Mapping the technological evolution of generative AI: a patent network analysis Authors: Navid Mohammadi, Jalil Heidary Dahooie, Amir Ali Bengari, Arash Rahimi Published: 2025 ABSTRACT AbstractThis study systematically maps the technological evolution and innovation landscape of Generative Artificial Intelligence (GAI) through a large-scale analysis of patent networks. Leveraging advanced text mining, network modeling, and community detection techniques on over 20,000 patents from Lens.org, we identify key technological domains, trace conceptual shifts, and highlight emerging trends across three major periods (pre-2016, 2016–2020, 2021–2025). Results reveal a pivotal transition post-2016 from early modular and domain-specific innovations, such as neuromorphic computing and bioengineering, toward integrated generative frameworks, API-driven platforms, and multi-modal capabilities. The post-2016 era is characterized by rapid growth in patent volume, increased conceptual diversity, and diminishing modularity, reflecting the convergence of previously distinct subfields. Key emergent clusters include generative AI frameworks, advanced medical imaging, personalized interactive systems, media authentication, and privacy-preserving AI. Notably, growing attention to content authenticity and user personalization underscores the interplay between technological maturation and societal concerns. Comparative analysis with baseline and alternative clustering models demonstrates the robustness and interpretability of the chosen network approach. While limited by reliance on English-language patents and TF–IDF extraction, this work provides an actionable roadmap for R&D leaders, policymakers, and investors to navigate the dynamic GAI landscape and informs future research directions involving multilingual and semantic-rich analyses.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-26810-7.Keywords: Generative artificial intelligence, Patent analysis, Technology roadmapping, Innovation clusters, Network modelingSubject terms: Complex networks; Complex networks; Information systems and information technology; Mathematics and computing; Science, technology and society Introduction Artificial intelligence is a broad concept and is defined as a set of technologies that can perform human cognitive tasks, such as perception, reasoning, planning, and prediction1. Conceptually, the development of artificial intelligence can be seen in three stages. The first stage is Symbolic AI, which has a knowledge-based approach. The second stage, which is based on deep learning, is called the data-driven approach. The third stage is Generative Artificial Intelligence, which is a combination of the previous two approaches, namely, knowledge-based and data-driven2. Artificial Intelligence has transformed industries and altered human-technology interactions. Generative Artificial Intelligence(GAI) is a novel and developing subset of Artificial Intelligence that represents an intersection of creativity and computation, fundamentally altering the essence of creation3. Since the release date of ChatGPT in November 2022, Generative Artificial Intelligence with the ability to enhance, facilitate, and simplify complex procedures has become popular for both individuals and business organizations4–6. ChatGPT is a prominent and good example of generative AI, and numerous studies and research predict that generative AI will have potentially significant impacts on the way companies work and interact with customers7. McKinsey & Company predicts that Generative AI is set to drive the subsequent surge in productivity, with the potential to create between $2.6 trillion and $4.4 trillion in value across various industries8. GAI systems have garnered considerable interest from various stakeholders, inciting vigorous discourse regarding their influence on business and society. Various fields such as business, education, Healthcare, and Content Creation have benefited from GAI9. One critical aspect of GAI is the theoretical and mathematical aspect, which plays an essential role in GAI models. The outputs of GAI are directly dependent on the models and machine learning techniques applied to it, which are mainly VAEs, GANs, transformers & autoencoders, and Large Language Models10. These models underpin various generative tools across multiple modalities, including text, image, code, video, speech, and music generation. Well-known instances include ChatGPT, Midjourney, Dall-E2, Copilot, Sora, and MusicLM, which are being launched by leading technology companies and emerging AI start-ups worldwide11. With this rapid advancement of this technology, it is essential to chart the evolving landscape of GAI12. In Particular, a patent analysis approach can be used to better understand the speed and direction of GAI progress and track innovation trajectories13,14. Since Patents are widely recognized as a reliable indicator of innovation and technological progress12, conducting patent analysis becomes an essential tool for identifying technology trends and gaps, monitoring advancements through time, and understanding technology landscapes15. Moreover, the integration of machine learning techniques into patent analysis has emerged as a cutting-edge approach, enabling researchers to extract deep insights from large and complex patent datasets16. Some papers from the literature have studied the adoption of this technology in various sectors such as manufacturing, Operations, and Supply chain6,17,18, Service industry19, hospitality and tourism20, healthcare and medicine21,22, education23,24, Business and management innovation25–28. The impact of GAI on decision-making has been studied by29. Brüns and Meißner30 studied GAI in content creation and marketing. Alongside the benefits gained from GAI, ethical and legal issues have arisen regarding the use of this novel technology. Authorship and Academic Integrity, Regulatory Issues, Privacy, Trust, Bias, Misinformation, Deepfakes, Educational Ethics, Transparency and Accountability, Authenticity and Attribution, and Social and Economic Impact are the challenges of GAI31. GAI also has some technological issues that are worth mentioning. These issues include hallucination, training data quality, interpretability of GAI outputs, authenticity, and prompt engineering. For instance, hallucination refers to the generation of inaccurate or fabricated information by GAI models, which poses a challenge for reliability. Likewise, the quality and biases in training data significantly impact the fairness and accuracy of generated outputs. Furthermore, the interpretability of GAI outputs remains limited, making it difficult for users to understand or validate the reasoning behind generated content. These challenges are further compounded by the need for effective prompt engineering, where small changes in input prompts can lead to vastly different outcomes, raising concerns about consistency and control in real-world applications32–34. Despite the fact that extensive attention has been paid to applications and implications in prior works, there is still a notable gap in understanding how GAI technologies themselves are evolving from a technological and innovation perspective. Particularly, technological trajectories reflected from patent data have been largely missed in the literature, which inhibits our understanding of future development paths of GAI tools and platforms. Although previous patent-based studies have examined technological evolution in domains such as blockchain35, Quantum technologies36, Internet of Things37, and Artificial Intelligence in construction38, Systematic efforts focusing on Generative AI remain scarce. To address this gap, this paper seeks to systematically map the technological evolution of Generative AI through patent analysis, integrating scientometric visualization techniques such as network mapping, co-classification analysis, and clustering of technological areas to identify dominant technological areas, track their evolution, and uncover emerging innovation clusters. This provides a conceptual foundation for analyzing the technological trajectory and identifying emerging innovation clusters within Generative AI. In contrast to prior works that primarily focused on sectoral adoption and ethical implications, this study shifts attention to a perspective that investigates the technological evolution of GAI by employing patent-based scientometric analysis, which is absent from the current discourse. Therefore, this paper seeks to answer the following questions: What are the dominant technological areas in GAI patent filings? What are the dominant technological areas in GAI patent filings? How has innovation in Generative Artificial Intelligence evolved? How has innovation in Generative Artificial Intelligence evolved? What technological trends can be identified for the future of generative AI? What technological trends can be identified for the future of generative AI? The key contributions of this study are threefold. First, this study advances methodological frontiers by integrating patent analysis with scientometric visualization and community detection, offering a novel mixed-methods framework for exploring technological evolution in GAI. Second, it contributes to theory by addressing the overlooked gap in understanding the technological trajectory of GAI, moving beyond prior research that mainly emphasized sectoral adoption and ethical implications. Third, it provides foresight and practical value by outlining strategic implications for policymakers, industry leaders, and academic researchers, enabling evidence-based navigation of the rapidly evolving GAI landscape. Further sections of this paper proceed as follows: First, a literature review that contextualizes previous research on GAI technologies, followed by a methodology section outlining a mixed methods approach combining text mining, network modeling, and community detection to analyze patent data. The results section presents key classifications and emerging technological trends. The discussion section, as the fourth part of this article, further interprets the results, and finally, the main contributions alongside practical and theoretical implications are addressed in the conclusion. Literature review An overview of generative AI Artificial Intelligence (AI) has emerged as one of the transformative technologies of the twenty-first century. It refers to the ability of machines to perform tasks that typically require human intelligence, such as learning, perception, reasoning, and decision-making. Over the past decade, AI has undergone significant developments, enabling machines to understand conversations, process images, and translate languages39. Generative Artificial Intelligence (GAI) refers to AI systems designed to create content, including text, image, audio, and video, by producing innovative and original outputs based on their training data. It represents a specialized domain within the broader field of artificial intelligence that empowers machines to autonomously produce original content, encompassing text, images, audio, and even programming code, derived from intricate patterns identified within extensive datasets. Unlike