Research on artificial intelligence governance and policy

Research on artificial intelligence (AI) policy has gradually shifted from a techno-economic perspective toward a broader governance-oriented perspective. Early studies mainly treated AI policy as an institutional instrument for mobilizing innovation resources, improving industrial competitiveness, and supporting national strategic development. As AI technologies have become increasingly embedded in economic and social systems, however, scholarly attention has expanded to questions of ethics, regulation, public governance, and social legitimacy. In this context, AI is no longer understood merely as a technological agenda, but as a governance issue involving strategic planning, institutional design, and social coordination [5–11].

At the national strategic level, existing studies show that AI policy is closely connected to the socio-technical construction of state priorities and political legitimacy. Comparative analyses suggest that AI strategies vary substantially across political and developmental contexts: developed economies tend to emphasize ethical governance, transparency, explainability, and human-centric values, whereas emerging economies more often prioritize technological upgrading, industrial transformation, and modernization goals [7,8]. At the same time, many studies point out that current AI policy frameworks still suffer from major weaknesses, including ambiguous concepts, insufficient implementation detail, and limited evaluation mechanisms [6,8,9]. These limitations indicate that AI governance cannot be adequately understood through policy declarations alone, but must also be examined in terms of whether policy systems are institutionally coherent and operationally actionable.

At the level of ethics and social governance, scholars have further highlighted the tension between regulatory discourse and technological reality. Studies show that AI experts often adopt function-oriented understandings of AI, whereas policymakers frequently frame AI in anthropomorphic terms, thereby shifting policy attention toward future imaginaries rather than currently deployed systems [9]. Related research in emerging and developed contexts alike has shown that AI governance frameworks continue to face structural challenges in addressing accountability, transparency, risk management, and cross-sector coordination [10,11]. In addition, studies on global AI indices and benchmarking practices suggest that national AI trajectories may become locked into pre-existing competitiveness frameworks, thereby narrowing the space for alternative governance models [12].

With the increasing use of computational methods, AI policy research has also moved from qualitative discourse analysis toward more data-oriented approaches. Studies using machine learning, natural language processing, and semantic analysis have identified latent themes, policy priorities, and discursive patterns in AI policy texts [13–15].‌‌ These studies have significantly enriched understanding of the content and orientation of AI policy systems. However, the existing literature remains largely content-oriented. Most studies focus on policy themes, ethical concerns, discourse patterns, or policy tool classifications, while relatively limited attention has been paid to the structural consistency and internal coordination of AI policy systems. This limitation is particularly important in the case of China, where central-level AI policy documents function not only as strategic texts but also as key institutional carriers of AI governance.

Research on the PMC index model for policy evaluation

The Policy Modeling Consistency (PMC) Index Model emerged within the broader evolution of policy evaluation from result-oriented assessment toward more systematic, reflexive, and integrative forms of analysis [16,17]. Rooted in the Omnia Mobilis hypothesis, the PMC model evaluates the internal consistency of policy texts by constructing a multi-input–multi-output matrix that captures the degree of coordination among policy objectives, instruments, and implementation elements [18]. Compared with purely descriptive approaches, the PMC framework offers a structured, parameterized, and visualized means of examining the internal logical configuration of policy design.

Existing studies have applied the PMC model to a wide range of policy fields, including ecological governance, industrial development, regional policy, public service reform, and social governance [19–24]. Across these applications, the model has demonstrated particular value in identifying structural imbalances within policy design, such as weaknesses in timeliness, incentive mechanisms, implementation support, or cross-sector coordination. In this sense, the PMC model is especially useful when the analytical objective is not only to identify policy content, but also to evaluate whether policy structures are internally coherent and logically coordinated.

Despite its growing use in policy evaluation, the application of the PMC framework to AI policy remains limited. This is an important gap, because AI governance is inherently multidimensional, involving innovation support, industrial development, regulatory design, ethical governance, implementation arrangements, and coordination across governance levels. These characteristics make AI policy especially suitable for PMC-based structural evaluation.

Summary

Overall, existing studies have generated important insights into AI governance, policy discourse, ethical regulation, and policy instrument design, but they have rarely examined the structural consistency of AI policy systems in a systematic way. At the same time, the PMC Index Model has shown clear methodological value in evaluating the internal coordination of policy structures, yet its application to AI policy remains underdeveloped.

Accordingly, this study introduces the PMC model into the analysis of China’s central-level AI policies in order to move from thematic description toward structural evaluation. By examining policy consistency across multiple dimensions, the study seeks to reveal the internal logic, coordination patterns, and potential structural imbalances of China’s AI governance framework.

Research framework

The overall logical foundation of this study lies in the fact that artificial intelligence (AI) policy has increasingly become a core component of China’s national strategy and governance practice. How to conduct a systematic and quantitative evaluation of the policy texts issued at the central government level constitutes not only a methodological challenge within academic research but also a practical necessity for policy refinement. Anchored in this awareness, the study first reviews existing research on AI policy and policy evaluation methods, identifying that current analytical approaches remain limited and have yet to fully reveal the structural characteristics and developmental logic of China’s AI policy system.

Accordingly, this study establishes the Policy Modeling Consistency (PMC) Index Model as its core methodological framework, supplemented by text-mining techniques for auxiliary content analysis. A total of fifty-four AI-related policy documents issued by China’s central government between 2016 and 2025 were selected as the sample, and the ROSTCM6.0 software was employed to perform word frequency and semantic analysis. On this basis, a PMC index variable indicator system was constructed to ensure the rationality and reliability of the indicators, thereby minimizing subjectivity in variable formulation and supporting the objective identification of policy content within China’s AI policy corpus.

