"I think this is the most disruptive technology": Exploring Sentiments of ChatGPT Early Adopters using Twitter Data

I. INTRODUCTION

• ChatGPT can remember conversational context, reject inappropriate requests, challenge incorrect responses, and address complex questions, and it attracted more than 1 million users in one week.The introduction also notes predictions that it could replace Google and make some content-creation professions obsolete, alongside demonstrations of coding and exam-query performance.Source: §I · p. 1
• Studying early adopters’ sentiments matters because their enthusiasm and influence can shape broader perceptions and provide insights into a product’s potential success or failure.Their early encounters with problems can also help identify and fix issues before they become widespread, increasing the tool’s chances of market success.Source: §I · p. 1
• The researchers chose Twitter because unregistered users can read tweets and because the platform has been used to study public perception, health, and other topics.They describe Twitter as a large microblogging site and a source of information about potential customers.Source: §I · p. 1
• The study analyzes 10,732 early-adopter tweets, using topic modelling to identify discussion topics and manual qualitative sentiment analysis on selected tweets from each topic.The paper also reports topic-level tweet counts and classifies sentiments as positive, negative, or neutral.Source: §I · pp. 1–2
• The paper contributes topic-level descriptions with tweet examples and counts, sentiment categories for each topic, and future research directions concerning ethical and societal implications.It frames these contributions as relevant to both researchers and users.Source: §I · p. 2

II. RELATED WORK

• Prior Twitter studies analyzed IoT topics and sentiment, hybrid-work sentiment with RoBERTa, movie success, and Pakistan election prediction, reporting 98% accuracy.The IoT study found greater interest in business and technology and favourable sentiment; the hybrid-work study found majority-positive sentiment.Source: §II · p. 2
• The authors identify their study as the first qualitative analysis of early adopters’ ChatGPT sentiments and feedback, offering a snapshot of early public responses.They note that substantial research directly related to ChatGPT was not yet available, while related work covered GPT-family text-generating AI and Twitter data mining.Source: §II · p. 2

III. RESEARCH METHODOLOGY

• The methodology covers research questions, dataset construction, preprocessing, topic identification, and sentiment analysis, with its overall process shown in Fig. 1.Source: §III · p. 2

A. Research Questions

• RQ1 asks what characterizes ChatGPT early adopters.Source: A. Research Questions · p. 2
• RQ2 asks which main ChatGPT topics are discussed on Twitter.Source: A. Research Questions · p. 2
• RQ3 asks what sentiments Twitter users express about ChatGPT topics.Source: A. Research Questions · p. 2

B. ChatGPT Tweet Dataset Construction

• Researchers collected English-language tweets containing “ChatGPT” from December 5–7, 2022, using Python and the Twitter API, and recorded text, location, occupation, verification status, date, and hashtags.The dataset contained 18K tweets; location, occupation, and verification status were analyzed for RQ1.Source: B. ChatGPT Tweet Dataset Construction · p. 3

C. Data Pre-processing

• After removing retweets and duplicate tweets, the analysis dataset comprised 10,732 tweets.Source: C. Data Pre-processing · p. 3

2) Lowercasing:

• They lowercased tweets so differently capitalized forms of words were represented in lowercase.For example, “StackOverflow” and “Stackoverflow” were mapped to the same lower-case form.Source: 2) Lowercasing · p. 3

3) Noise Removal:

• They removed punctuation, URLs, emojis, and Twitter handles, retaining only alphanumerical data.Source: 3) Noise Removal · p. 3

4) Stop words Removal:

• They removed NLTK English stop words and the three most frequent domain-specific words: ChatGPT, OpenAI, and AI.Source: 4) Stop words Removal · p. 3

5) Lemmatization:

• They used WordNet-based lemmatization with NLTK to reduce inflected word forms to dictionary roots.The paper describes this as a common practice in existing literature.Source: 5) Lemmatization · p. 3

D. Identification of ChatGPT Topics

• They used MALLET’s LDA implementation to group tweets by word co-occurrence and frequency, assigning each tweet probabilities of belonging to topics.LDA was used to answer RQ2.Source: D. Identification of ChatGPT Topics · p. 3
• They tested topic counts from 5 to 60 in steps of 2, with 100, 500, and 1,000 iterations, and found comparatively high coherence for 9 to 12 topics.Coherence was used to assess topic understandability and human comprehensibility.Source: D. Identification of ChatGPT Topics · p. 3
• The researchers selected nine topics after coherence testing and reviewing 15 randomly selected tweets from each candidate topic count between 9 and 12.They judged nine to balance dataset comprehensibility.Source: D. Identification of ChatGPT Topics · p. 3
• They ran LDA with nine topics and generated a CSV for each topic, sorting each file by documents’ relevance to that topic.Source: D. Identification of ChatGPT Topics · p. 4

