Guidelines for the Deployment and Application of Large Artificial Intelligence Models in the Government Affairs Field

This guideline is formulated to thoroughly implement the decisions and plans of the Central Committee of the Communist Party of China and The State Council, regulate and guide the development and application of AI large models in the government field, and enhance the level of digital and intelligent governance and services in government affairs. This guideline mainly provides work guidance and basic reference for the deployment and application of AI large models for government departments at all levels. It will be dynamically adjusted in accordance with the progress of practice and in light of the new situation and new requirements for the development and application of AI large models.

I. General Requirements

Guided by Xi Jinping Thought on Socialism with Chinese Characteristics in a New Era, all relevant parties shall further implement the spirit of the 20th CPC National Congress and the Second and Third Plenary Sessions of the 20th CPC Central Committee, thoroughly implement General Secretary Xi Jinping's important thought on building China into a cyber powerhouse, implement the new development philosophy in a complete, accurate and comprehensive manner, ensure both high-quality development and greater security, adhere to systematic planning and intensive development, uphold a people-centered approach and standardize applications, promote joint construction and sharing and efficient collaboration, ensure safety and stability and deliver solid outcomes, advance in an orderly manner the deployment, application, and continuous optimization of large AI models in the field of government affairs, maximize the advantages of large AI models in complex semantic understanding and reasoning, multimodal content generation, and knowledge integration and analysis, among others, provide efficient assistance to staff members, provide facilitation services for the public and enterprises, promote the innovative development of government affairs, and enhance governance effectiveness, optimize service management, and assist in scientific decision-making.

II. Application Scenarios

Government affairs departments may explore the application of large AI models in selected typical scenarios based on local conditions and in light of the actual circumstances by centering on the common and high-frequency needs of government services, social governance, office work of relevant organs, assisted decision-making, and other work. The main reference scenarios are as follows:

1. Government services

(1) Intelligent Q&A. Data such as business resources and knowledge bases of local regions and corresponding departments and fields are integrated, and natural language understanding, retrieval-augmented generation, knowledge graphs, and other technologies are utilized to provide facilitative online government affairs consultation services, enhance accurate understanding of public demands, generate reference answers in real time, assist in addressing public doubts, and improve the facilitation of obtaining information.

(2) Assisted handling. Government service guidelines, FAQs, user evaluation, historical handling records, and other data are integrated, and technologies such as intelligent matching and automated processing are utilized to provide one-stop assisted government affairs handling services such as intelligent handling guidance, personalized guidance, pre-completion of forms, assisted review, and progress inquiry and reminders, assist staff members in efficiently examining materials, and support the facilitative handling of matters by the public and enterprises.

(3) Direct access to and quick enjoyment of policy services. Policy service knowledge bases are constructed; policy requirements, policy labels, push conditions, application and redemption processes, and other related content are specified; and algorithmic models enabling “policies to reach and benefit corresponding people and enterprises” are utilized to strengthen the analysis of public and enterprises' demands, achieve intelligent policy matching, and advance the proactive and precise delivery and one-stop handling of services that benefit both the public and enterprises.

2. Social governance

(4) Intelligent monitoring and patrol inspections. Drones, video surveillance, smart sensors, and other equipment and computer vision and other technologies are utilized to conduct real-time analysis of surveillance video, images, and Internet of Things (“IoT”) perception data, among others, assist staff members in conducting real-time monitoring of infrastructure such as buildings, roads, gas, bridges, water supply, drainage, heating, and integrated pipeline corridors, discover abnormal behavior, environmental issues, or facility failures, among others, in a timely manner, automatically identify potential risks and hidden dangers, give reminders in a timely manner, and provide disposal suggestions based on the circumstances and severity of anomalies to improve the efficiency of monitoring and patrol inspections.

(5) Assistance in law enforcement and regulation. Technologies such as speech recognition, video analysis, knowledge graphs, and logical reasoning are adopted to assist law enforcement personnel in entering case information into the system in real time, identifying in a penetrating manner clues to problems, generating case reports, rapidly searching legal bases and judicial interpretations, and querying about similar representative cases, among others, provide pertinent suggestions on case handling, and improve the efficiency and standardization of law enforcement and regulation.