Further, the PMC index model is applied to conduct a quantitative evaluation of the sampled policies, and the results are visualized through three-dimensional surface mapping to illustrate the degree of coordination and divergence across different dimensions of China’s AI policy design. Based on the evaluation outcomes, the study proceeds to analyze and discuss the findings, identifying both the strengths and deficiencies of existing AI policy arrangements and offering targeted recommendations for future policy optimization. The entire research process unfolds in a progressive and reproducible manner, encompassing the origin of the study, literature review, methodological determination, data selection, model construction, result visualization, and interpretive discussion, thereby forming a coherent and verifiable research pathway. Figure 1 presents the overall research framework.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Research framework of the PMC index model for the quantitative evaluation of China’s artificial intelligence policies.

https://doi.org/10.1371/journal.pone.0337024.g001

Research method

This study combines text mining with the Policy Modeling Consistency (PMC) Index Model to evaluate the structural consistency of China’s central-level artificial intelligence (AI) policies. The methodological logic is sequential rather than purely automated. Text mining is first used to explore the thematic features of the policy corpus and to provide an empirical basis for variable construction. On this basis, the PMC indicator system is developed through the integration of prior policy-evaluation research, corpus-based interpretation, and researcher judgment. The model is then applied to representative policy cases through structured binary coding, PMC index calculation, and visual analysis.

The PMC Index Model serves as the core analytical framework of the study. Proposed by M. A. R. Estrada (2011), the model was designed to evaluate the internal consistency of policy texts through a structured multi-input–multi-output framework [25]. Its analytical value lies in its ability to transform complex policy content into a measurable configuration of variables, thereby revealing the strengths, weaknesses, and internal coordination of policy design. In the present study, the PMC model is used not simply to rank policies, but to assess whether key elements of AI governance—such as strategic orientation, policy content, incentive arrangements, target actors, and levels of influence—are configured in a coherent and balanced manner.

More specifically, the research process consists of four stages. First, a corpus of 54 AI-related policy documents issued by China’s central government between 2016 and 2025 was constructed through systematic retrieval, screening, and validation. Second, the full texts of these policy documents were merged into a single corpus and processed through Chinese word segmentation, word-frequency analysis, and semantic co-occurrence analysis. This stage was conducted primarily to identify recurrent policy terms, thematic clusters, and semantic associations within the corpus. Third, based on the text-mining results, relevant policy-evaluation literature, and close reading of the policy texts, a PMC indicator system containing nine primary variables and forty-three secondary variables was established. Fourth, five representative policies were selected for detailed PMC evaluation. Each secondary variable was coded using binary assignment criteria (1 = Yes; 0 = No) according to the predefined scoring matrix, and the resulting values were used to calculate the PMC Index and generate visual outputs.

Accordingly, the text-mining component of the study should be understood as supportive rather than determinative. It was used to reveal the semantic structure of the policy corpus and to assist in the identification of analytically relevant dimensions, but it did not automatically generate the final policy scores. The final construction of the indicator system, the interpretation of policy elements, and the assignment of binary values all relied on structured researcher judgment guided by explicit coding rules. In this sense, the study combines automated corpus processing with interpretive policy analysis in order to achieve both analytical transparency and methodological rigor.

As illustrated in Fig 2, the analytical procedure can be summarized into four main steps. First, the evaluation indicators and corresponding coding criteria were established, including the definition of primary and secondary variables and the scoring basis for each secondary indicator. Second, the policy corpus was processed through text mining in order to identify high-frequency terms, semantic associations, and major thematic tendencies relevant to AI governance. Third, based on these results and the predefined scoring matrix, the selected representative policies were organized into a multi-input–output structure and coded using binary assignment criteria. Fourth, the PMC Index was calculated for each policy, and the results were visualized through PMC surface mapping to reveal the relative strengths, weaknesses, and structural characteristics of policy design.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Construction process of the PMC index model for China’s artificial intelligence policies.

https://doi.org/10.1371/journal.pone.0337024.g002

Taken together, these steps form a transparent and reproducible analytical workflow that combines automated corpus processing with structured policy interpretation, thereby enabling the systematic evaluation of the structural consistency of China’s AI policy system.The reporting of the text-mining procedure was further refined with reference to prior methodological studies on transparent text-analytic workflows, including Danesh et al. (2021) [26].

Data sources

This study draws on policy documents related to artificial intelligence (AI) issued by China’s central government between 2016 and 2025 (up to August 31, 2025), comprising a total of 54 valid samples. The choice of 2016 as the starting point is based on two main considerations. First, although no overarching AI policy framework had yet been released that year, several national development and sectoral plans—such as the 13th Five-Year Plan for National Informatization, the 13th Five-Year Plan for Strategic Emerging Industries Development, and the “Internet Plus” Artificial Intelligence Three-Year Action Implementation Plan—explicitly referenced AI development goals or directions in certain sections, signifying the initial inclusion of AI within China’s national policy system. Second, the formal issuance of the New Generation Artificial Intelligence Development Plan by the State Council in 2017 marked the beginning of a new phase characterized by systematic planning and strategic deployment of AI policy in China. Therefore, setting 2016 as the baseline year carries both retrospective and practical significance, providing a comprehensive foundation for analyzing the trajectory of policy evolution.

