A study on the dynamic governance mechanism of digital publishing policies driven by generative AI technology—based on an analytical framework of technological-institutional co-evolution

Abstract

Introduction:

The rapid evolution of generative artificial intelligence (GAI) is disrupting the digital publishing sector, creating governance challenges such as ambiguous copyright ownership and unclear platform liability. Existing research often interprets the technology-institution relationship through a unidirectional causal lens, lacking empirical analysis of their interactive mechanisms. This study aims to analyze the co-evolutionary dynamics between GAI and institutional responses to understand how policy systems adapt to technological change.

Methods:

This study employs a technological-institutional co-evolutionary framework using a mixed-methods approach. The methodology integrates natural language processing (NLP) topic modeling, judicial case coding, and a breakpoint test. The analysis compares 48 policy documents and 14 judicial cases from China, Europe, and the United States, spanning the period from 2016 to 2025.

Results:

The findings reveal that GAI has driven a structural shift in policy agendas toward AI governance and copyright issues. Comparative analysis shows divergent evolutionary trajectories: China exhibited administration-led catching-up characteristics with a policy lag of approximately 12 months, whereas Europe and the United States demonstrated collaborative adaptation patterns with a longer lag of approximately 24 months. Legal conflicts were predominantly concentrated in the attribution of copyright for AI-generated content (40.63% of cases) and platform liability (35.94%).

Discussion:

This study reveals the non-linear structural breaks and divergent evolutionary trajectories of institutional responses to GAI. By providing empirical evidence of how different governance systems navigate the balance between technological change and institutional inertia, the findings contribute to the development of adaptive AI governance strategies.

1 Introduction

In recent years, generative artificial intelligence has developed rapidly, profoundly impacting the digital publishing industry (Ryzhko et al., 2024). The mismatch between rapid technological iteration and policy and regulatory adjustments has created governance challenges, such as copyright attribution and platform liability, leading to multiple dilemmas, including market disorder, stifled innovation, and heightened ethical risks. Analyzing the co-evolutionary path of technology and policy and exploring agile, adaptive governance mechanisms have become urgent priorities for both theoretical research and policy practice.

Existing research often interprets the relationship between technology and publishing from a unidirectional causal perspective, overlooking their bidirectional interaction. Technological determinism emphasizes the disruptive effects of artificial intelligence on traditional publishing processes, arguing that automated content creation and intelligent recommendations are reshaping the publishing industry chain (Guo, 2018; Chang, 2020). For example, if AI can efficiently generate text, images, and even complex cross-modal content, it can significantly improve publishing efficiency and potentially reduce labor costs (Chauhan, 2024). However, such analyses often fail to fully account for constraints and the shaping effects of the institutional environment on technology adoption, implementation pathways, and their socioeconomic consequences. Institutionalist research, on the other hand, focuses on a normative examination of existing policies and regulations, pointing out that the current governance framework lags significantly behind in addressing issues such as copyright definition, information authenticity, and ethical risks (Samuelson, 2023; Deng et al., 2025; Zhang et al., 2025).

However, while both types of research offer valuable insights, neither has succeeded in revealing the complex mechanisms of interaction and co-evolution between technological innovation and the institutional environment. This gap in understanding has created a fundamental puzzle in observing the convergence of artificial intelligence and digital publishing: while some applications integrate seamlessly into existing workflows, others, such as the establishment of authorship rights, continue to spark intense ethical and legal controversies. To address these challenges, this study focuses on the following core questions: How does the development of GAI technology drive the evolution of digital publishing policy agendas in China, Europe, and the United States? What temporal differences emerge in policy priorities? What core legal and ethical conflicts does this trigger within the digital publishing sector? How do different governance models influence the speed and efficiency of technological-institutional co-evolution? This study applies a technological-institutional co-evolution framework to this rapidly iterating technological field. It identifies the evolutionary trajectories of policy foci through natural language processing and topic modeling, reveals core conflicts and stakeholder bargaining patterns through qualitative coding of judicial cases, and compares the lag periods and evolutionary stages of policy responses by combining technological development data with breakpoint tests. The article is structured as follows: Section 2 presents a literature review; Section 3 outlines the research design; Section 4 presents the empirical results; Section 5 discusses the theoretical implications; and Section 6 presents the conclusions and implications.

2 Literature review

2.1 The technical characteristics of generative AI and its impact on digital publishing

As a key advancement in deep learning, generative AI (GAI) is driving profound changes in the digital publishing industry through its technical capabilities. The core technological foundations of GAI, particularly the Transformer architecture (Vaswani et al., 2023) and large-scale foundational models trained on massive amounts of data (Radford et al., 2019; Bommasani et al., 2022), through self-supervised learning, have acquired powerful cross-modal content generation capabilities and are regarded as “creative automation.” This enables AI not only to perform repetitive tasks but also to generate novel, complex content, thereby directly intervening in and reshaping the digital publishing value chain. First, in the content production phase, GAI tools can automate or assist with tasks such as writing, editing, image generation, and translation, significantly improving publishing efficiency and reducing production costs (Bommasani et al., 2022). Tools like MidJourney have already been used for illustration design in digital publishing, significantly shortening the creative cycle. Second, in the content distribution phase, drawing on successful practices from other media industries (Gomez-Uribe and Hunt, 2015), AI-driven recommendation algorithms enable more precise personalized content delivery, optimizing user experience and enhancing platform user retention. For instance, Amazon’s Kindle platform has increased e-book sales through AI recommendation algorithms (Amazon, 2025). Finally, GAI’s multimodal generation capabilities have also driven innovation in publishing formats, giving rise to new forms such as interactive e-books and dynamic data visualizations, thereby enriching the digital publishing content ecosystem (Zhu et al., 2018). However, these early studies have largely focused on the direct efficiency gains of the technology, and there remains a lack of systematic evaluation of the quality and originality of the generated content, as well as its deeper impacts on cultural diversity.