(6) Market risk prediction. Using generative time series analysis models and anomaly detection algorithms, etc., monitor and conduct in-depth analysis of various market data to capture market trends, including fluctuations in economic indicators, anomalies, etc., predict potential market risks, assess the impact on the economy and society and economic trends, and issue early warnings in a timely manner to support government management and social governance.

3. Office in government agencies

(7) Assist in document drafting. Leverage the generative power of large language models by building local knowledge bases and preset templates to provide staff with writing suggestions, assist in drafting documents, check, proofread and optimize formats and contents, improve productivity, and reduce the burden on the grassroots level.

(8) Data retrieval. By using technologies such as knowledge graph construction and information retrieval, accurately understand the data retrieval needs of staff, achieve rapid retrieval, precise positioning, multi-dimensional sorting, intelligent association and comparative analysis of government affairs information, and help staff improve the efficiency and accuracy of data retrieval.

(9) Intelligent distribution. Using technologies such as natural language understanding and multimodal recognition, construct multi-dimensional task classification and assignment rules to automatically identify, accurately classify, assist in filling, and prioritize tasks such as incoming documents, incoming calls, and work orders, to achieve assisted distribution and intelligent assignment, and improve task assignment efficiency.

4. Decision-making assistance category

(10) Disaster warning. Big data correlation and comprehensive analysis and judgment of multi-source, multi-dimensional, multi-modal data from satellites, ground sensors, geological monitoring stations, as well as forecast and warning, disaster risk census, etc., identify abnormal fluctuations, predict possible natural disasters, issue early warnings in advance to assist government departments in taking effective measures in a timely manner to mitigate disaster risks and reduce disaster losses.

(11) Emergency response. Use technologies such as reinforcement learning to analyze and assess the nature, characteristics, extent of harm, scope of impact, development trends, and public response of public security emergencies, identify and warn of potential risks in a timely manner, quickly simulate the effects of emergency response plans based on emergency scenarios, distribution of forces and resources, provide scientific and reasonable emergency response suggestions, optimize the allocation of rescue resources, improve the speed and efficiency of emergency response.

(12) Policy Evaluation. Use the inference and analysis capabilities and data mining capabilities of large AI models to analyze public feedback, market response, economic indicators, and social satisfaction, construct multi-dimensional indicators to assess the extent of policy goal achievement, policy impact, and potential problems, and support policy-making departments in policy optimization.

(13) Intelligent-assisted review. By leveraging capabilities such as self-learning generalized cognition, humanized review reasoning, and multimodal intelligent analysis, conduct project reviews in accordance with relevant requirements, conduct in-depth scanning and intelligent analysis of project document contents, provide review suggestions and opinions, and assist in improving the efficiency and scientificity of project reviews.

III. Standardized deployment

Government departments should, in light of the actual work and the characteristics of the scenarios, fully demonstrate the application requirements, implementation paths, functional designs, etc. of large AI models, select appropriate deployment models, promote the implementation in a coordinated manner, promote co-construction and sharing, and enhance the efficiency of construction and management.

1. Reasonably select the implementation path

Government departments should carefully select the implementation path of large artificial intelligence models based on the requirements of different government scenarios and the existing technological foundation. For scenarios with strong universality and rich data resources such as intelligent question answering and assisting in document drafting, model products and services that are mature in the market and have been filed with the cyberspace administration should be adopted. For specialized scenarios with complex business logic such as assisting law enforcement supervision and market risk prediction, domain expert knowledge and professional data can be utilized for targeted training to create vertical models. Under the premise of ensuring security and not disclosing state secrets, work secrets and sensitive information, make full use of Internet computing power and model resources to deploy and apply large artificial intelligence models in the government affairs field. Encourage the exploration of innovative applications such as government intelligence agents and embodied intelligence.

2. Deploy in a coordinated and intensive manner

Government departments should carry out the deployment of large artificial intelligence models in the government field in a coordinated and intensive manner, relying on “East Data West Computing” and the national integrated computing power network, promote the layout of intelligent computing power infrastructure in a coordinated manner, and implement centralized and unified security management and systematic technical protection measures to avoid “fragmented” security risks. Where conditions permit, central and state organs and provinces (autonomous regions and municipalities directly under the Central Government) may uniformly deploy intelligent computing power resources and AI large models to provide AI large model services in the e-government extranet environment to subordinate units or regions under their jurisdiction. Prefectures and cities should carry out deployment and application in accordance with the unified requirements of provinces (autonomous regions and municipalities directly under the Central Government), and county-level and below should, in principle, reuse the intelligent computing power and model resources of their superiors for application and service, and no longer independently build and deploy government large models.