The study covers AI policy documents released between 2016 and 2025, with the specific data retrieval period spanning August 30–31, 2025. Policy retrieval and screening were conducted primarily through authoritative official sources, including the Chinese Government Website (www.gov.cn) and PKUlaw (pkulaw.com). During the cross-checking and validation process of AI-related policies, the following selection principles were applied:

1. Keyword relevance: Policies were identified using “Artificial Intelligence” and “New Generation Information Technology” as core keywords. Only texts exhibiting strategic guidance, institutional construction, or industrial support functions were retained, ensuring strong thematic relevance to the research topic.
2. Authority and validity of data sources: All policy texts were obtained from official and authoritative central government or ministerial websites, ensuring legality, reliability, and completeness. Only currently valid policies were included, while expired or superseded documents were excluded.
3. Normative document types: The study primarily selected formal and regulatory policy documents, including plans, opinions, guidelines, notices, implementation schemes, and development strategies. In contrast, non-normative materials—such as leaders’ speeches, internal memos, training notices, meeting summaries, annual reports, or public name lists—were excluded to maintain analytical rigor and policy comparability.

Through this systematic process of retrieval, screening, and validation, the final corpus was constructed to reflect the central policy architecture of China’s AI governance framework. For purposes of transparency and reproducibility, the complete list of the 54 policy documents is provided in S1 File in S1 Data.

According to the selection principles described above, a total of 54 valid policy documents were included in the sample. As shown in Table 1 and Fig 3, the annual distribution of China’s AI policies exhibits distinct stage characteristics. The release of the New Generation Artificial Intelligence Development Plan in 2017 marked a critical turning point, after which the number of AI-related policy issuances increased significantly. Two noticeable peaks occurred in 2018 and 2021, reflecting a cyclical policy trajectory characterized by planning, implementation, and upgrading.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Selected sample of China’s artificial intelligence policy documents.

https://doi.org/10.1371/journal.pone.0337024.t001

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Trend of the number of artificial intelligence policy documents issued by the Chinese central government (2016–2025).

Note: The dashed line indicates the overall trend in annual policy issuance.

https://doi.org/10.1371/journal.pone.0337024.g003

The publication of the Opinions on Deeply Implementing the “AI+” Action Plan by the State Council in 2025 represents the most recent central-level document, signifying China’s strategic deployment of AI integration within the broader context of “new productive forces.” This policy continues the institutional trajectory established by previous documents and implies the beginning of a new phase of systemic policy upgrading.

Overall, the time-series characteristics of the sample demonstrate both continuity and representativeness, providing a comprehensive reflection of the evolutionary stages and structural features of China’s AI policy system. These publicly available policy documents constitute the sole data source of the study and provide a solid empirical foundation for the subsequent phases of text mining and structural evaluation within the PMC index modeling framework. Accordingly, all coding, interpretation, and scoring procedures were conducted by the authors within a document-analysis framework. No human participants, external experts, interviewees, survey respondents, personal data, medical records, biological materials, or field research were involved at any stage of the research; therefore, ethics approval and informed consent were not required.

Variable setting and parameter identification for the valuation of China’s AI policies

Policy text mining and word frequency analysis.

The preprocessing of policy texts constituted the starting point for variable construction in this study. The input for text mining consisted of the full texts of the 54 central-level AI policy documents included in the corpus, rather than only titles or selected excerpts. Text-mining techniques were employed to identify recurrent terms, semantic associations, and major thematic tendencies within the corpus, thereby providing an empirical basis for subsequent variable setting. In this sense, text mining was used to support corpus exploration and variable identification, rather than to automatically generate the final PMC scores.

Specifically, the text-mining procedures were conducted using ROSTCM6.0 software. First, all 54 AI-related policy documents were merged into a single plain-text corpus and imported into the software for Chinese word segmentation. In the word-segmentation stage, this study used the lexical resource available in the ROSTCM6.0 environment as the initial segmentation basis, and further refined the segmentation output through corpus-oriented manual checking, selective retention of AI-relevant policy expressions, and filtering. Based on the initial segmentation results, structurally repetitive or analytically uninformative expressions—such as article numbers, chapter labels, and procedural terms—were identified and removed, while policy-relevant terms associated with AI development and governance were retained and better aligned with the analytical purpose of the study. This process improved both the accuracy of policy-text identification and the validity of subsequent word-frequency analysis.

Second, word-frequency analysis was performed on the processed corpus to identify the most frequently occurring substantive terms. Expressions with limited analytical value, such as “shall not,” “process,” and “should,” were excluded during preprocessing. The twenty most frequent terms in China’s AI policy texts were then retained to construct the high-frequency word list (see Fig 4).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Top 20 high-frequency terms in China’s artificial intelligence policy documents.

https://doi.org/10.1371/journal.pone.0337024.g004

Third, to provide a more intuitive representation of the internal structure and relational patterns within China’s AI policy texts, semantic co-occurrence analysis was conducted and visualized using NetDraw. The segmented and filtered terms were used to generate a policy keyword co-occurrence network (see Fig 5). In the resulting network, lines between nodes represent co-occurrence relationships between keywords, while nodes with a higher number of connections exhibit greater degree centrality and therefore occupy a more prominent position in the overall policy discourse structure.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Semantic co-occurrence network of high-frequency terms in China’s artificial intelligence policy documents.

https://doi.org/10.1371/journal.pone.0337024.g005

As shown in Fig 4, the word-frequency distribution of China’s AI policy corpus indicates that terms such as “artificial intelligence,” “technology,” “data,” “development,” “application,” and “intelligence” appear most frequently, followed by “service,” “capacity building,” “innovation,” “platform,” “system,” “security,” and “management.” This pattern suggests that China’s AI policy framework is organized around three relatively stable thematic dimensions—technological capability, scenario implementation, and institutional support—reflecting a consistent orientation toward linking innovation capacity with practical deployment and governance mechanisms.

