Understanding the Extent of Automation and Process Transparency Appropriate for Public Services: AI in Chinese Local Governments

Many countries are exploring the potential of artificial intelligence (AI) to improve their operations and services, and China is no exception. However, not all AI techniques or automation approaches are suitable for every government service or process since transparency and accountability are paramount in the public sector. In this context, automation via expert systems (ES) is still a vital complement or even an alternative to AI techniques, because they can be more easily audited for potential biases. This paper analyzes the smart examination and approval (SEA) process use in China and explores how different forms of automation could be better options for certain services or specific processes within services, considering their level of transparency as an important characteristic. Based on these results, the authors argue that governments could consider hybrid approaches combining, for example, machine learning, for verification processes, and ES, which are more easily auditable, to make final decisions on individual cases. They also propose a classification of services by considering the extent of automation and process transparency needed. The classification considers a hybrid approach such as SEA, but also include other alternatives such as the exclusive use of AI techniques, as well as traditional online delivery and face-to-face procedures.


INTRodUCTIoN
In the age of artificial intelligence (AI), governments around the world are exploring new techniques and applications to deliver services at lower cost and with better quality to citizens. There are many types of AI techniques and applications, each with its own advantages and disadvantages. In China, the rapidly spreading initiative "Smart Examination and Approval" (SEA) is an interesting in the public sector, and which techniques are better suited for certain government services or specific tasks, stages, or processes within a single service. This is very important because systems used by governments are increasingly using some form of ADM, including AI technologies and applications, and public managers do not have enough knowledge to make decisions about IT investments and what specific characteristics of these systems are the most appropriate for certain types of services.
The paper is organized into seven sections, including the foregoing introduction. Section 2 includes a short literature review about AI and ADM in the public sector. Section 3 describes our research design and the methods used for this study. Section 4 presents our main results, and Section 5 describes some of the most important characteristic of SEA, which include the use of machinelearning-based AI and expert-systems-based AI for different steps of the service-delivery process. Section 6 provides some implications for research and practice, presents some limitations of this study, and suggest a few ideas for future research. Finally, Section 7 provides some concluding remarks.

LITERATURE REVIEw: ARTIFICIAL INTELLIGENCE ANd AUToMATEd dECISIoN-MAKING
Many scholars have analyzed the principles, applications, and results of government agencies' use of AI for service delivery. Wirtz et al. (2019) suggested ten AI application areas: AI-Based Knowledge Management (KM) Software, AI Process Automation, Virtual Agents, Predictive Analytics and Data Visualization, Identity Analytics, Cognitive Robotics and Autonomous Systems, Recommendation Systems, Intelligent Digital Assistants (IDA), Speech Analytics, and Cognitive Security Analytics and Threat Intelligence (Wirtz et al., 2019). In terms of Wirtz's categories, SEA in China fits within AI Process Automation Systems. However, this label is not a widely used term in this field. According to the literature, studies of automated services provision in government can be included in the field of ADM (Kraft, 1964).

what is Automated decision-Making?
The concept of ADM was created 59 years ago when Ivor Kraft critiqued Thomas A. Cowan's paper about decision theory. Cowan indicate that "If now we reflect that conscious decision making always, or at least usually, involves the making of inferences, we see the part that the computer plays in the process of decision making" (Cowan, 1963). So, in essence, Automated Decision Making means to make decision using a computer. Kraft disagreed with Cowan in many aspects. But Kraft's viewpoint that "it has long been the impression of many of us that at least 90 percent of what is traditionally considered the 'creative' employments of physicians, judges, lawyers, editors, is no less sheer hackwork than the physical exertions of day laborers" (Kraft, 1964) is still radical nowadays and seems to predict the coming of the age of AI. Even though computer technology has undergone many disruptive changes in the past 60 years, the definition of Automated Decision Making is still simple: to make decision using software without human intervention.
Before being connected with AI and included in GDPR (the General Data Protection Regulation of the European Union; see GDPR, 2018), ADM was widely used in many fields, such as medical care, aerospace, and optometry (Clayton et al., 1988;Huang, 1990;Madsen et al., 1993), for dozens of years. ADM is sure to gain tremendous attention when it is adopted in public administration, and it will inevitably be questioned by professionals and experts from many disciplines because it will significantly influence the masses. As one scholar notes, When ADM is used in the service of public administration, the objective is to produce a decision that involves the exercise of public law in a manner that defines, for an individual or for a private legal entity, a particular right, duty or benefit on the basis of material legislation (Suksi, 2020).

