The Social Impact of Generative AI: An Analysis on ChatGPT

1 Introduction

• Generative models offer practical possibilities across sectors, while concerns range from privacy risks to deepening social inequalities.The introduction also notes observed effects in service provision, education, and scientific analysis, alongside concerns involving democracy, inequality, security, and military technology.Source: §1 · pp. 1–2
• The paper calls for comprehensive analysis of generative AI’s positive and negative social consequences, emerging trends, and areas for improvement.Such analysis is presented as needed to address vulnerabilities and account for diverse social contexts and realities of deployment.Source: §1 · p. 2
• The paper examines generative AI’s societal ramifications and future trajectories, with particular emphasis on ChatGPT.It previews analysis of impacts across social sectors, model developments, study design, and literature-review findings.Source: §1 · p. 2

2 BACKGROUND

• The background describes ongoing debates about technology’s benefits and hazards, including regulation, risk mitigation, liability, transparency, accountability, and socialization.It notes that emerging applications pervade societal domains from education to diplomacy and reshape social processes.Source: §2 · p. 2
• ChatGPT’s potential social transformations are debated through resistant, skeptical, and optimistic perspectives, amid unequal and complex social environments.The background links these debates to potential effects on power dynamics from individual interactions to broader social structures.Source: §2 · pp. 2–3
• Global Google searches for “ChatGPT” stayed consistently low until December 2022, then rose significantly and peaked in April and May 2023.The passage associates the shift with the months following ChatGPT’s prototype launch on 30 November 2022 [Gordon, 2022]; Figure 1 notes a Google data-collection improvement on 1 January 2022.Source: §2 · p. 3, Figure 1 · p. 3
• The paper’s goal is to assess ChatGPT’s evolution and examine its perceived impact on social aspects during the current wave of AI evolution.It emphasizes examining scientific knowledge across localities and contexts.Source: §2 · p. 3

3 STATE OF THE ART

• A language model assigns probabilities to word sequences, while a large language model uses a neural network with typically billions or more parameters.The section traces generative pre-training, transformers, and early language models, distinguishing pre-trained transformers such as BERT and XLNet from generative models.Source: §3 · p. 3

3.1 Foundational models

• Foundational models are trained on broad data at scale so they can be adapted to a wide range of downstream tasks.GPT-4 is described as OpenAI’s most recent model in the passage, with training details and size withheld for business reasons.Source: §3.1 · p. 4

3.2 Task-Specific Models

• Task-specific models adapt foundational GPTs for particular tasks or subject domains, including through fine-tuning on new data.The section also describes instruction-tuned InstructGPT and text-to-image and text-to-video models that generate media from prompts.Source: §3.2 · p. 4

3.3 Domain-specific models

• Domain-specific GPT systems and software add-ons target applications such as finance, biomedical text, code development, and spreadsheet use.Examples include BloombergGPT, BioGPT, CoPilot, and GPT for Sheets and Docs.Source: §3.3 · p. 4

4 STUDY DESIGN

• The review follows a systematic protocol covering research questions, search strategy, eligibility criteria, data extraction, and analysis and synthesis.The authors cite the protocol of and review guidance of.Source: §4 · p. 4

4.1 Goal and Research Question Definition

• The study asks about ChatGPT’s perceived positive and negative impacts, emerging development trends, and areas for improvement.These three questions guide the literature review.Source: §4.1 · p. 5

4.2 Search strategy

• The search combined grey literature intended to capture media and technology-sector sentiment with academic literature across multiple fields.The grey-literature search produced 1230 sources and the academic search produced 86; both sets were screened in three stages by two authors, with disagreements referred to a third author.Source: §4.2 · pp. 5–6