The results presented in Fig 5 directly reflect the primary emphases of China’s AI policy discourse. Structurally, the co-occurrence network is centered on the core keywords Technology, Artificial Intelligence, Application, Development, and Innovation, which exhibit the highest frequency and the widest range of semantic connections. These terms occupy the central hub of the network and form three relatively stable cohesive subgroups.

From the perspective of content generation and thematic clustering, the semantic network can be interpreted as comprising three major clusters:

1. (1) The “technological foundation”cluster(Data–Platform–System–Foundation–CapacityBuilding–R&D/Research) demonstrates a high degree of connectivity, revealing China’s integrative policy logic that links data, platforms, systems, basic research, and capacity building into a unified supply structure. The nodes platform and data play a prominent bridging role, connecting technological innovation with real-world application.
2. (2) The “governance and assurance” cluster (Security–Standard–Management, extending to Risk, Framework, and Evaluation) reflects the principle of “balancing development with security.” This cluster highlights the regulatory boundaries of China’s AI policy system, wherein standardization, management, and evaluation mechanisms are used to control potential risks within an institutional framework.
3. (3) The “application diffusion” cluster (Enterprise–Service–Field/Scenario) directly couples Application with Technology, pointing toward large-scale industrial deployment and the cultivation of public service ecosystems. Around this core, numerous action-oriented nodes—such as Implementation, Promotion, Strengthening, Enhancement, and Acceleration—are densely distributed, revealing an execution-oriented policy tendency driven by organizational advancement, scheduling, and capacity enhancement.

At the outer periphery of the network lie relatively marginal terms such as Risk, Society, and Management, which mainly carry conceptual or macro-level contextual meanings rather than operational functions. This core–belt–periphery structure indicates that China’s AI policy network embodies a dual core of technology and application, bounded by the external constraints of capacity building as the foundation and security regulation as the boundary.

Based on these analytical findings, the study preliminarily identified several macro-level policy dimensions that informed the subsequent construction of the PMC variable system. These dimensions were not treated as automatically generated indicators, but were further refined through reference to prior policy-evaluation studies and structured interpretation of China’s AI policy texts. They include:

• Strategic Orientation (Development,Innovation,Priority);
• Technology and R&D (Algorithm,Platform,System,Foundation,R&D);
• Data and Basic Capacity (Data,Standard,Capacity Building,Resource);
• Application and Industrialization (Application,Enterprise,Service,Field,Scenario);
• Governance,Security and Ethics (Security,Management,Framework,Risk,Evaluation,Ethics);
• Capacity Building and Implementation Mechanisms (Talent,Implementation,Coordination,Incentives,Strengthening,Enhancement).

Variable classification and parameter identification.

In developing the hierarchical variables, this study drew upon three sources of evidence: relevant academic theories, existing empirical studies on policy evaluation, and close interpretation of China’s AI policy texts. The text-mining results presented above were used to identify recurrent themes, semantic clusters, and policy emphases within the corpus, thereby providing an empirical basis for variable design. However, the indicator system was not treated as a direct output of automated text mining. Rather, it was constructed through a researcher-led process that combined corpus-based findings with prior policy-evaluation literature and structured interpretation of the policy texts.

On this basis, a total of nine primary variables and forty-three secondary variables were established to form the evaluation indicator system for the PMC model. The complete framework is presented in Table 2. Accordingly, Table 2 should be understood as a researcher-constructed indicator system informed by both text analysis and theoretical reference, rather than as an automatically generated classification scheme. In substantive terms, the indicator design seeks to capture the multidimensional structure of AI policy, including policy nature, timeliness, policy level, policy content, policy domain, incentive measures, policy evaluation, target actors, and level of influence.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Quantitative evaluation indicator system and sources for China’s artificial intelligence policiesprimary variable.

https://doi.org/10.1371/journal.pone.0337024.t002

Following the construction of the indicator system, each secondary variable was operationalized using binary assignment criteria. A value of 1 was assigned when the corresponding policy element was explicitly present in the policy text, and 0 otherwise. This coding procedure enabled the transformation of qualitative policy content into a structured input format suitable for PMC evaluation. For transparency and reproducibility, the detailed indicator scoring matrix, including evaluation criteria, operational coding notes, binary scoring rules, and source bases for all secondary variables, is provided in S3 File in S1 Data.

Construction of the multi-input–output table.

The multi-input–output table operationalizes the indicator system presented in Table 2 by reorganizing the nine primary variables and their corresponding forty-three secondary variables into a standardized structure for PMC coding and calculation. In other words, Table 3 should be understood as the formal coding framework derived from Table 2, rather than as an independently generated analytical result.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Multi-input–output table of PMC model variables for China’s artificial intelligence policies.

https://doi.org/10.1371/journal.pone.0337024.t003

Within this structure, each secondary variable is assigned a binary value according to the coding criteria defined in the indicator system. If a given policy text explicitly contains the policy element represented by a secondary variable, the value is coded as 1; otherwise, it is coded as 0. This binary assignment procedure allows the qualitative content of policy texts to be transformed into a structured and comparable input format for PMC evaluation. Based on this logic, the multi-input–output table constructed for China’s AI policies is presented in Table 3.