E. Sentiment Analysis on ChatGPT Topics

• They chose qualitative manual analysis after finding NLTK misclassified some tweets and wanting to assess images accompanying tweets.They cite prior work [27] for qualitative analysis’s depth and nuance, and propose automated analysis for a larger future dataset to assess trends over time.Source: E. Sentiment Analysis on ChatGPT Topics · p. 4
• For sentiment analysis, the authors manually labelled nine topic datasets, each containing 100 randomly selected tweets, as negative, neutral, or positive.Neutral covered tweets discussing both positive and negative aspects or too ambiguous to assign a specific positive or negative sentiment.Source: E. Sentiment Analysis on ChatGPT Topics · p. 4
• The authors used open coding to identify topic-specific discussion patterns, recording key points and concise codes in spreadsheets.All five authors participated; two labelled each topic dataset, and disagreements were resolved through discussion.Source: E. Sentiment Analysis on ChatGPT Topics · p. 4

IV. RESULT ANALYSIS

• The paper presents its RQ1, RQ2, and RQ3 results in Sections IV-A, IV-B, and IV-C, respectively.Source: §IV · p. 4

A. RQ1. What are the characteristics of ChatGPT early adopters?

• Early adopters were geographically dispersed, with tweets mainly from North America and Asia and the USA, India, UK, Canada, and Germany as the top five countries.Tweets from Europe, Australia, and South America were similar in number to one another.Source: A. RQ1. What are the characteristics… · pp. 4–5
• Only 2% of early adopters were verified Twitter users, indicating adoption was not confined to verified accounts.Source: A. RQ1. What are the characteristics… · p. 5, Figure 4 · p. 5
• The top occupations were software practitioners, academics, and students, alongside a broad range of other professions.Other identified occupations included data scientists, investors, business analysts, entertainers, and journalists; users without a specified job were classified as “other.”Source: A. RQ1. What are the characteristics… · p. 5, Figure 5 · p. 5

being discussed about ChatGPT in Twitter?

• ChatGPT attracted researchers, practitioners, managers, entertainers, business analysts, and educationists.The paper frames identifying common topics as a way to understand how diverse communities express their experiences across domains.Source: being discussed about ChatGPT in… · p. 5
• ChatGPT reached 1 million registered users within five days of its beta release.The passage contrasts this with approximately 300 days for Facebook, 1,200 for Netflix, and 75 for Instagram; the authors say the topic analysis can help assess ChatGPT’s capability, effectiveness, and usability.Source: being discussed about ChatGPT in… · p. 5
• Authors selected topic names by reviewing the top 20 keywords and 30 randomly selected tweets per topic, then discussing until consensus.They first used LDA to identify the optimal number of topics; the naming approach is described as common in the literature,,,,,.Source: being discussed about ChatGPT in… · p. 5
• Nine topics were identified among early ChatGPT adopters, and Fig. 6 shows their distribution by tweet count.The supplied passages do not report the individual topics or their tweet counts.Source: being discussed about ChatGPT in… · p. 5, Figure 6 · p. 5

1) Disruptions to software development:

• Users described ChatGPT generating and debugging code, as well as summarizing and translating it.Source: 1) Disruptions to software development · pp. 5–6

2) Entertainment and exercising creativity:

• Users used ChatGPT for entertainment by prompting poems, jokes, and humorous writing, often combining traits of different characters, personalities, or concepts.Source: 2) Entertainment and exercising… · p. 6

3) Natural Language Processing:

• Early adopters discussed ChatGPT’s natural-language capabilities for generating and understanding human-like text.Source: 3) Natural Language Processing · p. 6

4) Impact on Educational Aspects:

• Users discussed ChatGPT for education, including early childhood learning, syllabus development, literature review, and crisis-management learning.Source: 4) Impact on Educational Aspects · p. 6

5) Chatbot Intelligence:

• Early adopters highlighted ChatGPT’s ability to address complex problems, including writing and debugging complex code, optimization, and complicated queries.Source: 5) Chatbot Intelligence · p. 6

6) Impacts on Business Development:

• Users discussed ChatGPT for startup pitches, business use cases and plans, and financial advice.Source: 6) Impacts on Business Development · p. 6

7) Implications for Search Engines:

• Early adopters described ChatGPT as a search alternative that selects relevant information and explains it simply, offering a novel experience compared with Google Search and Bing.Source: 7) Implications for Search Engines · p. 6

8) Q&A Testing:

• Q&A Testing was the dataset’s third-largest topic, covering users’ efforts to learn, compare, and verify answers across academic, conceptual, technical, and emerging subjects.Users also asked open-ended and analytical questions to test ChatGPT’s capabilities.Source: 8) Q&A Testing · p. 6

9) Future Careers & Opportunities:

• Future Careers & Opportunities was the largest topic, with adopters discussing how AI could affect careers and what people and organizations need to stay current and support sustainable progress.They considered the skills, knowledge, values, and behavior needed to adapt, as well as collaboration between industry and research organizations and progress aligned with human social, emotional, and cognitive needs.Source: 9) Future Careers & Opportunities · p. 6

are being expressed about ChatGPT topics on Twitter?