conversational AI, generative AI responds to queries and generates the content within those response40. In contrast to traditional AI, which predominantly emphasizes tasks such as analysis, classification, and prediction, Gen AI engages in the active generation of innovative outputs without the necessity for explicit programming directives41. This capability differentiates it from analytical AI and conventional machine learning, which rely on structured data to make decisions but do not produce entirely new content. With the advancements in AI during the Fourth Industrial Revolution, tools like ChatGPT have become increasingly popular, transforming the way content is produced and created42. Understanding basic concepts such as natural language processing (NLP), large language models (LLMs), and Transformers is necessary to comprehend Gen AI technology better. Natural language processing (NLP): Natural Language Processing (NLP) aims to process, comprehend, and generate human languages, enabling computers to ultimately interpret speech in a way that mimics natural human understanding43. Natural language processing (NLP): Natural Language Processing (NLP) aims to process, comprehend, and generate human languages, enabling computers to ultimately interpret speech in a way that mimics natural human understanding43. Large language model (LLM): A language model (LM) enables the prediction of the most probable following words in a text, making it possible to suggest completions for emails or messages44. Large language model (LLM): A language model (LM) enables the prediction of the most probable following words in a text, making it possible to suggest completions for emails or messages44. Transformer: A transformer, or transformer model, is an artificial neural network that converts input from one format to another. Simply put, it processes an input and generates an output, which may sometimes be in a different format. For instance, a transformer can take a written prompt as input and produce an output such as an essay, an email, an image, or even audio45. Transformer: A transformer, or transformer model, is an artificial neural network that converts input from one format to another. Simply put, it processes an input and generates an output, which may sometimes be in a different format. For instance, a transformer can take a written prompt as input and produce an output such as an essay, an email, an image, or even audio45. The applications of Gen AI span multiple domains, including business, healthcare, education, finance, and manufacturing. In healthcare, it facilitates pharmaceutical discovery and medical imaging analysis by producing molecular configurations and enhancing diagnostic accuracy46. In education, it powers personalized learning, adaptive assessments, and intelligent tutoring systems47. Gen AI has also disrupted creative fields like digital art, music, and content writing by generating human-like, high-quality outputs48. In business, it enhances customer service, marketing, and financial modeling, optimizing operations, and boosting productivity49. Despite its vast potential, Gen AI poses several challenges. Ethical concerns are prominent, particularly regarding misinformation, deepfakes, and intellectual property issues50,51. Socioeconomically, Gen AI’s rapid adoption has sparked debates on job displacement, with concerns that automation could replace creative and knowledge-based professions52. Additionally, the digital divide may widen as access to cutting-edge AI technologies remains limited to well-resourced organizations and economies31. In conclusion, Generative AI represents a groundbreaking technological advancement with far-reaching implications across industries. While it fosters innovation and efficiency, the challenges associated with its implementation, ethics, and societal impact necessitate careful consideration. Addressing these challenges requires robust regulatory frameworks, ethical AI design principles, and interdisciplinary collaboration to ensure the responsible deployment of Gen AI across industries53. GenAI applications Generative AI models are revolutionizing content creation across various formats, such as text, images, videos, and audio. We will explore different areas in the following: Text generation: Large language Models (LLMs) such as GPT, T5, and BERT enable Generative AI to produce text-based content by leveraging Transformer Architectures and self-attention mechanisms54. Text generation: Large language Models (LLMs) such as GPT, T5, and BERT enable Generative AI to produce text-based content by leveraging Transformer Architectures and self-attention mechanisms54. Image and video generation: Generative AI in the field of image and video generation creates or modifies visual content by leveraging models such as GANs, Diffusion Models, and Transformer-based vision Models11,55. Image and video generation: Generative AI in the field of image and video generation creates or modifies visual content by leveraging models such as GANs, Diffusion Models, and Transformer-based vision Models11,55. Speech and music generation: Generative AI in Audio processing focuses on producing human-like Speech, Music, and sound effects. This involves using deep learning models trained on large data sets of speech or musical compositions. One of the most important areas in this field is Text-to-Speech(TTS) generation, where AI models convert written text into natural-sounding speech56. Speech and music generation: Generative AI in Audio processing focuses on producing human-like Speech, Music, and sound effects. This involves using deep learning models trained on large data sets of speech or musical compositions. One of the most important areas in this field is Text-to-Speech(TTS) generation, where AI models convert written text into natural-sounding speech56. Code generation: Generative AI provides tools for writing code using large language models (LLMs) like GPT. These models can turn plain language instructions into working code because they have learned from many code examples in different programming languages. For example, OpenAI’s Codex can automatically create code blocks, understand descriptions of what a user wants a program to do, and generate the appropriate code57. Code generation: Generative AI provides tools for writing code using large language models (LLMs) like GPT. These models can turn plain language instructions into working code because they have learned from many code examples in different programming languages. For example, OpenAI’s Codex can automatically create code blocks, understand descriptions of what a user wants a program to do, and generate the appropriate code57. Recent breakthroughs have significantly enhanced the capabilities of Generative AI. For instance, the emergence of large-scale pretrained models like GPT and DALL-E has showcased unmatched scalability and adaptability across various domains such as Education, Healthcare, Agriculture, Lifestyle, Marketing, Gaming, Entertainment, E-commerce, Finance, Industry, and Travel & Transport58. The applications of generative artificial intelligence in various industries are summarized in Table 1. Table 1Generative AI applications across industries.SectorKey applicationsBenefitsChallengesValidation studiesEducationPersonalized learning, automated content creation, intelligent tutoring systemsReal-time feedback, complex concept summarization, and providing study guidesReliability and precision, plagiarism, accurate performance assessment59HealthcareMedical research, clinical decision-making, and personalized medicineDrug discovery, improved disease prediction, and patient support systemsData privacy and regulatory concerns, clinical validation60AgricultureCrop disease detection, soil health analysis, and precision farming techniquesData-driven yield improvementQuality of data infrastructure dependencies, ethical issues61,62Lifestylediet planning, travel recommendations, wardrobe selection, and hobby curationHyper-personalized user suggestionsPotential privacy intrusions58,63,64MarketingContent automation, targeted advertising, and sentiment analysisCampaign optimization, personalized messagingBrand alignment, data bias, ethical transparency65,66GamingDynamic narratives, NPC interactions, personalized contentEnhanced player immersionCompute-intensive, Data scarcity, Unpredictable outputs,67,68EntertainmentScriptwriting, content recommendation, interactive storytellingCreative content generation, personalized user experiences, and Production process streamliningCopyright ambiguity11E-commerceAutomated support, product recommendations, and fraud detectionSearch optimization, user experience enhancementBias in recommendation algorithms27,69FinanceFraud detection, risk assessment, and algorithmic tradingReduction in banking fraud cases, faster market analysis, and portfolio optimizationInterpretability limits of complex models, Bias risks in historical data, regulatory hurdles70,71IndustrySupply chain optimization, production automationOperational efficiency gains assist sustainability effortsOver-automation risks, inventory distortion from demand misprediction72,73Travel and transportRoute optimization, itinerary planning, and real-time traffic managementCustomer service enhancementLiability allocation issues74 Generative AI applications across industries. The transformative impact of generative AI on creative and operational processes across sectors, from improving educational experiences and streamlining healthcare innovations to personalizing user engagement in lifestyle and marketing, all demonstrate that generative AI is redefining traditional workflows. Despite the significant benefits of this technology, there are also challenges such as ethical concerns, data quality, and regulatory compliance that need to be addressed. As adoption of this technology accelerates, a balanced approach that considers both innovation and responsibility is needed. Key technologies and components of GenAI Generative Artificial Intelligence is a transformative class of AI enabling the creation of novel human-like content in response to given prompts75. GAI represents a form of unsupervised learning that generates new data samples without direct instructions regarding the desired outputs. The model receives training from an extensive dataset of examples, facilitating its ability to detect and internalize the fundamental patterns and structures embedded in the data. The primary classifications of generative models encompass: transformers, generative adversarial networks (GANs), variational autoencoders (VAEs), autoregressive models, and Diffusion Models76. The definitions of models and their comparison are mentioned in Table 2. Table 2Comparison of key generative AI models.ModelDescriptionArchitectureApplicationsStrengthsLimitationsReferencesTransformersNeural networks for sequential data analysis using self-attentionEncoder-decoder with self-attention mechanismText generation, translation, summarizationHandles long-range dependencies; scalableRequires large datasets and computing power77GANsModels that generate new data through adversarial training between the generator and discriminatorTwo networks: generator