Two Types of Automated decision-Makings in Public Services
When automation in government was first implemented, most cases referred to office automation, the computerization of office routines, or the use of an electronic card instead of a paper voucher. As technology has continued to develop, both expert systems and machine learning have been adapted for use in government services.

ADM Based on Rules: Expert Systems
Rule-based ES were the most common application of ADM for a long time, until machine-learning began to capture public attention. ES are a branch of applied AI and were developed by the AI community in the mid-1960s (Liao, 2005). They are designed and implemented to: aid in government document reference services (Harley & Knobloch, 1991), advise on the causes of sanitary problems in public water sources (Swayne et al., 1992), evaluate energy use in public buildings (Gatton, 1995), help supervisors determine appropriate disciplinary actions for employee offenses (Berry et al., 1998), create public transportation dispatching plans (Yin & Peng, 2000), distribute legislation to industries online (De Meyer & Lavrysen, 2004), automatically classify public investments (Gutiérrez Vanegas, 2017), and conduct efficient and effective public accounting (Cooper et al., 2019). From these cases, we find that, in the area of government services, ES is useful to deliver expertise and support decision-making when a task is too difficult to be handled by the staff of the relevant organization.

ADM Based on Machine Learning
Machine learning is an information-processing technique concerned with deriving concepts from data. A well-explored machine-learning approach is learning from examples, also referred to as inductive machine learning. In this approach, examples of problem situations are submitted to a learning system that induces a general description of the underlying concepts that are useful for problem solving (Filipič & Junkar, 2000). Machine learning is implemented in ADMs to: detect bursts and other abnormal flows in water resources (Mounce et al., 2010), spot motorcyclists who are not wearing a helmet (Silva et al., 2012), automatically detect and sanction speed-limit violations (Hamelin, 2010), analyze text reports automatically (Ku & Leroy, 2014), and automatically recognize handwritten text and numbers and complete emotion classifications (Al-Mushayt, 2019). From these studies, we could have a basic understanding of machine learning, including its suitability for some steps in providing services to citizens and businesses, such as identity recognition of human beings and material things, and identifying the relative behaviors of certain subjects.