Calculation of the PMC index.

The calculation of the Policy Modeling Consistency (PMC) Index proceeded on the basis of the standardized coding structure shown in Table 3. First, the binary values assigned to all secondary variables were entered into the multi-input–output table. Second, the values of the primary variables were calculated according to the aggregation rules specified in Equations (1)–(3). Third, the PMC Index of each policy was obtained by summing the values of all primary variables in accordance with Equation (4).

In this process, the role of Table 4 is interpretive rather than computational. Specifically, Table 4 provides the policy classification standard used to interpret the PMC scores after they have been calculated. Referring to previous studies [34], this study adopts four levels of PMC-based policy evaluation, ranging from “Poor” to “Excellent.” Accordingly, Table 4 should be understood as the analytical basis for score interpretation, rather than as a table generated directly from text-mining output or coding procedures.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. PMC index–based policy classification standards.

https://doi.org/10.1371/journal.pone.0337024.t004

The calculation of the PMC Index therefore involved two linked stages: the transformation of coded policy content into measurable index values, and the interpretation of those values according to the PMC classification standard. The specific equations used in this study are presented below.

(1)(2)(3)(4)

To facilitate substantive interpretation, the PMC Index in this study is understood as an indicator of the structural consistency of policy design across the selected evaluation dimensions. A higher PMC score does not imply that a policy is universally superior in every respect; rather, it indicates that the policy exhibits a relatively more complete, balanced, and internally coordinated configuration across dimensions such as policy nature, timeliness, content, incentive measures, target actors, and level of influence. By contrast, a lower PMC score suggests that, although the policy may still be valid and practically relevant, its structural design is relatively less comprehensive or less balanced, with one or more dimensions receiving weaker coverage or support.

In practical terms, relatively strong structural consistency refers to a policy design in which strategic orientation, temporal planning, substantive content, implementation support, and governance targets are aligned in a comparatively coherent manner. Relatively weak structural consistency, by contrast, refers to a policy structure in which one or more of these dimensions are underdeveloped, weakly connected, or insufficiently supported, thereby limiting the overall internal coherence of the policy. Accordingly, policies classified as “Excellent” can be interpreted as exhibiting a comparatively high level of structural completeness and coordination across major dimensions; those classified as “Good” indicate a generally coherent policy structure with some remaining weaker aspects; those classified as “Qualified” reflect a more basic or uneven structural configuration; and those classified as “Poor” suggest a relatively fragmented or weakly supported policy structure.

PMC surface equation.

To present the PMC Index in a more intuitive three-dimensional form for facilitating policy evaluation and improvement, the PMC surface visualization in this study was implemented using Python (version 3.8). The Matplotlib and SciPy libraries were employed to generate three-dimensional plots.

In this model, nine primary variables were established, forming a 3 × 3 PMC matrix, and the calculation of the PMC surface is expressed as shown in Equation (5) below:

(5)

Transparency and reproducibility materials

To enhance methodological transparency and reproducibility, this study provides supplementary materials documenting the core steps of the analytical process. These materials cover corpus construction, text processing, indicator operationalization, and PMC scoring, making the analytical procedure explicit and verifiable. Because the main manuscript prioritizes analytical clarity and representative findings, detailed supporting documentation is provided separately in the supplementary files.

First, the supplement includes a complete list of the 54 central-level AI policy documents comprising the corpus, along with essential metadata (e.g., policy title, issuance year, issuing authority, and source). While the main text highlights selected representative policies for readability, the full corpus is provided to support sample verification and future comparative research.

Second, to demystify the text-mining stage, the text-processing notes and the segmentation-related lexical resource used in the text-mining stage are included. These materials clarify how raw policy texts were segmented, filtered, and matched to relevant policy features, thereby improving the interpretability of the transition from raw data to analytical indicators.

Third, the files detail the operational scoring matrix for the PMC evaluation system. This encompasses the definitions of the nine primary variables and 43 secondary indicators, the binary scoring criteria for each, and the coding notes applied during the assignment process. Providing this information ensures the scoring logic remains transparent and minimizes ambiguity in indicator interpretation.

Finally, the complete PMC evaluation outputs are reported in the supplement. Beyond the representative findings discussed in the main text, providing the full set of scoring results facilitates independent verification and methodological replication.

Taken together, these materials extend the study’s “text–indicator–metric–visualization” pipeline, establishing a robust foundation for verification, replication, and future cross-jurisdictional comparisons in AI policy research.

Selection of evaluation objects

The full corpus of 54 central-level AI policy documents was used for text mining and variable identification, whereas five representative policy documents were selected for detailed PMC evaluation. The evaluation period for these representative policy samples corresponds to China’s 14th Five-Year Plan period. This period was chosen because it represents a comparatively mature and policy-intensive stage in the recent development of China’s AI governance framework, during which AI policy became more systematically linked to national strategies of digital transformation, industrial upgrading, and governance modernization. The specific information on the five selected policies is presented in Table 5.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Selected AI policy samples in China.

https://doi.org/10.1371/journal.pone.0337024.t005

The selection of evaluation objects was based on the following considerations.

First, policy level. This study focuses on AI policies issued at the national level of China. Accordingly, all five selected cases are central-level policy documents, ensuring consistency in governance hierarchy and institutional authority.

Second, policy issuance time. The selected documents were all issued during the 14th Five-Year Plan period (2021–2025), which ensures temporal continuity and reflects the most recent stage of central-level AI policy development.