3. Explore the realization of unified management and reuse

Government departments should explore the construction of an intensive deployment model of “building in one place and using in multiple places and departments”, and promote the deployment and application of large government models in a coordinated manner to prevent the formation of “model islands”. Provinces (autonomous regions, municipalities directly under the Central Government) should build a unified service platform for artificial intelligence large models in the field of government affairs, and integrate it with the government affairs cloud management platform, government affairs application and component management platform, etc. They should incorporate resources such as the intelligent computing power of the e-government extranet, government affairs large models, and government affairs data sets within the region into unified management to form a “one account” of element resources to support the operation monitoring of government affairs large models. Provide resource application and scheduling services to promote efficient reuse. National industry authorities explore unified training and construction of large vertical models of government affairs in specific fields in accordance with business requirements and development needs, strengthen coordinated deployment with provinces (autonomous regions and municipalities directly under the Central Government), and deepen intelligent empowerment across levels and regions in industry fields. Vertical administrative departments should strengthen the overall deployment and management of resources such as models, computing power and data to avoid waste of resources.

4. Continuously strengthen the data foundation

Government departments should enhance the governance of government data, continuously improve data quality, and accelerate the construction of high-quality government data sets and knowledge bases that objectively reflect public policies, institutional norms, business processes, and governance effectiveness, to support the optimized training of large government models. Classify and grade the data involved in government large models, strengthen the management of training data, fine-tuning data, knowledge bases, etc., establish ledgers and record in detail information such as data sources, types and scales to ensure reliable and traceable data sources and accurate and effective content. Relying on the government data sharing coordination mechanism, we will coordinate the results of data governance and promote the co-construction and sharing of high-quality government data sets and the collection and governance of generated data. Explore the governance path of government knowledge based on large models, build a trusted knowledge base, and ensure the authority, accuracy and timeliness of data sources.

IV. Operation and Management

Government departments should strengthen the operation and management of artificial intelligence large models in the government field, improve management systems, operation models and security requirements, and promote the deployment and application of artificial intelligence large model technologies, products and services in the government field in an orderly manner.

1. Clearly define the requirements for application management

Government departments should coordinate burden reduction and empowerment, strictly implement relevant requirements such as “Several Provisions on Rectifying Formalism to Reduce Burden on the Grassroots” and “Several Opinions on Preventing and Controlling ‘Formalism on fingertips’”, avoid blind pursuit of technological leadership and conceptual innovation, avoid repetitive and ineffective construction, avoid construction before approval and construction without management, avoid forced use and ineffective use, avoid multiple data collection and repeated requests, and effectively prevent “digital formalism”. The deployment and application of large artificial intelligence models in the government affairs of central and state organs should be incorporated into the overall planning of national government informatization. Government departments should establish and improve a full-cycle management system covering the deployment and application of AI large models in the government field, clarify the application methods and boundaries, implement the “auxiliary” positioning of AI large models, and promptly solve new problems that arise during the deployment and application process. Risk warnings should be placed prominently on the application interface of the government large model, clearly informing of the limitations of the large model’s services and marking the output content of the large model. For AI large model application scenarios that represent government departments in providing services to the public and enterprises, such as intelligent question answering and assisted processing, the content review system process should be strictly implemented, and measures such as manual review, real-time risk control of generated content, and cross-validation of multiple models should be reasonably adopted in combination with the characteristics of the scenarios and technical capabilities to prevent risks such as model “illusion”, Ensure that the output content does not exceed the business scope, guarantee the accuracy of the content, and maintain the credibility of government departments.

2. Continuously promote iterative optimization

Government departments should consider continuous iterative optimization as a key link in the deployment and application of large AI models, establish a regular update mechanism, accelerate functional optimization, and deepen scenario application. Keep a close eye on technological developments and continuously update and optimize the basic models and security capabilities of AI large models in the government field. Establish an efficient data collection and processing mechanism, update the input data and knowledge base that support the operation of the AI large model in a timely manner, and clean, label, supplement and optimize the training data set in a timely manner to continuously enhance the model’s capabilities. Establish a user evaluation and feedback mechanism for AI large models in the government affairs field, collect and process user demands in a timely manner, and drive iterative optimization with user feedback.