Potential Biases when Using Automated decision-Making in Public Services
Recent AI studies show that algorithm bias can come from several aspects of the real world. Some emerge from popular preferences, such as AI virtual assistants (AI VAs), which reflect characteristics of white femininity in voice and cultural configurations (Moran, 2020). Others come from the intrinsic limitations of the algorithm (Varona et al., 2020) or the quality of the algorithm, including a deterioration of accuracy in the presence of algorithm errors, which was found for example in facial recognition applications (Howard et al., 2020). Another set of issues emerges from the data, which is essential for the prediction result of an algorithm. These diverse challenges include missing data, unidentified individual data, insufficient sample size, and underestimation, misclassification, and measurement error, which can cause bias (Gianfrancesco, 2018). Such bias could cause serious issues, especially when machine learning is used in sensitive applications, such as those in the health care industry, where it may perpetuate or even exacerbate racial and ethnic disparities (Obermeyer, 2019;Noseworthy et al., 2020). In contrast, some research argues that AI may help overcome human biases by providing information and decision support based on quality data and appropriate algorithms (Noriega, 2020).
The most common concern about ADM involves algorithm bias, which frequently occurs with the AI technique of machine learning. Suksi (2020) states that "there should be a provision that prevents the use of machine-learning ADM and requires the use of rule-based ADM in administrative decision-making." According to the GDPR (2018), individuals should have the right not to be subject to a decision based solely on automated processing, including profiling. To prevent algorithm bias, diverse approaches have been adopted in many countries, such as the United Arab Emirates (Halaweh, 2018) and Canada (McKelvey & MacDonald, 2019).
Even though ADM is frequently based on ES and the decision rules are relatively clear, there are still some limitations and challenges. For example, the use of ADM in the Canadian immigration system has caused a significant backlash and some of the most controversial practices have been stopped. In Poland, the practice of profiling that divided unemployed people into three categories earned significant backlash and was officially ended (Kuziemski & Misuraca, 2020). However, the key point is not that the ADM systems might be violating citizens' rights, but that the algorithms applied in these systems do not fully comply with the applicable laws and requirements. Wihlborg et al. (2016) discussed the role of professional officers in the ADM program used by the Swedish National Board of Student Aid and found that legitimacy was the most important characteristic for all actions. ADM based on ES rely on decision rules that follow the existing laws and requirements for specific government programs and services. Therefore, it is also valid to consider that, if a result is not fair to a group of people, the cause might not be the data or the algorithms, but the policies and other rules that are coded by programmers into the systems (Eubanks, 2018).

RESEARCH dESIGN ANd METHodS
This paper explores the use of SEA in China. We collected data from commercial search engines for analysis. The following section briefly describes our overall research design, the strategies for data collection, and the data analysis performed. The study follows a methodology based on library and information science. There are three principles guiding the data collection: (1) whether the data about SEA news is online; (2) whether the search engine and crawler software can be used to collect the data; (3) whether the news can be correctly coded for further analysis.The first principle is related to the motivation of local government officers when applying new technology. Li and Zhou (2005) find that the likelihood of promotion of provincial leaders increases when their economic performance is good, while the likelihood of termination decreases (Li & Zhou, 2005). This logic could be applicable to the use of information technologies in government services. So, if they spend a significant amount of money in SEA applications, they will also have news published online through all kinds of media. In addition, mass media are also very interested in government initiatives. Therefore, it is reasonable to obtain the necessary information from Internet-based news sources.
In relation to the second guiding principle, we have chosen the top search engine in China and the best crawler software in such a way that online news about SEA can be effectively collected. The most popular search engine in China is Baidu.com, which has a market share of 71.13% (Statcounter, 2020). In addition, the professional software Octoparse (Octopus Data, Inc., 2020) has full ability for internet data collection. Therefore, Baidu and Octoparse are very good choices to achieve an effective data collection of online news. In terms of the third principle, there is no need to code the result for further analysis because the data will provide the necessary evidence. The content of the news will help us identify whether there is any SEA case in a specific province and how many cases are already in use.

data Collection
The data were scraped for two levels, the provincial level, and the city level. As most of the newly updated information is found on the first page of a search result, we only scraped the items on the first page, which include approximately ten results. Because several phrases refer to the initiative of SEA, we selected four keywords for the search: Smart Examination and Approval, Intelligent Examination and Approval, Examination and Approval in a Second, and Service Provision in a Second. We used the keywords together with the name of the province to conduct the search. Some example search terms are: Smart Examination and Approval Beijing, Intelligent Examination and Approval Shanghai, Examination and Approval in a Second Zhejiang, Service Provision in a Second Guangdong.
As there are 34 provinces, autonomous region, municipalities, and Special Administrative Regions in China, we needed to search 136 times for the result. The searching and scraping was carried out in August, 2020. There were 1,360 results collected. As there are 346 cites, including municipalities, in China and some have an abbreviated name, we searched 2,768 times to cover the four statements about SEA and the two possible names for each city (full name and abbreviated name). This round of searching and scraping was carried out in September 2020. Ultimately, 27,618 results were obtained. For each search result, information was collected from six fields: (1) title, (2) URL (the website address), (3) abstract, (4) date, (5) website name, and (6) keywords.