versus discriminator in a game-theoretic setupImage synthesis, super-resolution, image-to-image translationProduces sharp, high-fidelity samplesTraining is unstable; risk of mode collapse78,79VAEsProbabilistic generative models encoding data into a latent spaceEncoder maps to distribution; Decoder reconstructs samplesData compression, image reconstruction, anomaly detectionLatent space allows interpolation; stable trainingGenerates blurrier outputs than GANs76Autoregressive modelsModels that predict the next value from Previous values in a sequenceLinear regression Over past data pointsLanguage modeling, forecasting, Speech synthesisSimple, interpretable Good for time seriesAssumes stationarity; limited to Short-range patterns80Diffusion modelsGenerate data via an iterative noise addition/removal processForward: noise added; Reverse: noise removed step-by-stepImage, video, and audio generationProduces diverse, high-quality outputsComputationally expensive and slow sampling81LLMsDeep learning models trained on massive text data to generate coherent textTransformer-based; trained with an autoregressive objectiveChatbots, summarization, translation, and question-answeringContext-aware, human-like output; generalizableProne to hallucination; high resource demands82 Comparison of key generative AI models. Related works The potential for using generative AI has increased due to rapid technical advances and innovations in the architecture and methods of training AI models, leading to more scalable, efficient, and multi-modality generative AI tools. Given the rapid changes taking place in the field of Gen AI technology, it is necessary to outline its future vision. Patent data is a reliable source for examining technological advances and innovation. By text-mining records and extracting patents related to Gen AI technology, patent data can be used to outline this vision. Several previous studies have successfully employed patent data to identify technological opportunities across various domains, as summarized in Table 3. Table 3Summary of key studies using patent analysis for identifying technological opportunities.Article author(s)Research methodImportant resultsSong et al.13Text mining of patent documents using F-term classificationIdentified new technological opportunities based on technical features in patent documentsKim and Bae83Patent analysisForecast promising technologies through detailed information in patentsYoon and Magee84Visualization of patent information using generative topographic mapping and link predictionExplored technology opportunities by visually mapping patent data relationshipsKline et al.85Empirical study of patent applications and economic indicators in U.S. firmsDemonstrated link between patenting activity and outcomes like employee compensation and productivityZhang et al.35Integration of patent analysis with technology roadmapCreated a technology roadmap in the blockchain domain based on patent dataNoh et al.86Patent-based opportunity analysis for roadmap constructionProposed an opportunity-driven technology roadmap for the 5G mobile service industrySu et al.87Patent analysis on 5G-V2X autonomous driving technologiesPredicted growth trends, identified top countries and assignees, and tracked regional technology development over timeMetzger et al.88Patent analysis and text mining focused on circular economy technologiesFound increased patent activity in battery technologies supporting the circular economy, especially in Asian and European innovation clustersSingh and Sai Vijay89A combination of patent analysis, semantic network analysis, and topic modelingProvided a multifaceted forecast of e-commerce technologiesMaghsoudi et al.90Patent extraction and expert evaluation using the S-curve modelIdentified 11 key technology clusters in the circular economy, analyzed their maturity stages, and highlighted a mix of emerging and mature technologies for targeted investment and research Summary of key studies using patent analysis for identifying technological opportunities. Studies above show that patents are a valid proxy measure for technological developments. As we have seen in the Related Works section, various studies have used patents to examine the path of technology and its prospects in various fields, such as blockchain, augmented reality, circular economy technologies, inventions and technologies related to self-driving cars, and technologies in the field of e-commerce. However, by reviewing the literature related to generative artificial intelligence in Table 4 We realize that despite the rapid pace of change in generative artificial intelligence technology and the wide applications of this technology in various fields, including education, health, e-commerce, etc., few studies have examined the path, prospects, and stages of maturity of this technology using the patent method. In this study, we seek to fill this gap. Table 4Summary of generative AI related works.Article author(s)Research methodImportant resultsPathak and Pallasena97Bibliometric reviewThe study reveals four core themes from the review, which relate to (a) Advancements in Gen-AI Architecture and Frameworks, (b) Gen-AI Applications, (c) Model Validation and Benchmarking, and (d) Systemic ConsiderationsKim et al.98Scientometric and text mining techniquesExamining the security implications of generative artificial intelligence (GAI), focusing on models such as ChatGPT. As GAI technologies are increasingly integrated into industries like healthcare, education, and media, concerns are growing regarding security vulnerabilities, ethical challenges, and potential for misuseSingh et al.12Structural topic model (STM)Identifying generative AI applications & developing a roadmap for the near, mid, and far future using Structural topic modeling (STM) to discover latent semantic patterns and identify the key application areas from a text corpus comprising 2398 patents published between 2017 and 2023Ahmed et al.99Systematic literature reviewGAI applications in education, personalization, task automation, teacher assistance, and learning efficiencyDwivedi and Elluri100Bibliometric analysis, text analysis, social network analysisExplore the generative AI academic research in the past decade. Specifically, research themes and their relationships, the evolution of research themes over time, and prominent authors, articles, journals, institutions, and countries publishing in generative AI are identifiedBalasubramaniam et al.101Literature reviewproviding a comprehensive review of generative AI’s evolution, significance, and applications, including the foundational architecturesRay102Literature reviewDelving into the background, applications, key challenges, and future directions of ChatGPT by conducting a comprehensive reviewAbadi and Pecht103Patent analysisThe paper analyzes AI-related patents, highlighting trends such as the significant increase in patents from 4598 in 2008 to 20,639 in 2018, indicating a growing focus on generative AI techniques and applications within research and development activities Summary of generative AI related works. To contextualize methodological approaches adopted in this study, scientometric and visualization techniques applied in other emerging technological domains have been reviewed. Liu et al.91 proposed an integrated approach combining scientific papers and patent data to map technology evolution paths, demonstrating that analyzing multiple data sources and performing gap analysis can enhance forecasting of technology trends and identification of innovation opportunities, as shown in their case study on all-solid-state lithium-ion batteries. Technology evolves through a cumulative process, which is driven by knowledge. Hence, studying these knowledge-guided paths known as technology trajectories reveals the gradual direction of technological advancement and provides a foundation for technology forecasting92. Beyond the domain of generative AI, several studies based on scientometric and visualization have been conducted to map the technological trajectories of other emerging fields. Rawat and Yadav93 analyzed quantum computing applications across seven scientific domains using trends and visual analytics, uncovering key hotspots and emerging research frontiers. Similarly, Rawat and Yadav94 investigated the intellectual landscape of artificial intelligence within the domain of environmental sustainability, emphasizing thematic evolution, citation bursts, and the growing role of hybrid AI approaches. Rawat and Sharma95 mapped the knowledge structure of quantum computing, identifying influential contributors and collaboration networks, and Coccia96 examined technological trajectories in quantum computing using both scientific publications and patent data, identifying path-breaking directions such as quantum optics, quantum information, and quantum communication, thereby emphasizing the role of technology trajectory analysis in guiding industrial and societal transformation. Although scientometric and patent-based mapping have been successfully applied to other domains such as blockchain, quantum computing, and circular economy technologies, comparable systematic efforts are scarce in the context of generative AI. Considering rapid advancement, diverse applications, and cross-sectoral impact of GenAI, a patent-based network analysis can provide valuable insights into emerging technological clusters, innovation hotspots, and maturity stages. This study seeks to address this gap by integrating patent data with scientometric visualization to comprehensively capture the technological evolution and innovation trajectories of GenAI. While earlier works have applied patent analysis to domains such as blockchain, 5G, quantum computing, and the circular economy, no systematic effort has been made to investigate the technological trajectory of Generative AI through patent data. Addressing this gap, the present study integrates patent analysis with scientometric visualization to capture the evolution, maturity stages, and innovation hotspots of GenAI technologies. Methodology This study aims to conduct a conceptual analysis of generative artificial intelligence and identify technological trends by using text mining techniques, term weighting, network modeling, and community detection. This integrative approach was selected to leverage the strengths of both quantitative and qualitative methods. Therefore, the research process has been designed and implemented in several systematic and sequential steps, as seen in Fig. 1. Research process. Patent data collection and preparation Patent retrieval from Lens.org In the initial phase, the required data was collected from the Lens.org database. Lens is a reliable and openly accessible platform that integrates scholarly and patent information from various scientific and industrial fields. As one of the most reputable patent databases with global coverage, it has a convenient API for extracting structured data. Lens.org has been increasingly utilized in recent scientific studies and systematic reviews for retrieving and analyzing patent data due to its comprehensive coverage and advanced search capabilities104. Keyword selection and query strategy