2.2 Current state of research on digital publishing policies

Existing research on digital publishing policy primarily focuses on several key issues, and the rise of GAI has introduced new dimensions and challenges to these issues: the first concerns copyright protection and ownership. The limitations of traditional digital rights management (DRM) have become even more apparent in the era of GAI. Key challenges include determining the boundaries of fair use for copyrighted materials used in GAI training data, the copyrightability of AI-generated content, and the definition of its ownership (Lu Han and Mohong Liu, 2024). Most existing copyright laws are based on the “human authorship” principle and are difficult to apply directly to AI-generated works (Gaffar and Albarashdi, 2024). However, some studies have explored potential solutions (Rasenberger, 2024); most analyses, however, remain at the level of legal principles, lacking detailed research and empirical testing of copyright allocation schemes across different technological approaches. Second, regarding content safety and information governance. The issues of information echo chambers and bias caused by algorithmic recommendations (Du, 2024) are exacerbated by GAI’s deepfaking capabilities, posing a greater threat to information authenticity and social trust (O'Sullivan and Passantino, 2023). At the same time, the black-box nature and potential biases of GAI models pose significant challenges for content attribution (Kocak, 2024), liability determination, and algorithmic auditing. While existing research largely calls for strengthened regulation, there is a lack of concrete, feasible proposals and cross-national comparative analyses for designing dynamic content governance mechanisms that balance efficiency, fairness, and transparency—particularly in the context of GAI. Third, the level of industrial competition and market structure. The issue of platform monopolies in the digital publishing sector has long been a concern; large platforms have been accused of leveraging their dominant market positions to squeeze out small and medium-sized publishers, thereby undermining cultural diversity (Cramer, 2017). The development of GAI further exacerbates market concentration, as the development and deployment of advanced GAI models require massive computational power and data resources. Large tech companies hold a significant advantage, potentially creating a “winner-takes-all” scenario that puts even greater pressure on the survival of small- and medium-sized publishing organizations (Bender et al., 2021). Although some regions (such as the European Union) have attempted to regulate platform power through measures like the Digital Services Act (DSA) (Veale and Zuiderveen Borgesius, 2021), existing policy analyses pay little attention to the specific mechanisms through which GAI affects the concentration of the digital publishing industry, or how adaptive policy tools can be used to maintain fair competition and an innovative ecosystem. In particular, there is a lack of research systematically comparing the EU and the United States as a whole with China, making it difficult to reveal the divergent pathways of different governance traditions in the technological-institutional co-evolution.

2.3 Core concepts of the theory of technological-institutional co-evolution

The theory of technological-institutional co-evolution offers an integrated perspective for analyzing the governance challenges posed by generative artificial intelligence; at its core, it emphasizes the dynamic relationship between technological innovation and the institutional environment, in which the two mutually shape and evolve in tandem (Nelson and Winter, 1983; Geels, 2002). This theoretical framework was established by Nelson and Winter in the field of evolutionary economics, which posits that there is a continuous process of interaction and selection between technological change and organizational structure (Nelson and Winter, 1973); North expanded upon this from the perspective of institutional change, emphasizing that institutions shape the trajectory of economic evolution by reducing uncertainty and providing incentives (North, 1971); Giles, meanwhile, applied this framework through a multi-level perspective to the study of the interaction between technological innovation and social institutions, revealing how Niche innovations disrupt and reshape socio-technical systems. Technological breakthroughs often challenge existing laws and norms, forcing institutional adjustments; conversely, the institutional environment guides the direction of technological R&D and social applications through regulation and incentives. This bidirectional causal feedback mechanism constitutes the core of symbiotic evolution. Building on this, North further introduced the concept of adaptive efficiency to measure an institutional system’s ability to respond to uncertainty and promote sustainable development, emphasizing that institutions should possess the resilience to learn from practice and continuously optimize themselves. Previous research has largely focused on long-term macro-level transformations, such as the energy transition, or has addressed general AI governance; few studies have delved into the specific interaction mechanisms between GAI and digital publishing policies, and there is a particular lack of empirical comparisons of differences in adaptive efficiency across different institutional environments (Saikaly, 2023). Existing theoretical frameworks are largely based on long-term technological change and often implicitly assume gradual, linear evolution, making it difficult to fully account for the rapid iteration and high uncertainty characteristic of generative artificial intelligence. To overcome these theoretical limitations, this study aims to develop a practical, empirical analytical framework grounded in symbiotic evolution theory. By incorporating breakpoint tests, lag period calculations, and a typology of differentiated pathways, we translate theoretical concepts into measurable indicators, thereby providing a methodological foundation for comparing adaptive efficiency across different institutional environments (see Table 1 for related studies).

3 Data sample and methodology

3.1 Research design and analytical framework

This paper employs a mixed-methods approach, primarily qualitative, with quantitative analysis as a supplement, to reveal the symbiotic evolution of generative artificial intelligence and digital publishing policies. Based on the theory of technological-institutional co-evolution outlined in the literature review, we have transformed its core tenets into an operational, empirical, analytical framework, constructing a model comprising three core dimensions (see Figure 1). The technological dimension treats GAI as an external shock source triggering institutional responses, using indicators such as the Gartner Hype Cycle to characterize the intensity and pace of technological shocks. The institutional dimension focuses on the intensity and focus of the policy system’s response, measuring the allocation of institutional attention by tracking changes in the thematic weight of different issues across policy texts. The interaction dimension aims to reveal the speed of feedback and interaction patterns between technology and institutions. It introduces the indicator of policy lag period, quantifying feedback speed by comparing the time difference between technological inflection points and policy inflection points, while identifying whether interaction patterns are dominated by collaborative adjustment or conflictual competition through the distribution of core conflict types in judicial cases (Nelson and Winter, 1983; North, 1990).