• The qualitative sentiment analysis reports positive, negative, and neutral percentages for each topic in Fig. 7.Source: are being expressed about ChatGPT… · p. 7, Figure 7 · p. 7

1) Disruptions to software development:

• 81% of sentiment about software development was positive, with adopters praising ChatGPT’s help with coding, API generation, debugging, and troubleshooting.Users also described its potential to save developers time and effort in day-to-day tasks.Source: 1) Disruptions to software development · p. 7
• Only six users expressed concern about software-development disruptions, while others cautioned that ChatGPT can give plausible but wrong answers, especially for code using changing external libraries.Some users said developers should be mindful when using ChatGPT for development assistance.Source: 1) Disruptions to software development · p. 7

2) Entertainment and exercising creativity:

• 92% of sentiment in the entertainment and creativity topic was positive, with users generating poems, jokes, humorous write-ups, and combinations of familiar entities.These outputs typically began with a prompt or seed text.Source: 2) Entertainment and exercising… · p. 7
• Users exercised creativity with prompts for amusing crossovers and original stories, including a Shakespearean Friends scene and a Sherlock Holmes-style mystery about a dog.The story prompt featured Baxter, a dog in heaven investigating a stolen Christmas ham.Source: 2) Entertainment and exercising… · pp. 7–8
• Tweets sharing screenshots of ChatGPT’s entertaining outputs received high engagement, which the authors suggest might have contributed to the technology’s popularity.The authors point to the large initial number of subscribers as evidence of that popularity.Source: 2) Entertainment and exercising… · p. 8

3) Natural Language Processing.:

• In our Twitter dataset, 83% of tweets implied satisfaction with ChatGPT, 14% expressed concerns, and 3% reported a neutral viewpoint, reflecting diverse sentiments among early adopters.Early adopters seemed happy with ChatGPT’s amalgam of NLP and AI.Source: 3) Natural Language Processing · p. 8
• Early adopters praised ChatGPT’s natural-language processing and realistic, human-like text generation, while its trainers and supervised fine-tuning were described as improving generation and perception.Some users nevertheless said generated text lacked the depth and complexity of human thought.Source: 3) Natural Language Processing · p. 8
• Some users worried that ChatGPT could produce misinformation because it lacked critical-thinking skills, nuance, and ethical decision-making ability.The section also reports concerns about the quality of generated text.Source: 3) Natural Language Processing · p. 8

4) Chatbot Intelligence.:

• 78% of tweets expressed favorable views of ChatGPT’s intelligence, 20% raised harmful AI impacts, and 2% were neutral.The harmful-impact concerns included possible increases in terrorism, hacking, and unemployment.Source: 4) Chatbot Intelligence · p. 8
• Users cited clear React-app instructions as beneficial, but feared terrorism, hacking, unemployment, and harmful instructions, prompting a recommendation for effective usage policy and regulation.The paper recommends policy and regulation to mitigate negative impacts and improve future user acceptability.Source: 4) Chatbot Intelligence · p. 8

5) Impact on Educational Aspects:

• Educational views were 52% positive, 32% negative, and 16% neutral, reflecting both acceptance and concerns about ChatGPT’s use in education.The paper contrasts this mix with the overwhelming acceptance of ChatGPT for software development.Source: 5) Impact on Educational Aspects · p. 8
• Users saw educational applications in grading, assessing learning, and preparing syllabi, and some viewed ChatGPT as a personal teacher and a force for changing teaching and assessment.Some users hoped it could support assessment practices that better capture student learning.Source: 5) Impact on Educational Aspects · p. 8
• Users worried student essay and assignment use could impede learning, evade plagiarism detection, and yield shallow answers, while others disputed plagiarism efficacy and replacement of humanistic education.Some argued ChatGPT can identify AI-produced text, mitigating plagiarism concerns; others said it would be neither an effective way to plagiarize nor a replacement for humanistic education.Source: 5) Impact on Educational Aspects · p. 8

6) Impact on Business Development:

• 75% of business views were positive, 5% negative, and 20% neutral toward ChatGPT’s use in business.The analysis describes its business-domain prospects as positive across applications from business-plan development to customer-service support.Source: 6) Impact on Business Development · p. 9
• Business uses ranged from elevator pitches, technical co-founder support, and decision-making to business plans, trading strategies, customer support, and financial advice.Users also anticipated AI-assisted coding would make tech businesses smaller and flatter, boost good coders’ productivity, and help managers make better decisions.Source: 6) Impact on Business Development · p. 9

7) Implications for Search Engines:

• 54% of search-engine tweets were positive, 15% negative, and 31% neutral, with users often describing ChatGPT as a potential replacement for existing search engines.Some users called it a “#googlekiller.”Source: 7) Implications for Search Engines · p. 9
• Users saw ChatGPT as faster in search and more accurate than search engines, and as a potential risk to knowledge-sharing platforms such as Stack Overflow, Quora, and Wikipedia.The passage specifically associates speed with search and accuracy with knowledge-sharing platforms.Source: 7) Implications for Search Engines · p. 9
• Some users questioned ChatGPT’s information accuracy and awareness of its limitations, prompting the recommendation to verify its results against other sources before use.The passages link these concerns to its use of internet-sourced information.Source: 7) Implications for Search Engines · p. 9
• Browser extensions combining ChatGPT and existing search results were described as a way to support better-informed decisions, and the authors foresee improved search and knowledge-sharing interactions.The passages frame this as a potential future enhancement to user experience.Source: 7) Implications for Search Engines · p. 9

8) Q&A Testing:

• 38% of Q&A-testing tweets were positive, 40% neutral, and 22% negative; positive reactions reflected answer quality, human-friendly interaction, fast responses, and explanations.Users described value from quality answers and enthusiasm for how ChatGPT interacts while answering questions.Source: 8) Q&A Testing · p. 9
• Q&A adopters raised concerns about wrong or invalid answers, including a chemistry response that gave moles instead of grams, and confidently incorrect answers may hamper wider adoption.The passages also note Stack Overflow’s temporary ban on ChatGPT-generated answers, citing incorrect answers.Source: 8) Q&A Testing · pp. 9–10

• Sentiment was 75% positive, 9% neutral, and 16% negative, with positive views linked to ChatGPT’s fast, effective solutions for career success.Source: 9) Future Careers & Opportunities · p. 10
• Adopters saw AI coding and decision support enabling small teams to create valuable ventures, while ChatGPT’s innovation, learning, and adaptation could simplify job tasks.Source: 9) Future Careers & Opportunities · p. 10
• Adopters viewed personalized messages, services, and answers as ways to engage future customers, including through faster, accurate customer support.Source: 9) Future Careers & Opportunities · p. 10
• Negative views reflected fears of job loss as ChatGPT could perform software-development and coding work quickly, threatening programmer and other IT jobs.One adopter also framed ChatGPT as ready to replace product managers.Source: 9) Future Careers & Opportunities · p. 10

V. IMPLICATIONS

• The implications section uses the study’s topic and sentiment findings to inform future ChatGPT acceptance and discuss guidance for users and researchers.Source: §V · p. 10

A. Implications for Users

• The study identifies ChatGPT as a potential aid for software development and business tasks, including generating ideas that can support brainstorming.Users may use it to help create software and feasibility reports or business cases.Source: A. Implications for Users · p. 10
• Users should verify ChatGPT’s outputs before critical decisions, protect confidential data, and consider legal and ethical concerns when using its creations in academic or commercial work.The passage notes that outputs are not verified by an established authority and that adversarial attacks may expose users to data breaches.Source: A. Implications for Users · p. 10

B. Implications for Researchers

• Researchers should investigate ways to mitigate misuse in educational work and misinformation or inaccurate outputs, including protocols for ethical use across domains.The concerns were reflected in the Impacts on Educational Aspects and Q&A Testing topics.Source: B. Implications for Researchers · p. 10

VI. THREATS TO VALIDITY

• Because the study uses Twitter tweets to represent early adopters, future research could examine Stack Overflow and blogs to further generalize the findings.Twitter was described as an enriched knowledge base for sentiment analysis.Source: §VI · p. 11
• Manual topic-wise sentiment labels may reflect human judgment bias; two researchers labelled each topic’s sample, resolving disagreements through discussion.Source: §VI · p. 11
• Choosing the optimal number of topics N was a validity threat, addressed by experimenting with a broad range of N values.Source: §VI · p. 11

VII. CONCLUSION

• Early adopters expressed overwhelming excitement and limited concern, while users praised ChatGPT’s performance and potential to assist across domains, alongside ethical concerns about education.The conclusion calls for continued research and dialogue to develop responsible-use best practices.Source: §VII · p. 11