To collect all the data related to generative artificial intelligence, the keywords listed in Table 5 were selected and searched in the Lens.org database. This choice is aligned with similar patent-mining studies12, where GenAI-related queries were used to construct datasets for text-mining and roadmap analysis. Table 5The query searched in the lens database.Search queryDatabase(“Generative AI” OR "Generative Artificial Intelligence" OR “GAN” OR "Generative Adversarial Network" OR "Computer-Generated Content" OR “Transformer Architecture” OR “Diffusion Model” OR "Large Language Model" OR “LLM” OR “Midjourney” OR "DALL-E" OR “CLIP” OR "Contrastive Language-Image Pre-training" OR “GPT” OR "Generative Pre-trained Transformer" OR “BERT” OR "Bidirectional Encoder Representations") AND (“artificial intelligence” OR “machine learning” OR “neural network” OR “deep learning”)Lens.org The query searched in the lens database. Extracted fields and data description Among all the available information fields, the patent summary column (Abstract) was selected for textual analysis as a rich and concise text section that introduces the structure and innovation of the registered patent. Other columns of information, including publication date, priority numbers, technical classifications, and legal status, were also stored as auxiliary data, but the analytical focus of the study was solely on the content of the abstracts and the date of patent publication90. Baseline method for textual analysis and semantic mapping framework Text preprocessing and corpus cleaning After collecting the data, text preprocessing was applied to the abstracts. This stage included initial cleaning of the text from language noises such as punctuation marks and unnecessary characters, converting all letters to lowercase, and removing stop words. To maintain the conceptual accuracy of the vocabulary, stemming or lemmatization was intentionally avoided. Thus, a clean version of the abstracts’ text was prepared to be the basis of linguistic analysis in the next steps. Text preprocessing was performed using standard NLP libraries in Python, including stopword lists and regular expressions. Temporal analysis and inflection point identification Next, to identify conceptual developments over time, a temporal distribution analysis of patents was performed. For this purpose, the number of registered patents was extracted by year, and a publication trend graph was generated. To better detect inflection points in the growth trajectory, the vertical axis of the graph was set to a logarithmic scale. Given the nonlinear growth of the number of patents published in different years, a semi-logarithmic chart was used to determine the intervals under study. Using the Logarithmic Transformation Formula for the Y-axis (Eq. 1), the exponential growth patterns will be revealed: where Y is the original value on the vertical axis (e.g., number of patents per year), Y′ is the transformed value plotted on the log scale, logb denotes the logarithm with base b (commonly b = 10 or e). Analysis of the graph showed that 2016 was a significant change point in patent growth and possibly in technological leaps. Regarding the post-2016 era, the acceleration of generative AI developments following the public release of ChatGPT in late 2022 was a major motivation to further divide this period, to capture the impact of ChatGPT and similar products. Therefore, the data were divided into three periods, “before 2016”, “2016–2020”, and “2021–2025” to allow the comparison of conceptual trends in three different periods. In the next step, the Term Frequency–Inverse Document Frequency (TF–IDF) method was used (Eqs. 2–4) to extract index keywords in each period. where : Frequency of term t in document d, N: Total number of documents in the corpus, : Number of documents in which term t appears. The TF–IDF model, which is based on the relative frequency of words in a document compared to their frequency across the corpus, assigns a statistical weight to each word. For each interval, the TF–IDF weight of all words was calculated, and a certain number of words with the highest score were selected as important keywords. These dominant expressions represented the core concepts of patents of each period. The TF–IDF method was applied to the abstracts at both the unigram (single-word) and bigram (two-word combinations) levels. For each temporal segment, different values of top-N keywords were examined to assess robustness. Specifically, the top N = 15 keywords per document were used in the main method, while N = 10 and N = 25 were additionally tested for sensitivity analysis. Keyword extraction and semantic co-occurrence mapping After extracting the keywords, the co-occurrence technique was applied to examine the semantic associations. Co-occurrence means the simultaneous appearance of two words in a document-level co-occurrence (in this research, patent abstract), which can be a sign of a semantic or functional relationship between those two words. In this study, keyword pairs were enumerated from among the keywords, and if they occurred simultaneously in at least one abstract, an edge between them was defined. The Co-occurrence frequency count equation is shown in Eq. 5: where : Indicator function that returns 1 if the condition is true, 0 otherwise, D: Set of all documents (patent abstracts), : Keyword pair. This equation measures the number of documents in which two keywords co-occur. An edge is created between two nodes if their co-occurrence count exceeds a predefined threshold. To maintain the coherence and interpretability of the graph of the network, a minimum threshold of frequency was also considered to eliminate weaker relationships. The outcome of this process was a network of co-occurrence-based links between keywords for each period. The co-occurrence network was modeled as an undirected weighted graph, where each edge’s weight corresponded to the number of abstracts in which the keyword pair co-occurred. Word co-occurrence was computed at the document (abstract) level. An edge was established if two keywords appeared in the same abstract. During text preprocessing, frequent yet semantically non-informative words (such as domain-specific stopwords) were removed. The preprocessing phase included: (i) converting all text to lowercase, (ii) removing punctuation marks and special characters, and (iii) eliminating stopwords using an extended list combining the NLTK English stopword list with domain-specific terms. To preserve semantic granularity, stemming and lemmatization were intentionally avoided. The entire preprocessing pipeline was implemented in Python using the NLTK and re libraries. However, no minimum frequency threshold was applied during the construction of the co-occurrence network, and all remaining terms were included for link generation. Community detection and network visualization with gephi Then, these networks were prepared in the form of structured edge-node files for each period and transferred to the Gephi software environment. As one of the powerful tools in analyzing and visualizing complex networks, Gephi serves as a robust platform for network analysis and visualization. This software offers the Louvain algorithm, which was implemented for each network to identify conceptual clusters. The Louvain algorithm was selected for community detection due to its high scalability and ability to uncover densely connected clusters with optimal modularity. As reported in recent comparative studies105, Louvain demonstrates higher computational efficiency and scalability compared to alternative community detection techniques such as Girvan-Newman or label propagation. The Louvain community detection algorithm was executed with a resolution parameter of 1.0 and randomization enabled, aiming to maximize the modularity index for optimal cluster separation. Equation 6 describes the formulation below: where : Adjacency matrix of the graph, : Degrees of nodes i and j, m: Total number of edges in the network, : Communities of nodes i and j, : Kronecker delta (1 if , 0 otherwise). Modularity is a measure of the strength of division of a network into communities. The algorithm operates based on the maximization of the Modularity index and categorizes words into groups with the highest intra-cluster communication density. Also, to increase the visual clarity of the graph and the separation of the clusters, the Force Atlas 2 algorithm was applied in Gephi. Thematic clustering and conceptual labeling Finally, by examining the output from Louvain, the resulting thematic clusters were examined. For each cluster, the list of extracted words was sorted according to the amount of repetition. Then, by analyzing the meaning of the most frequent words in each cluster, an attempt was made to determine the semantic theme or conceptual label. These clusters represented the most frequent core thematic domains and dominant themes in each period. The comparative analysis of the clusters in two defined periods allowed the identification of emerging concepts, highlighted or reduced issues, and the evolution of technological paths. This methodological pipeline enabled a structured exploration of conceptual dynamics within generative AI patents, offering insights into evolving technological domains. Comparative models In order to evaluate the accuracy and stability of the results of the original model, several other methods for topic extraction and document clustering were used. These models were considered comparative tools to clarify the strengths and weaknesses of the original approach. In this regard, the following four models were selected and implemented as comparative models: K-Means algorithm: a classical and basic clustering algorithm, K-Means algorithm: a classical and basic clustering algorithm, Latent dirichlet allocation (LDA): a probabilistic model based on topic generation, Latent dirichlet allocation (LDA): a probabilistic model based on topic generation, Non-negative matrix factorization (NMF): a method based on matrix decomposition and non-negative factorization, Non-negative matrix factorization (NMF): a method based on matrix decomposition and non-negative factorization, BERTopic: an advanced model based on deep learning and semantic text representation, which combines language embeddings, dimensionality reduction, and density-based clustering to identify and extract latent topics. BERTopic: an advanced model based on deep learning and semantic text representation, which combines language embeddings, dimensionality reduction, and density-based clustering to identify and extract latent topics. This algorithmic variation provides the possibility of comparing the results from the point of view of the number of clusters, topic coherence, lexical diversity, and also the stability of clusters in different periods. In the following, each of these methods is introduced separately. K-means algorithm The K-Means algorithm is one of the most widely used clustering methods that divides data into k clusters. In this method, each data point is assigned to the nearest cluster center (centroid), and the cluster centers are iteratively updated to minimize the objective criterion. The