Figure 1

Based on this framework, the study proposes three criteria for policy comparison: response speed measures the agility of governance systems based on the length of the lag period; conflict focal points compare the distribution of conflict types, such as determination of ownership and responsibility allocation in judicial cases to identify core institutional bottlenecks across different jurisdictions; and evolutionary pathways compare the classification of evolutionary stages and preferences for policy instruments across different regions based on the results of breakpoint tests.

3.2 Data sources and sample selection

To populate the analytical framework’s various dimensions, the system collects three types of data: policy, judicial, and technical. Policy texts are used to measure the intensity and focus of institutional responses; judicial cases are used to analyze interaction patterns and points of conflict; and technical reports are used to characterize the trajectory and maturity of technological impacts.

3.2.1 Policy text data

This study examines digital publishing and artificial intelligence-related policies and regulations from three representative regions: China, the European Union, and the United States. The European Union and the United States are analyzed together as “Europe and the U.S.” primarily because both follow a governance logic centered on protecting fundamental rights, mitigating risks, and regulating markets in the field of GAI governance. This stands in stark contrast to China’s administration-led agile governance model; combining the two helps highlight the fundamental differences in governance approaches. Second, the sample sizes of policies from the EU and the U.S. are relatively limited; combining them for analysis enhances the effectiveness of statistical methods such as the breakpoint test. Chinese policy samples are primarily sourced from official channels such as the Chinese government website and the Cyberspace Administration of China (CAC), including core documents such as the Cybersecurity Law, the Data Security Law, the Interim Measures for the Management of Generative Artificial Intelligence Services, their related interpretations, and Local policy experiments. Policy samples from Europe and the United States are primarily sourced from the EUR-Lex database and the official websites of European and American institutions, with a focus on the General Data Protection Regulation (GDPR), the Digital Services Act (DSA), the Digital Markets Act (DMA), and the draft, adopted texts, and related guidance documents of the Artificial Intelligence Act (AI Act) (Table 2). A detailed list of all 48 policy texts (including titles, sources, and dates) is provided in Appendix 1.

3.2.2 Judicial case data

We screened legal databases such as Beida Fabao, Westlaw, and LexisNexis, as well as officially published judicial documents, to identify key dispute cases that occurred in China, the United States, and Europe between 2019 and 2025. The selection criteria included publicly available court judgments or administrative rulings that directly addressed copyright ownership, infringement liability, or data compliance regarding AI-generated content, and that were highly influential or representative. A total of 14 judicial cases were identified; see Appendix 2 for a detailed list.

3.2.3 Data on technological development

This section cites industry reports from authoritative third-party organizations to track changes in GAI’s technological maturity, adoption rates, and risk perceptions in the digital publishing sector. Key references include Gartner’s Hype Cycle and annual reports on AI and digital economic development issued by the China Academy of Information and Communications Technology and the China Internet Network Information Center. A detailed list is provided in Appendix 3.

3.3 Analysis methods and steps

3.3.1 Policy text analysis

To identify trends in the evolution of policy priorities and their alignment with technological developments, topic modeling was conducted on policy texts. During data preprocessing, Chinese texts were segmented using the Jieba library with a custom dictionary to filter out common stopwords; English texts were stemmed using the NLTK library and filtered for stopwords. All texts were uniformly encoded, with punctuation, numerals, and low-frequency words removed. Following preprocessing, a corpus consisting of 28,647 valid words was obtained. The Latent Dirichlet Allocation (LDA) model was employed for topic modeling. To determine the optimal number of topics, the K-value was tested from 2 to 10, and topic consistency and model confusion scores were calculated for each K-value. The results showed that when K = 5, the topic consistency score peaked (Cv = 0.58) and the confusion score stabilized. The five themes were named, based on high-weight keywords, as Data and Algorithm Ethics, Digital Content and Copyright, Platform Regulation and Market, AI Governance and Security, and Technology Development and Infrastructure. After theme identification, the probability distribution of each text across the five themes was calculated, and the average weight of each theme was derived by aggregating data by year to plot a thematic intensity evolution map. Spearman’s rank correlation coefficient was used to test the correlation between policy theme intensity and Gartner Technology Maturity Scores (Gacula and Singh, 1984).

3.3.2 Deconstruction of judicial cases

The core objective of judicial case deconstruction is to reveal conflict patterns and the logic of stakeholder interactions within the interplay between technology and institutions through qualitative coding. Claims are categorized into three dimensions: ownership determination, responsibility allocation, and technical compliance. Following grounded theory, a three-tier coding process is employed: open coding to extract original points of contention, axial coding to summarize conflict types, and selective coding to construct a framework of claims among developers, publishers, and regulators (Strauss and Corbin, 1990). In 14 judicial cases, we conducted a qualitative assessment of each party’s claim intensity and the court’s tendency to allocate liability in the judgment texts. We assigned quantitative scores to the claims of different actors and constructed a Stakeholder game intensity heatmap to visually illustrate the distribution patterns of each party’s legal positions across different conflict types (Barrett et al., 2005) (see Table 3). To ensure reliability, two coders independently processed all cases, Krippendorff’s α = 0.73 (>0.70 threshold) (Hayes and Krippendorff, 2007). Consensus was reached on three disputed cases following arbitration by a third expert. The complete coding table (including original points of contention and rationale for scoring) is provided in Appendix 4.

3.3.3 Identification of stages in technological-institutional co-evolution

The theory of technological-institutional co-evolution posits that a unidirectional causal relationship does not characterize technological progress and institutional change; rather, they co-evolve synergistically through bidirectional interaction (Duan and Dong, 2025). In the field of digital publishing driven by GAI, drawing on the stage-based framework of symbiotic evolution theory and combining quantitative indicators with statistical tests allows for a more objective depiction of the interactive process. To systematically analyze this dynamic process, this paper establishes key indicators to measure technological development and institutional responses, grounded in theoretical frameworks and prior analyses (see Table 4).