3. Do a solid job in security management

Government departments should establish a security responsibility system, clearly define the security responsibilities and tasks of the participants at each stage of data processing, large model training and scenario application, and do a good job in user identification and permission management. When providing artificial intelligence large model services, government departments should abide by relevant provisions such as the Interim Measures for the Administration of Generative Artificial Intelligence Services, use data and basic models with legal sources, fulfill obligations such as algorithm filing and security assessment in accordance with the law, and sign service agreements with users to clarify the rights and obligations of both parties. Establish a classified and graded governance system for AI large models in the government sector, improve security management processes, and formulate emergency response plans for potential security risks. Do a good job in detecting and dealing with adversarial attacks on government large models, and identify and intercept prompt injection, resource consumption attacks, etc. Strengthen the security management of government affairs large model content by integrating semantic recognition, rule bases, model algorithms, etc., to identify, analyze and control multimodal input and output content, establish reasonable response and rejection mechanisms, and promptly discover and handle illegal and bad information, sensitive content, etc. Give full play to the content review advantages of news media, do a good job in content review and control of the training data of the government affairs large model, and strengthen the content monitoring and management of the government affairs large model. Manage the application operation logs of government large models well and conduct regular audits of log records. Promote the formation of a security risk threat information sharing and emergency response mechanism, handle and report security incidents in a timely manner as required, and enhance the ability of artificial intelligence to respond to security risks.

4. Strictly enforce confidentiality requirements

Government departments should strengthen data security and confidentiality and personal information protection during model training, deployment and application, adhere to bottom-line thinking, strictly implement confidentiality discipline requirements such as “no access to the Internet for classified information and no access to classified information on the Internet”, and take measures such as installing confidentiality “guardrails” to prevent state secrets, work secrets and sensitive information from being input into non-classified artificial intelligence large models, Guard against the risk of information leakage caused by the aggregation and association of sensitive data. Establish and improve confidentiality management systems related to the application of AI large models in government affairs, and standardize confidentiality management throughout the entire process of AI large model selection, deployment, training, use, and decommissioning. The application of artificial intelligence large models in classified information systems shall be steadily advanced in accordance with the requirements of the state administrative department for confidentiality.

V. Safeguard measures

1. Strengthen organization and implementation

Strengthen overall coordination and steadily and orderly promote the standardized application of large artificial intelligence models in areas such as government services, social governance, government office work, and decision-making assistance. We will accelerate the establishment of a national standard system and the development of key standards for AI large models in government affairs, clarify work norms such as application effect assessment, system technical requirements, and intelligent technology application, and support the deployment and application to achieve practical results. Summarize and promote typical scenarios and innovative applications of the deployment and application of AI large models in the government field in a timely manner to promote reuse and efficiency improvement. Strengthen financial support for the deployment and application of AI large models in government affairs, introduce a market-based competition mechanism for products and services, explore an operational model where enterprises build and operate, the government purchases services, and charges are settled based on usage, and create an efficient and sustainable ecosystem for government affairs large models.

2. Conduct monitoring and evaluation

Build a full-process monitoring and evaluation system for the deployment and application of large artificial intelligence models in government affairs, and conduct monitoring and evaluation work in due course. Establish a security assessment mechanism for government large models, conduct thorough testing and verification of model algorithms, generated content, application functions, configuration environment, attached data, vulnerability risks, etc. before going live, and rectify and reinforce the identified problems and potential risks. Strengthen real-time monitoring and analysis of the operational status, response time, accuracy, security and potential risks of artificial intelligence large model systems in the government field, promptly identify problems and take effective measures to solve them. Evaluate the performance of large AI models, summarize experiences in a timely manner, continuously iterate and optimize, and promote the deployment and application to achieve practical results.

3. Training and publicity

Develop a training curriculum system covering the theory, technology, application, security, ethics, industry and other aspects of artificial intelligence large models, carry out artificial intelligence literacy and skills training, enhance the cognitive level of leading cadres on artificial intelligence, and improve the application ability and level of staff. Educate the public to enhance digital literacy for all, actively respond to user concerns, correctly guide society’s understanding and expectations of the applicable population, scenarios, and uses of AI large models in the government field, and objectively reflect the role of AI large models in optimizing government services, meeting the needs of the public and enterprises, and improving the level of social governance.