objective function is defined as follows(Eq. 7): where Ci is the set of cluster i points and μi is the mean of that cluster. Latent dirichlet allocation (LDA) LDA is a probabilistic generative model for discovering hidden topics in documents. In this model, it is assumed that each document is a combination of a distribution of topics, and each topic contains a distribution of words. In this method, for each document d, a topic distribution is chosen. Then, for each word in the document, the topic is first sampled. Then, a word is selected from the topic distribution . The overall likelihood function of the document set is expressed as Eq. 8: where α: is the Topic distribution parameter, β: is the Vocabulary distribution parameter, dn: is the Latent variable for assigning words to topics. Non-negative matrix factorization (NMF) The NMF method is a matrix decomposition technique that decomposes the word-document matrix X into two non-negative matrices W and H. This non-negativity constraint makes the results more interpretable. The formula is as Eq. 9: And the goal is to minimize the error, which is defined as Eq. 10: where X: Document–Word Matrix, W: Document–Topic Matrix, H: Topic–Word Matrix. BERTopic BERTopic is a modern topic modeling approach that leverages transformer-based embeddings and clustering methods to identify coherent latent topics from large text corpora. Unlike classical models such as LDA or NMF, BERTopic is based on semantic vector representations and density-based clustering, making it robust for short texts, multilingual corpora, and complex semantic patterns. The method involves four main steps: 1. Text embedding Each document di is transformed into a dense semantic vector using a pre-trained language model such as BERT, RoBERTa, or Sentence-BERT (Eq. 11): where ei is the embedding of document di, and m is the embedding dimension. 2. Dimensionality reduction (UMAP) Since embeddings are high-dimensional, BERTopic applies UMAP (Uniform Manifold Approximation and Projection) to reduce the dimensionality while preserving local semantic structure (Eq. 12): 3. Density-based clustering (HDBSCAN) The reduced embeddings zi are clustered using HDBSCAN, which identifies dense regions and assigns documents to clusters without requiring a predefined number of topics (Eq. 13): where T is the number of discovered topics, and each cluster cj corresponds to one topic. Documents not fitting any cluster are labeled as outliers. 4. Topic representation (c-TF–IDF) As shown in Eq. 14, for each cluster cj, BERTopic constructs a topic vector by aggregating the documents in the cluster and applying class-based TF–IDF (c-TF–IDF): where f(t,cj): frequency of term t in cluster cj, V: vocabulary, N: total number of clusters, nt: number of clusters in which term t appears. This formula highlights terms that are frequent within a topic but rare across other topics, producing interpretable topic keywords. Overall, BERTopic integrates transformer embeddings, UMAP dimensionality reduction, HDBSCAN clustering, and c-TF–IDF weighting to extract coherent topics with high semantic quality. It offers flexibility and interpretability while avoiding some limitations of probabilistic models like LDA. Results This section presents the findings from the analysis of the collected patents. The results include examining the time trends of patent registrations, extracting and analyzing conceptual keywords, identifying semantic clusters, and analyzing the network structure between key concepts. The section tries to provide an accurate picture of technological developments in generative artificial intelligence by relying on processed data. Collecting data from the Lens database In this research, the data were collected from the Lens database. To collect all the data related to generative artificial intelligence, a complex query was searched in the Lens.org database (Table 5). This choice is aligned with similar patent-mining studies12, where GenAI-related queries were used to construct datasets for text-mining and roadmap analysis. In this database, 172,489 patents in the field of Generative AI have been collected as of September 1, 2025. In Fig. 2, a graph of the patent publication trend can be seen. This chart illustrates the trend of patent issuance related to generative AI from 1986 to 2025. Since 2016, there has been a significant growth in the number of patents, peaking at 33,234 in 2024, indicating a considerable increase in research and development activities in this area. The number of patents published each year (indicates sharp growth in publication volume). Figure 3 illustrates the patenting activity of the leading companies in the field of Generative AI over time. As shown, there is a sharp increase in patent registrations starting around 2018, with a peak in 2021. Major technology corporations such as Adobe, Amazon, Google, Intel, Microsoft, NVIDIA, and Samsung demonstrate a dominant role in shaping this technological domain. In particular, NVIDIA and Microsoft have considerably intensified their efforts in recent years, reflecting their strategic focus on advancing Generative AI technologies and applications. This trend highlights both the growing competition and the consolidation of innovation among a limited number of key industrial players. Number of patents registered by top companies over time, showing a sharp rise in generative AI–related patenting activity after 2016, with significant contributions from firms such as Microsoft, Google, IBM, and Samsung. As discussed in the Section "Density-based clustering (HDBSCAN)", a logarithmic transformation was applied to determine optimal time intervals based on the distribution of patent registrations. The semi-logarithmic chart can be seen in Fig. 4. Considering the number of patents published in each year and the change in the growth rate in 2016, the time intervals are considered before and after this year and compared with each other. Based on the figure, the growth rate of patent publication after 2016 is higher than in the other periods. In particular, the rapid surge in developments after the public release of ChatGPT in late 2022 highlighted the need to distinguish this acceleration. To reflect these dynamics, the post-2016 period was further divided into two separate intervals, “2016–2020” and “2021–2025,” enabling a clearer comparison of technological and conceptual shifts over time. Semi-log patents publications over time. Based on the graph, the difference between the growth rate of the pre-2016 and post-2016 periods is significant. Co-occurrence network analysis of keywords in generative AI patents using TF–IDF and Louvain algorithms After selecting periods according to the growth of patent publication, the keywords of the patents in each period were extracted using the TF–IDF method. Then, the keyword co-occurrence graph extracted from patents was analyzed using Gephi software. To construct this graph, the TF–IDF weighting method was used to select important keywords, and the co-occurrence matrix was used to determine the relationship between words. This matrix was then plotted as a directed graph in Gephi. Finally, to identify conceptual structures in the network, Louvain’s algorithm for community detection was used. Before 2016 Before 2016 By applying the Louvain algorithm using Gephi software, a total of 11,283 nodes and 170,395 edges were identified in the pre-2016 network, reflecting the early-stage complexity and evolving semantic structure of generative AI concepts. The clustering output is stored as a Modularity Class in Gephi. The modularity index is 0.4768, which indicates the acceptable quality of cluster separation in the network. The closer this number is to 1, the better the separation between clusters. Figure 5 represents the network diagram in this period. Pre 2016 network diagram (The selected Nodes are the ones with a degree greater than 300). In the pre-2016 period, as illustrated in Fig. 6, the dominant keywords such as “data”, “user”, “content”, and “device” highlight that patents in this timeframe were primarily oriented toward information processing, user interaction, and digital content management. The frequent occurrence of terms like “image”, “signal”, “media”, and “video” further reflects the technological emphasis on multimedia systems, communication devices, and data-driven applications. Overall, the focus during this period was less on generative AI in its modern sense and more on foundational infrastructures for handling and transmitting digital information. Word cloud of the top 100 most frequent words before 2016. 2016–2020 2016–2020 The Louvain algorithm was used to identify conceptual clusters. The results of this algorithm were recorded as the Modularity Class variable, where the final graph consists of 10 distinct clusters. The 2016–2020 semantic network comprises 21,741 nodes and 348,227 edges, illustrating the vast scale and density of keyword co-occurrences in this period, and the Modularity Score is 0.3939, which is a minor decrease compared to the pre-2016 network. A network diagram of the 2016–2020 period is illustrated in Fig. 7 for selected Nodes. Network diagram of 2016–2020 period (nodes with degree greater than 500 are selected). In the 2016–2020 period, as shown in Fig. 8, the prominence of terms such as “image,” “device,” “network,” “model,” and “data” reflects a clear technological shift toward computer vision, connected systems, and AI-based modeling. Keywords like “neural,” “training,” and “processing” highlight the growing adoption of machine learning approaches, while emerging terms such as “virtual,” “sensor,” and “vehicle” point to the expansion of generative AI applications into areas such as autonomous systems and immersive technologies. Word cloud of the top 100 most frequent words between 2016 and 2020. 