These indicators were selected to characterize evolutionary stages across multiple dimensions, including technological maturity, market adoption, policy attention, institutional innovation, and interaction intensity (Geels, 2002; Teece, 2018). Threshold settings (Sun, 2005) are primarily based on literature reviews and exploratory analyses of the data in this study (Hair et al., 2018) to distinguish the typical characteristics of different stages. To objectively identify key turning points in the technology-institution interaction over time series and thereby delineate evolutionary stages, this study employs the Bai-Perron multiple structural breakpoint test. The core of this breakpoint test is to calculate the maximum value of the F-statistic (supF) to test whether the presence of m breakpoints is significantly different from the null hypothesis of zero breakpoints. Calculation of the F-statistic for a single breakpoint (see Equation 1):

Where T is the total number of observations, is the sum of squared residuals for the no-break model (null hypothesis), is the sum of squared residuals for the model, q is the number of constraints, and ρ is the number of parameters. When (the critical value), the null hypothesis is rejected, and a structural break is deemed to exist.

After identifying the breakpoints, the lag period in this paper is determined not through correlation analysis, but by comparing the years of significant structural breaks in the two key time series. We first use the Bai-Perron test described above to identify the significant structural break years for “GAI technological maturity” and “policy response intensity”和). The lag period is defined as follows (Equation 2):

Based on the identification of the aforementioned turning points and in conjunction with technological development milestones, this paper preliminarily divides the co-evolutionary process into three phases: the “technology triggering phase,” the “institutional feedback phase,” and the “collaborative adjustment phase” (see Table 4).

3.3.4 Integration of qualitative and quantitative approaches

This study employs triangulation to integrate data from multiple sources and methodologies. Specifically: (1) NLP and LDA are used to reveal how policy priorities have evolved; (2) judicial case coding is used to provide an in-depth explanation of why these core conflicts and power struggles arise; (3) the breakpoint test is used to identify when key shifts occurred and to correlate policy responses with the timeline of technological developments.

4 Results

4.1 The temporal evolution of policy focus

Topic modeling reveals a non-linear trajectory of institutional shifts as technology matures.

• Significant structural breaks in China’s policy themes occurred in 2020 and 2023 (see Figure 2). In the early stages, policies focused on digital infrastructure development and the governance of the platform economy; in 2016, the thematic weight for data and algorithmic ethics reached 0.999192, while in 2018, the thematic weight for platform regulation and the market stood at 0.999802. Starting in 2021, as GAI garnered global attention, the policy focus shifted toward AI governance and security, as well as digital content and copyright; in 2021, the weight of the digital content and copyright theme reached 0.999961. Around the time the Interim Measures for the Management of Generative Artificial Intelligence Services were issued in 2023, the thematic weight of “Data and Algorithm Ethics” rose from 0.741549 in 2020 to 0.999805 in 2023, and further increased to 0.999964 in 2024, representing a 34.7% increase.

• Policy developments in Europe and the United States have followed slightly different trajectories (see Figure 3). Between 2016 and 2019, policies centered on data and algorithmic ethics, with weights of 0.999785 and 0.999854, respectively, were closely tied to the implementation of the General Data Protection Regulation (GDPR). In 2020, the weight of the AI governance and security theme rose to 0.451940, as policies began to focus on risk management. In 2021, the weight of the platform regulation and market theme surged to 0.999861, coinciding with the legislative progress of the Digital Services Act and the Digital Markets Act. By 2024, the weight of the AI governance and security theme reached 0.999962, corresponding to the formal implementation of the AI Act. Unlike China’s reliance on ad hoc regulations, Europe and the United States place greater emphasis on systematic legislative frameworks. The Artificial Intelligence Act took nearly 3 years to progress from proposal to enactment, reflecting the emphasis on regulatory stability inherent in legislation-led governance. It should be noted that there are subtle differences between the EU and the United States: the EU prioritizes systematic legislation and the protection of fundamental rights, while the United States relies more on industry self-regulation and case law. However, both share a common core governance logic.

• Regarding the synchronization between technology and policy, Spearman’s correlation analysis shows that the weight of AI governance themes in Chinese policy exhibits a moderate positive correlation with technological maturity (ρ ≈ 0.4083). At the same time, in Europe and the United States, there is a strong positive correlation (ρ ≈ 0.5917). China’s policy responses are rapid but highly volatile; for example, the weighting plummeted to 0.000050 in 2021, failing to align with technological maturity fully. In contrast, the systematic and forward-looking nature of European and American policy frameworks ensures greater alignment with technological maturity. By 2024, when the AI Act was implemented, the weighting reached 0.999962, closely coinciding with the technology entering a trough period.

Figure 2

Figure 3

4.2 The core conflicts and key perspectives reflected in judicial precedents

An analysis of the coding of 14 judicial cases (see Table 5) shows that the three categories of dispute points—Determination of ownership, responsibility allocation, and technical compliance—accounted for 40.63, 35.94, and 23.43% of the total dispute points, respectively. By jurisdiction, ownership determination accounted for 41.94% of cases in China and 39.39% in Europe and the United States; regarding technical compliance, the proportion was 24.25% in Europe and the United States, higher than China’s 22.58%. Determination of ownership has become the most critical institutional bottleneck, while Europe and the United States place greater emphasis on technical standards. In conflicts over ownership, judicial reasoning across jurisdictions varies significantly. In China’s Tencent AI Writing Case, the court ruled that AI-generated articles possess originality and should be protected under copyright law. In the U. S. Talle Case, the federal district court denied the copyright registration request, ruling that copyright law applies only to human authors. In the UK Amber Heard deepfake case, the court held the defendant liable for infringement, emphasizing the protection of personality rights rather than the determination of the work’s attributes. Regarding conflicts over responsibility allocation, in the ByteDance AI recommendation algorithm infringement case in China, the court ruled that the platform bore partial liability for failing to fulfill its reasonable review obligations. The EU Digital Services Act imposes risk-assessment and content-moderation obligations on super-large online platforms, reflecting a trend toward stricter platform liability.