4.3 Eligibility Criteria Definition and Data Extraction

• The authors selected 71 papers from an initial pool of 1316, requiring English-language content mainly related to ChatGPT and LLMs.For academic articles, they sought diverse approaches rather than relying on one field; they preferred extensive analysis in news articles because early coverage was often repetitive and brief.Source: §4.3 · p. 6
• Two authors extracted relevant data and organized the selected papers in Atlas.ti, resolving extraction conflicts through a third author’s blind review.Atlas.ti was used to organize documentary resources with codes and categories and visualize content trends.Source: §4.3 · p. 6

• The two search phases ran from 29 November 2022 to 22 February 2023 and from 22 February to 19 May 2023; Atlas.ti organized and visualized the selected papers.Source: §5 · pp. 6–7
• The codes were developed mainly in vivo and organized into four categories: positive impacts, negative impacts, emerging trends, and areas for improvement.Source: §5 · p. 7
• Emerging-trend codes include copyright uncertainty, private-sector liability, AI Act coverage [European Commission, 2021], unintentional misinformation, and skepticism about impact.Source: §5 · p. 7
• Improvement codes cover mitigating adverse outcomes, data-governance uncertainty, limited up-to-date information, and limited medical terminology.The authors say current legal frameworks do not adequately account for models like GPT.Source: §5 · p. 7
• Across 71 documents, Figure 3 displays frequencies of the most repeated codes, while Table 3 identifies the most and least recurrent codes by category.Source: §5 · pp. 7–8, Figure 3 · p. 8, Table 3 · p. 8
• Most publication categories skew negative; positive perceptions are more prevalent in blogs and conference/symposium papers, and less prevalent in news articles.Scientific papers, blogs, conference symposiums, and other publication types encompass all coding groups; articles include only emerging trends, areas for improvement, and negative perceptions.Source: §5 · p. 8, Figure 4 · p. 9

Perceived positive and negative impacts

• ChatGPT is described as useful across customer service, education, healthcare, business, and journalism, including for repetitive tasks, personalized study support, and content creation.The review also notes cross-cultural dialogue, continuous availability, medical research and telemedicine, and potential environmental applications.Source: §6 · pp. 9–10
• The most recurrent concern is bias, alongside risks involving false information, privacy, job loss, over-regulation, and cybersecurity.The review links training data to sexist and racist views and warns that false information can affect access to accurate information and democratic stability.Source: §6 · p. 10

Emerging trends

• Emerging trends include copyright uncertainty, calls for ethical-use guidelines and transparency, and accountability shared by users and developers.The review also identifies timely regulation, political bias, AI-market changes, and environmental awareness as notable trends.Source: §6 · p. 11

Areas for improvement

• Proposed improvements include regulation that addresses generative models, people-centred design, human rights, and frameworks for navigating trade-offs.The review raises whether the AI Act’s risk-based approach covers generative models and calls for guidance spanning application through AI research and development.Source: §6 · p. 11
• Technical concerns include fictional references, repetition, hallucinations, and weak context understanding; suggested responses include filtering harmful outputs and improving training data.The review also identifies ethical development guidelines and broader risk-mitigation mechanisms as areas for improvement.Source: §6 · pp. 11–12
• Data governance should protect sensitive information and clarify third-party information exchange, while current data is sought to support accuracy and relevance.The review notes a trade-off between incorporating new data to improve performance and addressing data-governance issues first.Source: §6 · p. 12
• Other improvement areas include end-user responsibility, timely regulation, and greater awareness of environmental impacts.Source: §6 · p. 12

Conclusion

• The review provides initial insights into perceptions of generative AI and identifies privacy and potential bias as notable concerns, while evaluation remains at an early stage.Source: §7 · p. 12
• Addressing model flaws and biases requires examining the broader social context reflected in training data, not only repairing the model.Source: §7 · p. 12

Future work

• The authors call for responsible development, appropriate regulation, and more robust ethical frameworks, and propose future analyses of regulation and user perspectives.Planned work includes evaluating regulation in the US, Europe, and Latin America and distributing a survey to professionals and researchers.Source: §7 · p. 12