2021–2025 2021–2025 In the 2021–2025 network, the Louvain algorithm was applied to detect communities, yielding 80,760 nodes and 1,048,575 edges. This illustrates the substantial expansion and structural density of generative AI–related concepts in the most recent period. The modularity score was calculated as 0.3728, indicating a moderate quality of community separation within this highly interconnected semantic network. Figure 9 presents the network diagram for this time interval. Post 2021 network diagram, containing the nodes with degree greater than 1500. As illustrated in Fig. 10, in the 2021–2025 period, the dominant words such as “image”, “data”, “user”, and “video” highlight the centrality of large-scale multimedia content and user-centered applications in recent patents. Frequent appearance of terms like “training”, “neural”, “text”, and “language” further points to the rapid diffusion of machine learning and natural language processing technologies during this time. Moreover, the emergence of concepts like “medical”, “query”, and “prediction” indicates that generative AI methods are increasingly being applied to critical domains such as healthcare, information retrieval, and predictive analytics. Overall, this word cloud demonstrates the expansion of generative AI toward both broader multimedia ecosystems and specialized application contexts in recent years. Word cloud of the top 100 most frequent words after 2021. Comparison between networks in different time intervals Comparison between networks in different time intervals Table 6 represents a summary of the network’s results in different periods. In the pre-2016 period, the conceptual network contained 11,283 nodes and 170,395 edges. The modularity index is 0.4768, which indicates a relatively high quality of cluster separation and the existence of clear boundaries between different conceptual domains. The average network degree is 30.21, and the Jaccard similarity is 0.183, indicating that the overlap between terms is limited and each cluster focuses more on its own concepts. The network density (0.00223) also emphasizes that the connections are still scattered and unconcentrated. The dominant words in this period, such as data, user, content, image, and device, reflect the primary focus on data, user, and basic processing tools. Table 6Network statistics of generative AI patent clusters across three time intervals, showing growth in network size and decreasing modularity over time.Time intervalNumber of nodesNumber of edgesModularityAverage degreeJaccard similarityDensityPre 201611,283170,3950.476830.210.1830.002232016–202021,741348,2270.393932.030.14550.001472021–202580,7601,048,5750.372815.90.13710.00037 Network statistics of generative AI patent clusters across three time intervals, showing growth in network size and decreasing modularity over time. In the period 2016–2020, the network size has almost doubled to 21,741 nodes and 348,227 edges. The average degree increased slightly and reached 32.03, which indicates the growth of connections between concepts. However, the modularity has decreased compared to the previous period and reached 0.3939, which indicates more overlap of the clusters and a decrease in their resolution. Jaccard’s index (0.1455) and density (0.00147) also confirm that the network has become more complex and, at the same time, more interconnected. In this period, words such as model, network, and set have appeared alongside the words of the previous period, indicating the entry of concepts related to machine learning and neural networks into the generative intelligence patent ecosystem. Finally, between 2021 and 2025, the network experienced significant growth, reaching 80,760 nodes and 1,048,575 edges. This fourfold increase in the number of nodes and threefold increase in the number of edges indicate a serious leap in the diversity of concepts and the degree of interaction between them. However, the modularity index has decreased again, reaching 0.3728, indicating much greater overlap of clusters and gradual convergence of concepts. The average degree has decreased to 15.9, reflecting the greater dispersion of connections and the expansion of conceptual domains. The decrease in the Jaccard index to 0.1371 and the drop in density to 0.00037 also indicate that, despite the increase in the volume of concepts, connections are widely spread and the network has acquired a more open and dispersed structure. In this period, keywords such as image, data, user, video, and model indicate the importance of multimedia, big data, user-centricity, and deep learning models. Overall, the comparison of the three periods shows that the conceptual network of patents related to generative AI has transformed from a relatively concentrated and clustered structure in the pre-2016 period to a very extensive, diverse, and at the same time intertwined network in recent years. This trend simultaneously confirms both the expansion of the scope of concepts and indicates the gradual convergence of different subfields into a single and integrated ecosystem. Analysis of results based on the number of top keywords Analysis of results based on the number of top keywords To evaluate the sensitivity of the results to the number of selected keywords in each document, the Kruskal–Wallis statistical test was performed on the network indicators (modularity, average degree, density, and Jaccard similarity) in three values of N = 10, N = 15 and N = 25. The results showed that there was no statistically significant difference between these three cases in any of the periods studied (p > 0.9). Specifically, the p-value for the period before 2016 and the period 2016–2020 was 1, and for the period 2021–2025 it was 0.9392. These findings indicate that changing the number of selected keywords does not have a significant effect on the structure of clusters and network indices, and the results are sufficiently stable and consistent. Accordingly, the value of N = 15 was chosen as the final option and the basis for the analyses, because in addition to providing a balance between lexical coverage and preventing excessive network complexity, it is also statistically no different from the other options. As shown in Table 7, the results of the sensitivity analysis indicate that changing the number of selected keywords (N = 10, 15, 25) does not fundamentally alter the structural properties of the networks, confirming the robustness of the chosen value N = 15 for the main analysis. Table 7Sensitivity analysis of network indicators across different values of top-N keywords (10, 15, 25) for each time interval, showing that structural properties remain stable despite variations in N.Time intervalNumber of keywords(N)Number of nodesNumber of edgesModularityAverage degreeJaccard similarityDensityPre 2016108,07472,6550.54517.990.18870.0022Pre 20161511,283170,3950.476830.210.1830.00223Pre 20162513,825484,5240.418470.10.15860.00512016–20201014,969148,6150.449118.860.1470.001332016–20201521,741348,2270.393932.030.14550.001472016–20202526,974988,0860.345873.260.12710.00272021–20251053,915552,5520.305131.780.13710.000362021–20251580,7601,048,5750.372815.90.13710.000372021–202525104,3673,625,4670.393869.480.1930.00024 Sensitivity analysis of network indicators across different values of top-N keywords (10, 15, 25) for each time interval, showing that structural properties remain stable despite variations in N. Analysis of patent clusters To provide a clear overview of the conceptual areas identified, all clusters are presented in the appendix with detailed descriptions. However, for ease of review and quicker understanding of the results, the Table 8 summarizes the cluster titles along with their concise descriptions in the main text. This summary table facilitates direct comparison of the key areas and illustrates how the focus of patents has evolved, shifting from early topics toward more advanced and multifaceted systems. Table 8A summary of clusters with a description of each cluster.ClusterPeriodTitleDescription0Pre-2016Web-Based Digital Media and UIWeb and mobile-based platforms for digital content delivery and user interaction1Pre-2016Advanced Video ProcessingVideo capture, analysis, and delivery for surveillance and entertainment2Pre-2016Sensor-Based Data AnalyticsIoT-driven data collection and analytics for health and industrial applications3Pre-2016Wearable and Location-Based TechWearable devices with location-aware services for health and entertainment4Pre-2016Digital Commerce and Supply ChainsAI-driven e-commerce, payments, and supply chain systems5Pre-2016Personalized HealthcareIoT-based continuous health monitoring and risk assessment6Pre-2016Intelligent Vehicle SystemsConnected vehicles with AR displays and adaptive interfaces7Pre-2016Media and Telecom InfrastructureIntegration of broadcasting and internet-based content delivery8Pre-2016NLP and Knowledge ManagementText and multimedia semantic analysis for search and chatbots9Pre-2016Virtual Environments and CommunicationVirtualization and intelligent routing for enterprise communications10Pre-2016Computer Vision and Imaging2D/3D visual processing for autonomous systems and diagnostics11Pre-2016Speech Recognition and InterfacesVoice-driven search, commands, and cloud-based speech services02016–2020Deep Learning and Generative AI for VisionGANs and deep learning for image generation and medical analysis12016–2020Digital Content and Mobile MediaPersonalized streaming and mobile-first content delivery22016–2020Industrial IoT and AnalyticsContinuous monitoring using sensor fusion in industry and health32016–2020Gaming and Real-Time MediaInteractive gaming, VR, and live streaming technologies42016–2020Audio Processing and Signal AnalyticsAI-driven 3D audio, noise cancellation, and biomedical sound analysis52016–2020Mixed Reality DisplaysAR/VR/MR display systems with multimodal interaction62016–2020Surgical RoboticsAutomated and AI-assisted robotic surgery systems72016–2020Adaptive Manufacturing and Medical DevicesPersonalized device manufacturing with adaptive control82016–2020Autonomous Vehicle SystemsSelf-driving technologies with cognitive state monitoring92016–2020Edge Computing and WearablesDistributed processors for wearable and low-power devices02021–2025Generative AI and Medical ImagingDiffusion models for advanced image synthesis and diagnostics12021–2025Enterprise Data and Predictive AnalyticsReal-time heterogeneous data processing for Industry 4.022021–2025Extended Reality PlatformsXR and metaverse applications with adaptive content32021–2025Audio and Wireless CommunicationAdvanced wireless audio and multimodal communication42021–2025Large Language ModelsTransformer-based LLMs with knowledge graph integration52021–2025Video Analytics and StreamingReal-time video understanding and adaptive streaming62021–2025Robotic Surgery and Medical DevicesAdvanced robotic platforms and smart surgical tools72021–2025Digital Twins and Precision MedicineVirtual patient replicas for simulation and treatment optimization82021–2025Edge AI and Distributed NNEfficient inference at the network edge for real-time AI92021–2025Federated Learning and Privacy AIPrivacy-preserving federated and hybrid quantum–classical AI102021–2025Autonomous EV and Smart GridsIntegration of autonomous vehicles with electric grid systems112021–2025Advanced Video CompressionNext-gen codecs (e.g., H.266, AV1) for 8K and VR streaming A summary of clusters with a description of each cluster. An examination of the clusters shows that some of the identified conceptual trends are directly related to the ethical and societal implications of generative AI. In particular, Cluster 0 and Cluster 4 in the 2021–2025 period, as well as Cluster 0 in the 2016–2020 period, include patents addressing misinformation and methods to detect and mitigate fake content. These findings highlight the growing sensitivity to issues of trust and credibility in AI-generated outputs. Moreover, Cluster 9 in the 2021–2025 period is closely related to privacy, encompassing patents on federated learning and data-protection mechanisms. The presence of such clusters demonstrates that societal concerns such as trust, transparency, and data security are becoming integral elements in the technological development path of generative AI at the patent level. Evaluation across baseline and comparative methods The K-means model The K-means model The output of the K-Means model in three periods showed very low values of the Silhouette index, which indicates poor clustering quality. In the period before 2016, the best k was 20 with Silhouette = 0.063. In the