Third, policy type and structural comparability. The selected documents mainly take the form of opinions, action plans, and implementation-oriented policy measures aimed at promoting, regulating, or coordinating AI development. Their relatively comparable policy structure makes them suitable for PMC-based evaluation of internal consistency.

Fourth, representativeness. The selected policies differ in issuing authority, policy emphasis, and governance orientation, thereby enabling a comparative analysis of how structural consistency varies across different types of central-level AI policy arrangements within the same broader policy cycle.

Construction of the multi-input–output table

Based on the PMC indicator system established in Section 3, the multi-input–output table for the selected policy cases was constructed using nine primary variables and forty-three secondary variables. The purpose of this step was to translate the indicator framework into a standardized coding structure for subsequent PMC calculation.

More specifically, the coding process was carried out on the basis of the predefined binary scoring matrix presented in the revised Methods section and the corresponding Supporting Information. The identification of each secondary variable was informed by the policy texts themselves, together with the analytical dimensions derived from the preceding text-mining stage and relevant policy-evaluation literature. In this process, the text-mining results served as an empirical reference for identifying salient policy dimensions, whereas the final assignment of variable values relied on structured researcher judgment according to explicit coding criteria.

For each selected policy, a value of 1 was assigned when the corresponding policy element was explicitly present in the text, and 0 otherwise. This procedure transformed qualitative policy content into a structured input format that could be used for PMC evaluation. On this basis, the assigned values of the secondary variables for the five selected AI policies were organized into the multi-input–output table shown in Table 6, which provided the direct basis for subsequent PMC Index calculation and visualization.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Multi-input–output table for the five selected AI policies in China.

https://doi.org/10.1371/journal.pone.0337024.t006

Calculation of the PMC index

Based on the procedures described above, and following the establishment of the multi-input–output table, the PMC Index and corresponding mean values for each policy were calculated according to Equations (3) and (4). Subsequently, each policy was classified into its respective level based on the policy classification standards defined earlier. The detailed results are presented in Table 7.

Download:

• PNG
  larger image
• TIFF
  original image

Table 7. PMC index of the five selected AI Policies in China.

https://doi.org/10.1371/journal.pone.0337024.t007

Accordingly, differences in PMC scores should not be understood merely as numerical ranking differences, but as reflecting variation in the completeness and internal coordination of policy design. In this study, a higher PMC score indicates that the policy is configured more comprehensively and coherently across the selected dimensions, including policy nature, timeliness, content, incentive measures, target actors, and level of influence. By contrast, a lower PMC score suggests that the policy structure is relatively less balanced or less fully developed, even if the policy remains valid and relevant in practice.

In practical terms, relatively strong structural consistency refers to a policy arrangement in which strategic orientation, temporal planning, substantive content, implementation support, and governance targets are aligned in a comparatively coherent and mutually reinforcing manner. Relatively weak structural consistency, by contrast, refers to a policy structure in which one or more of these dimensions are underdeveloped, weakly connected, or insufficiently supported, thereby reducing the overall internal coherence of the policy.

This interpretive framework helps explain the differences among the five representative policies. For example, P5 received the highest PMC score because it shows a comparatively complete and coordinated structure across multiple dimensions, especially in policy timeliness, content coverage, incentive measures, target actors, and level of influence. By contrast, lower-scoring policies such as P4 reveal a more uneven structural configuration, with relatively limited performance in dimensions such as timeliness, content completeness, and influence scope. Therefore, the PMC results should be interpreted not only as quantitative rankings, but also as indicators of the relative structural completeness and coordination of AI policy design.

PMC surface and radar chart visualization.

The PMC surface provides a three-dimensional representation of the structural profile of each policy, allowing differences across the nine primary indicators to be visualized more intuitively. To generate the PMC surface, the normalized scores of the nine primary variables were first arranged into a 3 × 3 PMC matrix for each representative policy sample (see Table 8). Based on these matrices, the PMC surfaces were then plotted for P1–P5 (see Figs 6–10). In the resulting surface graphs, elevated areas indicate relatively stronger dimensions of policy design, whereas concave areas indicate dimensions with comparatively lower scores. This visualization makes it possible to identify not only overall policy quality, but also the internal distribution of strengths and weaknesses across policy dimensions.

Download:

• PNG
  larger image
• TIFF
  original image

Table 8. PMC matrices of the five selected AI policies in China.

https://doi.org/10.1371/journal.pone.0337024.t008

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. PMC surface of policy P1.

https://doi.org/10.1371/journal.pone.0337024.g006

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. PMC surface of policy P2.

https://doi.org/10.1371/journal.pone.0337024.g007

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. PMC surface of policy P3.

https://doi.org/10.1371/journal.pone.0337024.g008

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. PMC surface of policy P4.

https://doi.org/10.1371/journal.pone.0337024.g009

Download:

• PNG
  larger image
• TIFF
  original image

Fig 10. PMC surface of policy P5.

https://doi.org/10.1371/journal.pone.0337024.g010

The PMC surface visualization in this study was implemented in Python (version 3.8) using the Matplotlib and SciPy libraries. Specifically, the 3 × 3 matrices of primary-variable scores were smoothed through interpolation prior to three-dimensional plotting, so that the resulting surfaces would more clearly display structural variation across dimensions. In this representation, the x- and y-axes indicate matrix coordinates, while the z-axis represents the normalized scores of the PMC primary variables.