Figure 4 illustrates the claim intensity of three types of entities—developers, publishing houses, and regulators—across different dispute types. The values are based on a qualitative assessment of the intensity of each party’s claims as described in the court rulings and the courts’ tendencies in allocating liability (see Table 3 for the scoring criteria). In terms of distribution patterns, the claim intensity of developers is generally higher in cases involving determination of ownership, reflecting their core concern regarding the attribution of economic value to AI-generated content; in cases involving responsibility allocation, the claim intensity of publishers is relatively prominent, reflecting the practical dilemma they face regarding the pressure of infringement damages and compliance costs; in cases involving technical compliance, the claim intensity of regulators is higher, reflecting their focus on systemic risks such as algorithmic transparency and data security. The divergence in claim intensity across these three types of entities reveals the core dynamics of the technology-institution interaction. It provides differentiated guidance for jurisdictions seeking to establish adaptive governance mechanisms.

Figure 4

4.3 Empirical identification of the stage of technological-institutional co-evolution

By combining data on the evolution of policy themes, trends in judicial precedents, and technological developments, and applying the Bai-Perron breakpoint test and correlation analysis, we identify the stages of Sino-European technology-institution interaction.

• China’s Evolutionary Stages: The results of the Bai-Perron test indicate that the time series of China’s “AI Governance & Security” theme exhibits significant structural breaks in 2020 and 2023, clearly delineating two key stages of policy evolution (see Figure 5). The first breakpoint corresponds to the technology triggering phase, during which the policy theme weight rose from 0.000076 in 2019 to 0.258404 in 2020; the second breakpoint corresponds to the institutional feedback phase, during which the weight rose from 0.998294 in 2022 to 0.998959, stabilizing at 0.999624 in 2024. The lag period between the rapid technology diffusion phase (2021) and the significant policy response phase (2022) is 12 months.

• Evolutionary Stages in Europe and the United States: The evolutionary trajectory of GAI governance in Europe and the United States is characterized by an earlier start and sustained progress (see Figure 6). The time series of the weight of AI governance and security topics in Europe and the United States also exhibits significant structural breaks in 2020 and 2023. In 2020, the weight rose significantly to 0.451940, but fell back to 0.000050 in 2021. By 2023, the weight surged to 0.518138, reaching 0.999962 in 2024, indicating that European and American policies have entered a phase of high-level stability. This breakpoint aligns closely with the multi-year deliberation and eventual passage of the AI Act, reflecting Europe and the U. S. transitioning from an exploratory phase to a period of institutional feedback and regulatory framework development. The lag period between the rapid technology diffusion phase and the significant policy response phase is 24 months.

Figure 5

Figure 6

5 Discussion

5.1 The diversification of institutional responses driven by technology

Digital publishing policies in China, Europe, and the United States exhibit significant paradigm differences in their responses to the impact of generative artificial intelligence; the differences in lag periods revealed by empirical data (approximately 12 months in China versus approximately 24 months in Europe and the United States) essentially represent a quantitative manifestation of adaptive efficiency under different governance traditions (Geels, 2002). China’s administration-led model relies on top-down agenda-setting and local policy experiments, with a lag period of approximately 12 months, enabling the rapid establishment of provisional regulations following technological breakthroughs. The 2023 Interim Measures for the Management of Generative Artificial Intelligence Services were drafted and issued in only a few months, with Shenzhen, Hangzhou, and other cities simultaneously rolling out supporting experimental measures, demonstrating agile governance. Its advantage lies in its ability to promptly provide regulatory guidance for emerging technologies and reduce market uncertainty; however, this comes at the cost of significant policy volatility, which poses challenges to regulatory continuity and consistency. From an institutional perspective, China’s efficient administration relies on centralized decision-making combined with local policy experiments, achieving rapid responsiveness through a process of “central direction—local policy experiments—experience dissemination.” In contrast, the governance paradigms of Europe and the United States tend toward a collaborative adaptation model, with a lag period of approximately 24 months. The EU, through its supranational governance structure, has established a unified regulatory framework through systemic legislation, including the General Data Protection Regulation (GDPR), the Digital Services Act, and the AI Act. While the United States lacks unified federal legislation, it exhibits a diverse landscape characterized by parallel federal and state systems and a complementary approach that combines industry self-regulation and judicial precedent. Both emphasize the predictability of rules and procedural justice; however, their legislative cycles are relatively long—for instance, the AI Act took nearly 3 years from proposal to enactment—making it difficult to respond promptly to rapid technological iteration. From an institutional perspective, democratic governance in Europe and the United States prioritizes cross-departmental coordination and public participation, ensuring regulatory stability through the “legislation first, followed by judicial and administrative implementation” approach. It is worth noting that the paths of coordinated adaptation in the EU and the United States exhibit intrinsic differences: the EU relies on its supranational governance structure to establish a unified regulatory framework through systematic legislation such as the AI Act (European Commission, 2025); the United States exhibits a pluralistic governance structure characterized by parallel federal and state levels, with industry self-regulation and judicial precedents serving as complementary elements. However, both emphasize the predictability of rules and procedural justice. From the perspective of evolutionary economics, these two approaches have formed distinctly different trade-off logics in addressing technological uncertainty: the administratively-led model tends to trade short-term flexibility for the risk of early policy volatility. In contrast, the co-adaptive model is dedicated to leveraging institutional continuity to reduce long-term social transaction costs.