data Analysis
As the basis of our analysis, we wanted to learn whether government agencies in a certain region have implemented any SEA, and how many cases exist in each region. Specifically, at the provincial level, we collected the number of SEA cases in one province or municipality; at the city level, we simply judged whether there are any SEA cases or not, because there was not enough information available publicly to obtain an accurate count. Moreover, the amount of data in the provincial level is much smaller than that of the city level, so the analysis methods are different for each level of government.
For the provincial level, the amount of data is small and there are about 40 search results for each province or municipality. Two steps were applied: (1) read the title and abstract fields of the search results, one by one, to make a judgment about whether the news accurately describes the situation in a targeted region; (2) locate the URL to gather information about SEA use in a certain region. If a specific search result was not related to a specific region, the second step was not executed. Most results were news stories about the most popular and typical SEA cases throughout the country.
For the city level, the amount of data is large and there are approximately 80 search results for each city. Because there are many duplicated results in the data, the first step was to strike them. However, we faced the problem of how to define these duplicated results. To do so, we used a lowrisk algorithm that recognized items as duplicates only when the title field and the abstract field were the same. After removing the duplicates, we read the abstracts and titles to judge whether SEA had been applied in the targeted city. Third, if the information was insufficient, we visited the website to access the content on the webpage and reach a final conclusion.
In the process of collecting provincial information, we encountered another potential problem, which involved the relationship between provinces and their cities. If there were some SEA cases in one city, the online search results would find the news report successfully. However, within each news story, there was no guarantee that the name of the province to which the city belongs would appear. Moreover, even if the province name was included, it might not always be easily visible, or it might have not been included in the TDK tags (title, description, keywords) for the webpage, which makes it less likely to be grabbed and registered by search engines. Thus, it is possible that the progress of one city in a province was not found by the search engine and identified as related to that province.
To solve the problem, we modified the analysis of the provinces according to the data collected at the city level. This required us to obtain the number of SEA cases at the city level when needed, which increased the workload, but also the effectiveness of the search strategy.

Brief description of the Case: Smart Examination and Approval in China
The first SEA initiative in China was implemented in Guangxi Province in October 2017 (Tong & Zhou, 2017). It is described as a professional, efficient, and powerful "approval expert" that does not require the work of examination and approval to be done by people. This not only removes the 8-hour work limit, but also allows people to "be approved in seconds without visiting physical agencies, or even without submitting materials." This "approval expert" works as part of the 24-hour "non-closing" online government where an AI enhanced system automatically checks the submitted information and makes a decision about whether or not the applicant meets the requirements. Even when the service includes pre-approval items or requires information sharing from upper-level government agencies, the system can realize automatic data comparison, automatic approval, automatic generation of certificates, and the immediate display of the results.
SEA provides hope that administrators can break through the "ceiling" in government services provision. The rise of AI has made it possible to overcome the limitations of the traditional manualapproval model. After reducing and decentralizing examination and approval items, shortening the time limitation for processing, reducing examination and approval materials, optimizing the work process, and implementing electronic supervision, the reform of administrative examination and approval in government services has been significantly improved. The existing online service model in most government agencies is only an extension of the traditional manual approval model on the Internet, replacing offline processes with online ones. Whether a service is examined and approved online or in person, manual participation is required. As long as the examination and approval processes go through different approvers, there will be different interpretations of the same case by the different actors. Thus, there may be different approval results and potentially space for rent-seeking.
For example, in Shenzhen City, the most innovative city in China, the SEA features of "online application, automatic approval, instant results, and dynamic supervision" have taken on a more attractive conception: "Approval within Seconds," which means "zero physical spot visits, zero queuing, no meeting, and full automation" for applicants. On November 15, 2018, the General Office of the Shenzhen Municipal Government issued a specific policy, "Work Plan for Promoting the Approval within Seconds Model in Shenzhen," to promote the initiative widely and systematically introduce it into many services. Shenzhen is the benchmark of SEA use in China because it is the first city that published some policy to widely promote this type of initiative.