To complement the PMC surfaces, radar charts were also generated in order to display each policy’s performance across the nine primary indicators (X₁–X₉)in a two-dimensional comparative form. Using the normalized scores of the primary indicators as radial values, the profiles of P1–P5 were plotted sequentially to form nine-dimensional contours. In addition, the sample mean curve (shown as a dashed line) was included as a baseline for comparison, making it easier to observe the relative dispersion and concentration of policy performance across dimensions. Compared with the PMC surface, the radar chart provides a more concise representation of cross-policy differences in structural configuration. The results are presented in Fig 11.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 11. Radar chart of China’s artificial intelligence policies.

https://doi.org/10.1371/journal.pone.0337024.g011

Empirical results and analysis

Analysis of the distribution patterns of andicators in China’s AI policies.

To compare the differences in the performance of various primary indicators among the sample policies and to examine potential outlier samples, this study plotted boxplots for the nine primary indicators based on the assigned data of P1–P5 (see Fig 12). Overall, the indicator distribution presents a structural pattern characterized by “strategic and domain stability, timeliness and hierarchy dispersion, and implementation–incentive differentiation.”

Download:

• PNG
  larger image
• TIFF
  original image

Fig 12. Distribution of nine primary indicators across five representative Chinese artificial intelligence policies.

(Note: The black dots in the figure represent outliers in each data group.).

https://doi.org/10.1371/journal.pone.0337024.g012

First, the boxplots for (Policy Nature) and (Policy Domain) are tightly clustered, with small interquartile ranges and overall high scores, and few outliers. This indicates a strong consensus among the sampled policies in terms of strategic orientation and domain coverage. Most policies combine directive, advisory, and normative functions, and systematically encompass key domains such as economy, society, technology, institutions, and environment. This aligns with China’s positioning of artificial intelligence as a core driver within the “new quality productive forces–industrial system upgrading” framework, demonstrating that the top-level design of AI policies has reached a relatively mature stage regarding domain boundaries and functional definitions. Conversely, (Policy Level) shows almost no variance, concentrated at the ministerial or multi-ministerial level (0.33), reflecting that during the sampling period, few overarching central documents were issued. Most policy implementation relied on departmental regulations, explaining the observed insufficiency in cross-level coordination and vertical policy transmission.

Second, (Policy Timeliness) and (Level of Influence) display the highest degree of dispersion. The box for is overall low (mainly short- and medium-term), but includes an upper outlier of 1.0 corresponding to the State Council’s P5, which features long-term planning. This suggests that China’s current AI policies still rely primarily on phased promotion rather than systematic long-term planning. The range for spans 0–1, revealing significant disparities across the “national–regional–industrial–enterprise” transmission chain: P5 and P1 demonstrate strong alignment from national strategies to industrial and enterprise-level implementation, whereas P4, focused on educational scenarios, exhibits significant contraction in this dimension. This indicates that specialized policies lack institutionalized linkage to downstream actors—a finding consistent with the PMC surface and radar chart patterns discussed earlier.

Third, (Policy Content) and (Policy Evaluation) show a “strong overall performance with isolated outliers” structure. The median of is relatively high, but a low-end outlier (0.33, P4) suggests that when policies focus on a single scenario, they tend to exhibit uneven coverage across the full chain of “platform–R&D–data–application–industrialization–risk governance.” remains at a high overall level, reflecting a general trend toward normalization in goal-setting, planning, accountability, and evidence bases. However, the lower outlier (0.25, P1) reveals that even inter-ministerial policies can experience quality reduction if they lack measurable goals, clear accountability lists, or solid evidence foundations.

Fourth, the boxplot for (Incentive Measures) lies in the midrange but features long whiskers, indicating heterogeneity in the combination strength of policy tools such as fiscal subsidies, tax incentives, intellectual property protection, pilot projects, and talent mechanisms. The “non-modular policy toolkit” problem thus persists. Meanwhile, (Target Actors) scores high overall but extends downward to 0.5, showing that while multi-stakeholder inclusion (government, enterprises, academia, and the public) has become the norm, support intensity still varies notably across enterprises and research institutions.

In summary, the boxplot distribution provides three pieces of optimization evidence:

1. (1) Regarding the high dispersion of policy timeliness (), a unified mid- to long-term schedule and rolling evaluation mechanism should be established between central and ministerial policies to avoid fragmented policy pacing.
2. (2) Concerning the wide range of influence levels (), regional pilots and industrial platforms should be leveraged to improve the “national–regional–industrial–enterprise” feedback loop, enhancing terminal policy absorption and diffusion capacity.
3. (3) Addressing the outliers in content and evaluation (,) and the differentiation in incentives (), integrated policy packages should be developed that jointly incorporate scenario demonstration, data governance, risk regulation, and talent mechanisms, grounded in verifiable goals, accountability lists, and evidence-based frameworks to build a replicable and traceable evaluation system.

These findings align with China’s transition from intensive policy issuance to institutional stabilization in the AI domain and provide an actionable direction for future robustness tests and sensitivity analyses on larger policy samples.

Discussion

This study used the PMC Index Model to examine the structural consistency of China’s central-level AI policies and yielded three principal findings. First, the policy system displays a relatively stable core orientation centered on technological development, application expansion, and governance support. Second, despite this overall thematic stability, important weaknesses remain in temporal planning and policy hierarchy, especially with regard to long-term policy design and cross-level coordination. Third, marked variation persists in incentive arrangements and influence levels, indicating that the internal configuration of China’s AI policy system remains uneven across different dimensions of implementation.