5.2 Institutional tensions and adaptation dilemmas arising from core conflicts

Analysis of judicial case data reveals that ownership determinations account for over 40% of cases, making them the most prominent systemic bottleneck. This phenomenon reflects a fundamental contradiction between the human authorship-centered approach of the copyright legal system and AI-generated content. How different jurisdictions respond to this contradiction reveals deep-seated differences in their respective legal rationales. U.S. copyright law adheres strictly to the “human authorship” principle; in Thaler, the federal district court explicitly ruled that copyright law applies only to human authors, refusing to grant protection to AI-generated content. The U.S. model emphasizes the personal nature of the creative act, holding that the legitimacy of the copyright system stems from protecting the fruits of human intellectual labor. While this approach maintains the legal system’s logical consistency, it may stifle industrial investment. Chinese judicial practice follows a different logic: in the Tencent AI Writing Case, the court ruled that AI-generated articles possess originality and should be protected under copyright law. Chinese courts do not deny the “human authorship” principle; rather, they incorporate AI-generated content into the framework of corporate works. By recognizing platforms as rights holders, they protect industrial investment. Their core concern is not who created the content, but who made a substantial investment in its generation. The European Union, meanwhile, is exploring a third path in relevant legislation. Through institutional innovations such as the introduction of data producer rights and mandatory labeling obligations, it seeks to strike a balance between protecting investment and preserving the public domain (Cai, 2021; Millet et al., 2023).

Conflicts over platform liability also reflect profound institutional tensions. Traditional notice-and-takedown rules are premised on platforms’ passive role. However, GAI platforms not only disseminate content but also actively generate it, blurring the lines of their identity. In response to this tension, global regulation is becoming increasingly stringent: the proportion of joint liability rulings against platforms in China is rising, and the EU’s Digital Services Act imposes strict risk prevention and control obligations on super-large platforms. Strengthening platform liability may stifle innovation, particularly affecting small and medium-sized enterprises with limited resources. Striking a balance between safety standards and innovation potential has become a core challenge in institutional design. The institutional root of conflicts over technical compliance lies in balancing algorithmic transparency with the protection of trade secrets. In European and American cases, technical compliance disputes account for 24.25% of all disputes, higher than the 22.58% in China, reflecting the greater emphasis on technical standards in Western regulatory frameworks. From the perspective of complaint mechanisms, institutional adaptation faces a trade-off among judicial, administrative, and legislative mechanisms regarding agility, authority, and uniformity. The judicial mechanism excels in the continuity and precision of case-by-case adjudication but involves lengthy litigation cycles; the administrative mechanism offers rapid response but requires refinement of the boundaries of discretionary power; the legislative mechanism ensures rule uniformity but involves prolonged cycles and struggles to anticipate technological evolution. Future institutional adaptation must transcend the limitations of any single mechanism and explore synergistic coordination among the three. The application of adaptive governance tools, such as regulatory sandboxes, risk-based differentiated responsibility allocation, and algorithmic impact assessments provides a viable pathway for the coordinated operation of these mechanisms.

5.3 The impact of governance models on symbiotic evolutionary pathways

The non-linear structural breaks identified in this study constitute a substantial revision to the assumption of gradual co-evolution in traditional symbiotic evolution theory, revealing a threshold effect in institutional adaptability under scenarios of explosive technological change. The empirical finding that China’s policy weight surged from 0.000050 to 0.998294 between 2021 and 2022 clearly illustrates the structural shift in institutional response from a steady state to a discontinuous phase; this phenomenon directly reflects the “emergent” nature of GAI capabilities and the unpredictability of their social impacts. From the perspective of co-evolution, the institutional environment does not merely passively absorb technological shocks but, through feedback mechanisms, shapes the trajectory of technological development in a reciprocal manner. The administration-led model, by rapidly issuing compliance guidelines, objectively accelerated the adoption of AI technology in digital publishing; the legislation-led model, by establishing long-term rules, guided technological R&D toward technical compliance, thereby shaping technological path dependence. Based on this, this paper proposes the construction of a resilient governance framework as a high-level adaptive mechanism to address the risks of governance lag (Chowdhury, 2022). This framework advocates moving away from the pursuit of a single optimal rule and instead focuses on enhancing the governance system’s capabilities for perception, adaptation, and iteration. Specifically, resilient governance calls for establishing a cross-departmental risk assessment network to strengthen environmental awareness, dynamically adjusting regulatory intensity through a modular policy toolkit, and creating a dialogue platform involving multiple stakeholders to compensate for the cognitive limitations of a single regulator. Institutional designs should incorporate sufficient redundancy and learning mechanisms to avoid prematurely locking into specific technological pathways, ensuring that governance paradigms can be continuously optimized based on practical feedback, and ultimately guiding technological innovation and institutional norms toward a sustainable, virtuous cycle of interaction.

6 Conclusion

This study applies the theory of technological-institutional co-evolution to the rapidly evolving field of generative artificial intelligence. By employing breakpoint tests and a lag period, it corrects the implicit assumption of gradual linear progression in traditional theories, thereby providing a quantitative analytical tool for understanding how discontinuous technological shocks trigger institutional responses. The study identifies two governance paradigms—catch-up and collaborative adaptation model—and reveals the inherent trade-off between agility and stability in the two models of administrative leadership and legislation-led governance, thereby providing a new typological framework for comparative institutional analysis. Methodologically, the study integrates policy text topic modeling, judicial case coding, and breakpoint tests to construct a replicable mixed-methods approach, offering analytical tools and an operational paradigm for future research on technology governance.

The study proposes three actionable pathways for collaborative governance in digital publishing. (1) Establish a tiered algorithm compliance filing system, whereby the government implements differentiated filing requirements based on risk levels, and companies establish internal positions for algorithmic compliance review and submit periodic assessment reports; (2) Build a rapid mediation mechanism for copyright disputes and an industry self-regulation mechanism, with the government guiding industry associations to develop copyright guidelines for specific sectors, and companies establishing traceable copyright management databases; (3) Implement a linked model of regulatory sandboxes and compliance pilot programs, where the government establishes innovation pilot zones allowing for flexible regulation, while enterprises submit pilot proposals and regularly report operational data, transforming successful experiences into industry standards. These three pathways address three core conflicts—technical compliance, copyright ownership, and platform liability—respectively, internalizing external regulatory requirements into corporate compliance mechanisms.