period of 2016–2020, the optimal value of k equal to 20 and Silhouette = 0.054 was obtained. Finally, between 2021 and 2025, the best k is 24, and Silhouette = 0.050 was calculated. These numbers indicate that K-Means had a limited ability to separate semantic clusters in this data(Table 9). Table 9The results of the K-Means model.PeriodBest KSilhouetteCHDBPre 2016200.0632263.572016–2020200.0643983.852021–2025240.055704.13 The results of the K-Means model. The LDA model The LDA model The LDA model achieved the best number of clusters in the period before 2016, with a coherence value of (0.496). In the period 2016–2020, the optimal number of clusters was 8, with a coherence value of 0.485. Finally, in the period 2021–2025, the number of clusters increased to 20, and its coherence value reached 0.512. Overall, LDA has had a relatively stable performance with average coherence values of around 0.48 to 0.51(Table 10). Table 10The results of the LDA model.PeriodBest KCoherence c_vPre 2016160.4962016–202080.4852021–2025200.512 The results of the LDA model. The NMF model The NMF model The NMF model has shown higher values of convergence than LDA in all periods. In the period before 2016, the best number of clusters was 8, and the coherence value was 0.729. In the period 2016–2020, the number of clusters increased to 16, and the coherence value was calculated to be 0.676. In the period 2021–2025, the best value was obtained for 20 clusters, and its convergence was 0.672. These results indicate that NMF has the highest thematic coherence among the compared methods (Table 11). Table 11The results of the NMF model.PeriodBest KCoherence c_vPre 201680.7292016–2020160.6762021–2025200.672 The results of the NMF model. The BERTopic model The BERTopic model The BERTopic model extracted 60 topics in the period before 2016 and had a coherence value (0.625) and relatively high topic diversity (0.818), but a low Silhouette index (0.121), indicating poor clustering. In the period 2016–2020, only 2 topics were identified, which, although they had complete topic diversity (1.0), yielded a relatively low coherence value (0.589) and a very poor Silhouette (0.025). In the period 2021–2025, the number of topics increased significantly to 184, with high convergence (0.721) but still a very low Silhouette (0.0103). These results indicate that BERTopic performed well in terms of conceptual coherence but was weak in the quality of clustering resolution(Table 12). Table 12The results of the BERTopic model.PeriodNumber of main clustersCoherence c_vTopic diversityPre 2016220.62480.81832016–202030.589312021–2025100.72140.788 The results of the BERTopic model. Comparison with baseline method Comparison with baseline method The baseline method is based on extracting keywords with TF–IDF and building a word co-occurrence network, then applying the Louvain algorithm to detect clusters. Unlike models based on numerical clustering, this method was able to provide conceptual clusters with higher resolution and greater interpretability. The modularity index in this method was reported as 0.4768 (before 2016), 0.3939 (2016–2020), and 0.3728 (2021–2025) in three periods, respectively, which shows the quality of cluster separation. Also, sensitivity analysis by changing the number of top keywords showed that the cluster structure remained stable. Therefore, while NMF and BERTopic provided stronger numerical performance in certain respects, the baseline TF–IDF + network method ultimately offered more interpretable and stable conceptual clustering, making it the most suitable framework for addressing the research questions. Discussion In response to the first research question of identifying the dominant technological domains in generative AI patents, the semantic network results show that these domains have undergone significant evolution over time. Before 2016, the main focus was on concepts such as data, user, content, image, and device, which indicate a focus on data, user, and basic processing tools. In the 2016–2020 period, with the emergence of terms such as model, network, and set, the expansion of the fields of machine learning and neural networks is clearly visible. Finally, in the 2021–2025 interval, dominant terms such as image, data, user, video, and model have established their position, indicating the dominance of technologies related to image processing, big data, deep models, and interactive systems. In general, it can be said that the dominant trend has moved from basic and data-driven structures to more complex and multi-state generative systems; A transformation that has been reflected both in the significant increase in network volume (more than threefold growth in nodes) and in the gradual decrease in modularity and cluster convergence. Key trends of generative AI Transition from basic neural networks to advanced generative systems: Prior to 2016, patents focused mainly on early neural architectures and biological processes simulation. However, after 2016, a new generation of generative systems, such as GPT and Stable Diffusion, emerged that are capable of generating independent and multimodal data. This evolution marks a paradigm shift from mere data processing toward the intelligent generation of digital content. Transition from basic neural networks to advanced generative systems: Prior to 2016, patents focused mainly on early neural architectures and biological processes simulation. However, after 2016, a new generation of generative systems, such as GPT and Stable Diffusion, emerged that are capable of generating independent and multimodal data. This evolution marks a paradigm shift from mere data processing toward the intelligent generation of digital content. Moving from discrete components to integrated systems: In the early days, innovations were mostly focused on discrete components such as classification algorithms or sensing tools. However, in recent years, patents describe systems that provide integrated, end-to-end intelligent services. This integration has helped accelerate development and commercialization, thereby enhancing user experience and application flexibility. Moving from discrete components to integrated systems: In the early days, innovations were mostly focused on discrete components such as classification algorithms or sensing tools. However, in recent years, patents describe systems that provide integrated, end-to-end intelligent services. This integration has helped accelerate development and commercialization, thereby enhancing user experience and application flexibility. Increased focus on content authenticity: With the proliferation of models for text, image, and video generation, the issue of data authenticity and mitigating synthetic or fake content has gained particular importance. After 2016, a significant portion of patents have focused on developing systems to detect artificial media (such as deepfakes). This trend underscores rising concerns regarding data security, digital fraud, and the spread of misinformation. Increased focus on content authenticity: With the proliferation of models for text, image, and video generation, the issue of data authenticity and mitigating synthetic or fake content has gained particular importance. After 2016, a significant portion of patents have focused on developing systems to detect artificial media (such as deepfakes). This trend underscores rising concerns regarding data security, digital fraud, and the spread of misinformation. Shift from academic to commercial focus: Before 2016, many patents focused more on theoretical concepts and abstract algorithms. Since then, approaches have shifted toward practical applications, user experience, and industrial use. This shift reflects the maturity of the AI market and its move toward commercial deployment and economic exploitation. Shift from academic to commercial focus: Before 2016, many patents focused more on theoretical concepts and abstract algorithms. Since then, approaches have shifted toward practical applications, user experience, and industrial use. This shift reflects the maturity of the AI market and its move toward commercial deployment and economic exploitation. Implications Patent network analysis and conceptual clustering offer strategic insights into research gaps, investment avenues, and targeted technological policymaking. This study has led us to understand the conceptual evolution of generative AI technologies by comparing the pre- and post-2016 patent clusters. This comparison highlights the transition from early-stage neural and electromechanical clusters toward advanced generative and multimodal systems. Before 2016, clusters were centered on areas such as Web-Based Media and User Interfaces, Wearable and Sensor-Based Systems, and Electromechanical or Neuromorphic Architectures. In contrast, the post-2016 period, particularly after 2021, shows the dominance of clusters like Generative AI and Diffusion Models, Large Language Models and Knowledge-Enhanced AI, and Personalized Interactive XR Systems, reflecting the move from isolated component-level innovations to integrated, user-adaptive frameworks with direct societal and industrial impact. Unlike many previous studies in the field of patent-based technology roadmapping, which have generally focused on simple document counting or using traditional clustering algorithms, this study was able to identify the conceptual structure of patents with higher accuracy by utilizing a network approach and calculating indicators such as modularity, density, and Jaccard similarity. Also, performing sensitivity analysis for the selection of the number of keywords, and comparing several clustering models (LDA, NMF, KMeans, and BERTopic) demonstrates the innovation of this study compared to similar studies. Beyond foresight, the results support guiding investment strategies for startups and VCs to invest more efficiently. The study demonstrates that areas with dense patent activity often correlate with market maturity or growth potential. So, network visualizations and modularity analysis show concentrated clusters in personalized AI and automotive processing systems post-2016. Therefore, these can be interpreted as mature or competitive innovation spaces worth investment. Along with investors, this study might also benefit regulators and research scientists. Regarding regulators, the study provides tools to identify high-impact innovations. High-impact technologies often appear in dense clusters with many connections. In this study, co-occurrence networks make it possible to flag technologies like autonomous control and synthetic media, helping regulatory bodies monitor them. For example, the results of this research could be used by regulators to track and monitor sensitive technologies such as deepfakes by identifying conceptual clusters related to image processing and multimedia content production, enabling policymakers to formulate more precise preventive and regulatory rules. Besides, using the results section, R&D scientists can prioritize the efforts in their industrial labs. Labs must focus on problems with growing patent and competitive density, and the identified clusters help R&D teams align with the global innovation trajectory