, the number of positive entries surpasses that of negative ones. The graph reveals a moderately increasing trend in mid-February, where the daily entry count exceeded 350. This surge can be attributed to the launch of the Plus plan on February 9, offering Plus users the option to select from various versions of ChatGPT. Moreover, the worldwide release of ChatGPT Plus for purchase was announced on February 13th. In mid-March, there was a small peak in the count of entries, approaching nearly 400 daily, perhaps attributable to the announcement of GPT-4 on March 14, 2023. Additionally, there were 100 more positive entries than negative ones indicating that GPT-4’s launch offered advanced reasoning, complex instructions, and enhanced creativity. In May, discussions regarding ChatGPT reached a peak. This surge in activity can be attributed to several factors. Primarily, it could be due to OpenAI’s implementation of new privacy features on May 3rd, which introduced the option to "Turn off chat history and decline to use for model training." It addresses some privacy concerns and encourages user engagement. Furthermore, on May 12th, OpenAI released an update allowing ChatGPT Plus members to incorporate the Bing search engine for browsing web content. It transcended the previous limitations of ChatGPT’s database, which had been confined to information available only until 2021. Lastly, it could also be largely attributed to the momentous launch of the ChatGPT iOS app on May 18th in selected countries, including the United States, the United Kingdom, France, and others. It enriches users’ mobile experiences. However, due to the significant limitations imposed by these successive updates on the nation and the mobile operating system, most users have been unable to benefit from the conveniences. This has potentially led to public dissatisfaction, resulting in a surge of more than 300 negative comments and posts in a single day.

In mid-July, there was a slight increase in ChatGPT discussions, with more than 300 entries reflecting positivity. The surge of positive entries can be attributed to the widespread introduction of the Code Interpreter feature to all ChatGPT Plus users. This innovation allowed non-programmers to express intentions in everyday language, translating into executable Python code solutions, enabling the accomplishment of intricate tasks within a real-time working environment. The innovation not only streamlined the processes of code composition and data manipulation but also expedited the application of artificial intelligence across diverse domains. While the version updates of ChatGPT may spark heated discussions among Reddit users, users generally hold positive sentiments towards ChatGPT and there was no significant shift from positive to negative attitudes during the period between January 2023 and August 2023.