However, this study has certain limitations. The sample size of policy texts and judicial cases is limited. While combining the EU and the U.S. into a unified “Europe-U.S.” analysis helps highlight the divergent paths of governance traditions, it may obscure the institutional differences between the two. Faced with the ongoing interplay between technology and institutions, the adaptive efficiency of a governance system depends not only on response speed but also on the ability to learn from practice and continuously optimize. This study provides theoretical tools and an empirical foundation for future exploration of dynamic governance mechanisms.

References

• 1

  Amazon (2025). An article reveals the logic and evolution of the Amazon algorithm - a knowledge-sharing community for cross-border e-commerce. We are Sellery. Available online at: https://www.wearesellers.com/question/101972 (Accessed April 27, 2025).

  • Google Scholar

• 2

  BarrettP.HunterJ.MillerJ. T.HsuJ.-C.GreenfieldP. (2005). Matplotlib–a portable python plotting package., in astronomical data analysis software and systems XIV. Available online at: https://adsabs.harvard.edu/full/2005ASPC..347...91B (Accessed April 23, 2025).

  • Google Scholar

• 3

  BenderE. M.GebruT.McMillan-MajorA.ShmitchellS. (2021). On the dangers of stochastic parrots: can language models be too big?, In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, (Virtual Event Canada: Association for Computing Machinery), 610–623

  • Google Scholar

• 4

  BommasaniR.HudsonD. A.AdeliE. (2022). On the opportunities and risks of foundation models. doi: 10.48550/arXiv.2108.07258

  • CrossRef
  • Google Scholar

• 5

  CaiJ. (2021). Research on insider guarantees issues through the lens of bankruptcy: economic approach and legislative recommendations. J. China Univ. Polit. Sci. Law. 4, 155–181. Available online at: https://kns.cnki.net/kcms2/article/abstract?v=2FByMehzK2qKspKOLl_UV8hvl_JUDQ2k8I_ek1AIJIh3etPnFhNiUnKCKEBlXK7ud5wqUzRnzNSW1HcF6QM88BgwQG0gMsL4rafTDxIhwOBx4H0mZtd7QbRpF4aCxZe8hhIGsV8jvB7iJbdGddN6Ma6kVE40-zWPTi5XNch2uX7zvRi24V-17w==&uniplatform=NZKPT&language=CHS

  • Google Scholar

• 6

  ChangK. (2020). The impact of the development of artificial intelligence technology on the publishing process chain. Technol Markets27, 42–43. Available online at: https://kns.cnki.net/kcms2/article/abstract?v=XtSw_LC1RHmG5vwiD-hJqdc-ojHuvBQ17BgCn6iGuBKd0-pluw4_4NLmxfqg-hrXDHimKUctSksDf6fTGOqD3bwWy1R9G8VaJ8RIc-aXQqKsBGqG_krsgH_AOalJtDztuI2PaPPUSJ5VWji0kRxYHqRfD4uSrmI5AMeS2Plfv7i9Q9dg80Y5lY1tHlcUfTGK&uniplatform=NZKPT&language=CHS

  • Google Scholar

• 7

  ChauhanC. (2024). The impact of generative artificial intelligence in scientific content synthesis for authors. Am. J. Pathol.194, 1406–1408. doi: 10.1016/j.ajpath.2024.06.002,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 8

  ChowdhuryR. H. (2022). Digital government initiatives and national resilience: how digital governance frameworks were transformed post-COVID to maintain national services. World J. Adv. Eng. Technol. Sci.7, 224–240. doi: 10.30574/wjaets.2022.7.1.0098

  • CrossRef
  • Google Scholar

• 9

  CramerB. W. (2017). The attention merchants: the epic scramble to get inside our heads. J. Inf. Policy7, 268–271. doi: 10.5325/jinfopoli.7.2017.0268

  • CrossRef
  • Google Scholar

• 10

  DengS.DingW.WangF.WangH. (2025). A comparative study of generative AI regulatory policies at home and abroad from the perspective of “tool-structure”. J. Inf. Resour. Manag.15, 54–68. doi: 10.13365/j.jirm.2025.01.054

  • CrossRef
  • Google Scholar

• 11

  DuR. (2024). “News recommendation and information cocoons: the impact of algorithms on news consumption,” in Handbook of Applied Journalism, eds. BarkhoL.Lugo-OcandoJ. A.JamilS. (Cham: Springer Nature Switzerland), 43–61.

  • Google Scholar

• 12

  DuanY.DongX. (2025). Technology-system two-way adaptation: the logic of reconstructing government supervision mechanisms by digital-intelligence transformation--a case study based on digital reform of market supervision in Zhejiang Province. E-Gov. 5, 1–12. doi: 10.16582/j.cnki.dzzw.2025.11.006

  • CrossRef
  • Google Scholar

• 13

  European Commission (2025). The impact of the digital services act on digital platforms | shaping Europe’s digital future. European Commission. Available online at: https://digital-strategy.ec.europa.eu/en/policies/dsa-impact-platforms (Accessed April 23, 2025).