in AI. At the industrial level, large R&D companies could use the extracted clusters to prioritize investments in technologies such as large language models, interactive systems, or multimedia applications. This approach would align R&D strategy with global patenting trends. Taken together, these implications demonstrate how network-based patent analysis serves as a multifaceted tool for navigating the rapidly evolving landscape of generative AI. The results of network analysis and conceptual clustering, in addition to providing a scientific picture of technology trends, also have direct practical applications: for example, generative artificial intelligence companies will be able to align their research and development path with market priorities by identifying keywords and focused areas in each time frame. Limitations and future directions Despite the breadth and depth of this study’s analysis of generative AI technology through large-scale patent network modeling, it is important to acknowledge certain limitations that shape the interpretation of the results and point toward future research opportunities. A central limitation of the current work is its reliance exclusively on English-language patent documents. While English serves as the predominant language in the global technological landscape, this focus inevitably introduces a linguistic and geographical bias. Crucial insights from innovation hubs where patenting occurs primarily in other languages (such as China, Japan, or Germany) may be underrepresented. Given that countries like China are rapidly becoming leaders in generative AI research and commercialization, this linguistic constraint may obscure important regional trends and emergent clusters unique to non-English-speaking contexts. Future studies should consider integrating multilingual patent datasets, leveraging advances in machine translation and multilingual language models such as BERT or XLM-RoBERTa, to create a more globally inclusive and representative map of the generative AI innovation landscape. This would enable the identification of innovation hotspots, patterns of knowledge transfer, and region-specific priorities that are otherwise hidden in monolingual analyses. Another limitation relates to the study’s focus on the technological evolution of generative AI in isolation, without systematically comparing it to developments in adjacent domains such as robotics, healthcare, cybersecurity, or autonomous systems. The contemporary innovation ecosystem is increasingly characterized by cross-domain convergence and mutual influence, where advances in generative AI both draw from and contribute to neighboring fields. A comparative, multi-domain patent analysis would reveal the extent to which the trajectories observed in generative AI are unique or part of broader technological paradigms, and would clarify the dynamics of technological spillover, co-evolution, and interdisciplinary innovation. Such analysis could also support policymakers and investors in identifying which generative AI subfields have the greatest potential for impact beyond their primary domain. From a methodological standpoint, the study’s approach (combining TF–IDF-based keyword extraction, co-occurrence network modeling, and the Louvain community detection algorithm) offers a robust framework for capturing conceptual evolution. However, this method, like all clustering and topic modeling approaches, has its limitations. The reliance on TF–IDF may underrepresent the semantic richness and contextual nuance present in patent texts, especially as the field grows more complex. Although manual cluster labeling ensures interpretability, it introduces subjectivity and may hamper reproducibility. Future work could benefit from deploying advanced semantic embedding techniques (such as transformer-based embeddings, contextualized topic modeling, or graph neural networks) to automate the identification and labeling of clusters with greater depth and objectivity. Additionally, integrating other elements of patent documents (such as claims, citations, and technical drawings) would provide a more holistic understanding of technological novelty, impact, and the interconnectedness of innovation. The temporal granularity of the analysis, while suitable for capturing long-term shifts and major inflection points, may mask more subtle or rapidly emerging trends, particularly following disruptive events such as the release of ChatGPT or breakthroughs in diffusion models. Finer-grained, perhaps annual or even quarterly, temporal analysis could reveal the micro-dynamics of concept emergence, cluster evolution, and technological convergence. This would be especially valuable for real-time technology forecasting and for stakeholders seeking to respond rapidly to paradigm shifts in the field. Furthermore, while this study identifies the emergence of clusters related to societal and ethical issues (such as misinformation, trust, privacy, and regulatory compliance), it does not directly correlate patent evolution with external datasets, including investment flows, policy interventions, or startup activity. Future research could overlay patent network dynamics with data on venture capital investment, government regulation, or market adoption to better understand how technological, economic, and regulatory factors interact to drive or constrain innovation. Such integrative approaches would support more nuanced, actionable foresight for investors, policymakers, and industry leaders. Finally, the findings underscore the growing convergence of technological, commercial, and ethical considerations in generative AI innovation. As patents increasingly address challenges such as content authenticity, privacy-preserving learning, and regulatory compliance, the boundaries between technical progress and societal impact are becoming blurred. Future directions should thus include not only methodological and data enhancements, but also cross-sectoral collaborations that bring together technologists, ethicists, legal scholars, and industry practitioners. Such collaborations are essential to ensure that the rapid evolution of generative AI is aligned with societal values, regulatory expectations, and the broader public interest. In summary, while this study advances the understanding of generative AI’s technological evolution, addressing its current limitations (through multilingual analysis, cross-domain comparison, finer temporal resolution, advanced semantic methods, and integration with external datasets) will provide richer, more actionable insights. These efforts will support a more inclusive, responsive, and responsible roadmap for navigating the fast-changing landscape of generative AI. Conclusion This study provides a systematic and data-driven exploration of the technological evolution and innovation landscape of Generative Artificial Intelligence (GAI) through large-scale patent analysis. Methodologically, it contributes to the field by combining large-scale patent mining with scientometric visualization and community detection, a novel integration not previously applied to GAI. By leveraging advanced text mining, network modeling, and community detection techniques on over 20,000 patents, we have mapped the conceptual shifts, emerging domains, and innovation trajectories that have shaped GAI over the past decades. Our findings reveal a clear inflection point around 2016, marking a transition from early-stage, modular, and domain-specific innovations (such as neuromorphic computing, agricultural bioengineering, and electromechanical control systems) to a new era characterized by integrated generative frameworks, API-driven platforms, and multi-modal capabilities. This shift is quantitatively evidenced by a dramatic increase in patent volume, a threefold rise in conceptual diversity, and a significant decrease in network modularity, indicating the convergence and cross-pollination of previously distinct technological subdomains. The post-2016 landscape is dominated by clusters focused on generative AI frameworks, asynchronous processing for autonomous systems, personalized interactive experiences, media authentication, and advanced agricultural technologies. Notably, the growing emphasis on content authenticity and user personalization reflects both the technological maturation of GAI and the rising societal and ethical concerns associated with synthetic media and AI-driven decision-making. From a strategic perspective, our patent network analysis offers actionable insights for R&D managers, policymakers, and investors. The identification of dense innovation clusters and emerging hubs can inform targeted investment, guide regulatory oversight, and support the alignment of research agendas with global technological trends. Furthermore, the methodological framework presented here (combining TF–IDF-based keyword extraction, co-occurrence network analysis, and community detection) demonstrates the value of patent analytics as a foresight tool for navigating complex and rapidly evolving technology landscapes. However, this study is not without limitations. The reliance on English-language patents and TF–IDF-based keyword extraction may obscure semantic nuances and overlook innovations documented in other languages. The temporal granularity, limited to two broad periods, may mask more subtle or recent shifts, such as the post-2022 explosion of large language models. Manual cluster labeling introduces a degree of subjectivity, and future research should consider embedding-based semantic analysis and multilingual datasets for deeper and more nuanced insights. In addition to providing descriptive analysis, the suggested framework offers predictive capabilities that allow policymakers, business executives, and researchers to forecast future developments in generative AI and adjust their strategic plans to match these anticipated trends. In summary, this research advances our understanding of the dynamic and interconnected ecosystem of generative AI technologies. It highlights the field’s rapid expansion, increasing complexity, and the convergence of innovation domains. By providing a structured roadmap of technological evolution, this work supports informed decision-making and responsible innovation, ensuring that the transformative potential of generative AI is harnessed for broad societal benefit while addressing emerging risks and challenges. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Author contributions 1.Navid Mohammadi, Writing Final version, Data Gathering, Supervision, Methodology 2.Jalil Heidary Dahooie: Methodology, Validation, Visualization 3.Amir Ali Bengari, Writing Draft version, Data Gathering, Methodology 4.Arash Rahimi: Writing Draft, Data Gathering, Formal Analysis. Funding The authors did not receive support from any organization for the submitted work. Data availability The data used in this paper is available as requested. Declarations Competing interests The authors declare no competing interests. Ethical approval and informed consent All of the ethical issues have been considered in this research. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The data used in this paper is available as requested.