The first finding reinforces and extends previous studies that have characterized AI policy as shifting from a narrow techno-economic agenda toward a broader governance-oriented framework [5,6,11]. Existing content-based and computational studies have shown that Chinese AI policies place strong emphasis on innovation, development, and security [13–15]. The present study confirms this general pattern, but moves beyond thematic description by showing that these concerns are not merely frequent topics in policy discourse; rather, they constitute a relatively stable structural core within China’s central-level AI policy design. In this sense, our findings support prior observations regarding the developmental and governance-oriented character of China’s AI policy system, while also demonstrating that this orientation is embedded in the internal configuration of policy elements rather than only in rhetorical emphasis.

The second finding suggests that the main limitations of China’s central-level AI policies no longer lie primarily in the absence of policy content, but in the uneven structural integration of policy time horizons, governance hierarchy, and implementation linkage. This result is broadly consistent with earlier studies arguing that many national AI strategies suffer from insufficient implementation detail, limited evaluation mechanisms, and weak operational follow-through [8]. However, the present study adds a more differentiated structural diagnosis. Rather than treating implementation weakness as a general shortcoming, the PMC analysis shows more specifically that temporal planning remains concentrated in the short and medium term, while cross-level coordination is relatively underdeveloped. This indicates that current AI governance arrangements in China are not primarily constrained by thematic insufficiency, but by the incomplete institutionalization of long-range planning and vertical policy integration.

A third finding concerns the heterogeneity of incentive measures and influence levels. This result resonates with previous PMC-based policy studies showing that incentive mechanisms and coordination structures often constitute the most uneven dimensions of policy design [19–24]. In the case of China’s AI policy system, the present findings suggest that policy support is distributed asymmetrically across fiscal, tax, talent, service, and demonstration instruments, and that the transmission of policy influence across national, regional, industrial, and enterprise levels remains structurally uneven. This point also complicates existing accounts of AI governance that emphasize strategic ambition or policy intensity alone. A policy system may appear highly active or expansive at the discursive level, yet still exhibit structural imbalance in the ways support tools and governance influence are actually configured.

At the same time, a brief note on robustness is warranted. The PMC results reported in this study are inevitably shaped by both the selected indicator system and the sampling strategy used for representative policy evaluation. On the one hand, the nine primary variables and forty-three secondary variables were not arbitrarily defined, but developed through the combination of prior policy-evaluation studies, corpus-based findings, and structured researcher interpretation. This reduces, though does not eliminate, the possibility that the findings are overly dependent on a single indicator design. On the other hand, the full corpus of 54 policy documents was used for text mining and variable identification, whereas the detailed PMC evaluation was conducted on five representative policies. This means that alternative case-selection strategies or the inclusion of additional policies could lead to some variation in specific scores or rankings.Nevertheless, the main findings of the study remain analytically meaningful. The present discussion of robustness is intended as a conceptual qualification rather than an empirical sensitivity test. Accordingly, the findings should be interpreted as structurally informative rather than numerically absolute. Future studies may further examine robustness by testing alternative indicator configurations, different case-selection strategies, or broader policy samples.

Taken together, these findings suggest that the analytical value of AI policy evaluation lies not simply in identifying what policy documents emphasize, but in examining how policy elements are assembled into an internally coordinated governance structure. Compared with prior studies that focused primarily on policy themes, discourse, or instrument types, the present study further shows that the main challenge of China’s central-level AI policy system lies not simply in what policies emphasize, but in how policy elements are structurally coordinated across time horizons, governance levels, and implementation tools. In this sense, the study extends existing AI governance research by shifting the analytical focus from thematic description to structural consistency, thereby offering a more differentiated understanding of how policy coherence is configured within a central-level governance framework.

Implications

To improve the implementability and sustainability of AI policies at the policy level, this study suggests the following optimization directions for future implementation:

First, in terms of time governance, it is recommended to establish a 5–10-year mid- to long-term roadmap, linking phased milestones with budget allocation, pilot evaluation, and performance assessment. Rolling evaluations should replace one-time assessments, providing a stable temporal anchor for technological evolution and industrial diffusion without altering the existing division of responsibilities, thereby reducing fluctuations in policy execution.

Second, regarding hierarchical coordination, it is advisable to institutionalize “regional and industrial platforms” as bridging nodes between the overarching documents issued by the CPC Central Committee and the State Council and the regulations of individual ministries (e.g., national demonstration zones, sectoral public platforms, and pilot city clusters). Joint policy issuance and joint evaluation mechanisms should be promoted. By incorporating platform performance into the annual assessment of relevant ministries and local governments, the strength of cross-domain collaboration and vertical policy transmission can be significantly enhanced.

Third, in terms of policy tool configuration, the approach should shift from single fiscal subsidies to an integrated policy toolkit that combines fiscal support, taxation, intellectual property, data governance, scenario demonstration, and talent mechanisms. This toolkit should be implemented through a tiered diffusion path of “prototype verification – small-scale demonstration – large-scale deployment.” At the same time, clear rules for data compliance, ethical review, exemption boundaries, and exit mechanisms should be established to reduce uncertainty for enterprises.

Fourth, regarding content and evaluation standards, it is necessary to set “minimum coverage requirements” for specialized policies to ensure the inclusion of verifiable checkpoints across six key areas: platform, R&D, data, application, industrialization, and risk governance. Evaluation should be centered on verifiable objectives, accountability lists, and evidence bases, with standardized ex-ante, mid-term, and ex-post evaluation frequencies and methodologies. Furthermore, third-party evaluations and open data interfaces should be introduced to ensure that evaluation results are replicable and comparable across policies.