  • Google Scholar

• 14

  GaculaM. C.SinghJ. (eds.) (1984). “Nonparametric Statistical Methods,” in Statistical Methods in Food and Consumer Research, eds. GaculaM. C.SinghJ. (San Diego: Academic Press), 313–359. doi: 10.1016/B978-0-08-091831-0.50014-X

  • CrossRef
  • Google Scholar

• 15

  GaffarH.AlbarashdiS. (2024). Copyright protection for AI-generated works: exploring originality and ownership in a digital landscape. Asian J. Int. Law15, 23–46. doi: 10.1017/S2044251323000735

  • CrossRef
  • Google Scholar

• 16

  GeelsF. W. (2002). Technological transitions as evolutionary reconfiguration processes: a multi-level perspective and a case-study. Res. Policy31, 1257–1274. doi: 10.1016/S0048-7333(02)00062-8

  • CrossRef
  • Google Scholar

• 17

  Gomez-UribeC. A.HuntN. (2015). The netflix recommender system: algorithms, business value, and innovation. ACM Trans. Manag. Inf. Syst.6, 1–19. doi: 10.1145/2843948

  • CrossRef
  • Google Scholar

• 18

  GuoR. (2018). Research on the impact of the development of artificial intelligence technology on the publishing process chain. Friends Edit.10, 78–83. doi: 10.13786/j.cnki.cn14-1066/g2.2018.10.014

  • CrossRef
  • Google Scholar

• 19

  HayesA. F.KrippendorffK. (2007). Answering the call for a standard reliability measure for coding data. Commun. Methods Meas.1, 77–89. doi: 10.1080/19312450709336664

  • CrossRef
  • Google Scholar

• 20

  HairJ. FBabinJ. B.AndersonR. E. (2018). Multivariate data analysis 8th edition. US: Cengage Learning. Available online at: https://www.vitalsource.com/products/multivariate-data-analysis-joseph-f-hair-barry-j-babin-v9780357755228 (Accessed April 26, 2025).

  • Google Scholar

• 21

  KocakZ. (2024). Publication ethics in the era of artificial intelligence. J. Korean Med. Sci.39:e249. doi: 10.3346/jkms.2024.39.e249,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 22

  Lu HanMohong Liu (2024). Digital rights management (DRM) technologies and legal research: applications and regulations of encryption, digital watermarking, and copyright protection systems. Research82, 106–111. doi: 10.54254/2755-2721/82/20240957

  • CrossRef
  • Google Scholar

• 23

  MilletK.BuehlerF.DuG.KokkorisM. D. (2023). Defending humankind: anthropocentric bias in the appreciation of AI art. Comput. Hum. Behav.143:107707. doi: 10.1016/j.chb.2023.107707

  • CrossRef
  • Google Scholar

• 24

  NelsonR. R.WinterS. G. (1973). Toward an evolutionary theory of economic capabilities. Am. Econ. Rev.63, 440–449.

  • Google Scholar

• 25

  NelsonR. R.WinterS. G. (1983). An evolutionary theory of economic change. South. Econ. J.50:590. doi: 10.2307/1058233

  • CrossRef
  • Google Scholar

• 26

  NorthD. C. (1971). Institutional change and economic growth. J. Econ. Hist.31, 118–125. doi: 10.1017/s0022050700094109

  • CrossRef
  • Google Scholar

• 27

  NorthD. C. (1990). Institutions, Institutional Change and Economic Performance. Cambridge: Cambridge University Press.

  • Google Scholar

• 28

  O'SullivanDoniePassantinoJon (2023). 'Verified' Twitter Accounts share fake Image of 'explosion' near the Pentagon, Causing confusion | CNN Business. CNN. Available online at: https://www.cnn.com/2023/05/22/tech/twitter-fake-image-pentagon-explosion/index.html (Accessed April 23, 2025).

  • Google Scholar

• 29

  RadfordA.WuJ.ChildR.LuanD.AmodeiD.SutskeverI. (2019). Language Models are Unsupervised Multitask Learners. US: OPEN AI.

  • Google Scholar

• 30

  RasenbergerWill (2024). Copyright and creativity in the age of AI | the regulatory review. The Regulatory Review. Available online at: https://www.theregreview.org/2024/01/23/rasenberger-copyright-and-creativity-in-the-age-of-ai/ (Accessed April 22, 2025).

  • Google Scholar

• 31

  RyzhkoO.KrainikovaT.VodolazkaS.SokolovaK. (2024). Generative AI changes the book publishing industry: reengineering of business processes. Commun. Soc.37, 255–271. doi: 10.15581/003.37.3.255-271

  • CrossRef
  • Google Scholar

• 32

  SaikalyT. (2023). The transformative role of artificial intelligence in the publishing industry. Available online at: https://fadel.com/wp-content/uploads/2025/02/The-Transformative-Role-of-AI-in-the-Publishing-Industry_4.pdf (Accessed April 23, 2025).

  • Google Scholar

• 33

  SamuelsonP. (2023). Generative AI meets copyright. Science381, 158–161. doi: 10.1126/science.adi0656,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 34

  StraussA. L.CorbinJ. M. (1990). Basics of Qualitative Research: Grounded Theory Procedures and Techniques. Newbury Park: Sage Publications.

  • Google Scholar

• 35

  SunL. (2005). Principles and operations of structural equation modeling (SEM). J. Ningbo Univ. (Educ. Sci. Ed.)2, 31–34, 43.

  • Google Scholar

• 36

  TeeceD. J. (2018). Profiting from innovation in the digital economy: enabling technologies, standards, and licensing models in the wireless world. Res. Policy47, 1367–1387. doi: 10.1016/j.respol.2017.01.015

  • CrossRef
  • Google Scholar

• 37

  VaswaniA.ShazeerN.ParmarN.UszkoreitJ.JonesL.GomezA. N.et al (2023). Attention is All You Need. arXiv. [Preprint].

  • Google Scholar

• 38

  VealeM.Zuiderveen BorgesiusF. (2021). Demystifying the draft EU artificial intelligence act. Available online at: https://osf.io/38p5f (Accessed April 23, 2025)

  • Google Scholar

• 39

  ZhangY.DaiH.SuY. (2025). Research on problems in the process of journal publishing in the era of artificial intelligence. China Media Technol.1, 136–140. doi: 10.19483/j.cnki.11-4653/n.2025.01.029

  • CrossRef
  • Google Scholar

• 40

  ZhuJ.LiapisA.RisiS.BidarraR.YoungbloodG. M. (2018). “Explainable AI for designers: a human-centered perspective on mixed-initiative co-creation,” in 2018 IEEE Conference on Computational Intelligence and Games (CIG), (Maastricht: IEEE Press), 1–8.

  • Google Scholar