ANALySIS ANd RESULTS
This section presents the main results of our analysis. First, we provide an overview of the use of SEA by province and by region. Second, we present the results for a typical service for which SEA has been used frequently and extensively across China: registering a company. Our goal is to provide enough background information for the reader to better understand our argument in section 5, which is that combining machine learning and expert systems could be a good strategy to provide government services.

overall Use of Smart Examination and Approval in China
The results of this part of the analysis are presented in Table 1. A brief study was done in April 2019, and we found that only 14 provinces had applied some SEA cases. As of now (September 5, 2020), only three provinces have not utilized this procedure. Thus, the initiative is growing very rapidly. Table 1 shows the most recently updated number of SEA cases. For provinces in which Date is still 2019 such as Shanxi, there were no news related to SEA published online in 2020 and, therefore, the initiative might be facing some issues. In contrast, for provinces such as Zhejiang, news keeps emerging together with the fast spreading of SEA. Although the news was collected in September 2020, the period studied includes about one year and a half, from March 2019 to September 2020.
Among all the SEA cases, two are very popular. One is to register a company, and the other is to register an individual or household business, which includes family members only. In 15 provinces, the applicants can register a company within a few seconds and obtain the license on the condition that there is no pre-approval certificate. Similarly, in seven provinces, the applicants can register a household business and obtain the license within a few seconds.  District, 2020). SEA use at the city level is not as interesting. There are 129 out of 348 cities in which at least one SEA case has been applied; the municipalities are so large that it is easy to accomplish this. The five provinces with the greatest SEA use are Shandong, Zhejiang, Guangxi, Jiangsu, and Anhui. All, except Anhui, are also among the top five provinces in terms of average GDP, which is expected give the initial investment and expertise needed to implement SEA applications.

Analysis of a Typical Smart Examination and Approval Case: Registering a Company
There are 15 provinces out of 34 in which SEAs are applied in the service of registering a company. The first case was implemented in Qingdao City of Shandong Province on August 13, 2019. The most recent report was in Baoding City in the Hebei Province on August 24, 2020. This service is typical not only because of its popularity, but also because it requires the delivery of physical materials (i.e., business license).
The first case was described in the news as follows:

On the morning of August 13, 2019, the country's first intelligent registration system to establish an enterprise was put into operation in Qianwan Bonded Port of Qingdao City. The system can automatically examine and approve enterprises' establishment using software and it can completely replace the traditional manual-approval method. Thus, it can greatly reduce the time required for enterprise establishment by its immediate approval.
Entrepreneurs can choose to "get it right away" in the service hall or apply at home and have the business license mail to them. There is no appointment, no queuing, no need to submit paper materials, and no staff needed to review the materials item by item. "It is so convenient that I can register or close the business independently anytime and anywhere," Gao Lianxu said.

The system was produced through the cooperation of the Qingdao Qianwan Bonded Port Zone
Market Supervision Bureau and the Qingdao Administrative Approval Service Bureau. The system reduces the three examination and approval steps of background manual pre-examination, acceptance of the application, and approval of the registration into one automatic audit process to achieve the most simplified approval process with "zero manual intervention." Thus, the whole process can be done automatically and connect with the full electronic registration system for enterprises in Shandong Province. The automatic recognition of the company's registered address and the menu-style completion of the business's scope are the two major difficulties of such intelligent registrations. (Bai, 2019) The webpage for registering a company in Qingdao City also includes some explanations about the process.

The platform uses technologies such as a data center, facial recognition, electronic signatures, electronic business licenses, and other technologies to realize the "one network connection, one form filling, data exchanging, and automatic smart approval." Companies can obtain business licenses, seal engraving, tax invoices, social insurance participation, medical insurance participation, open provident fund accounts, open bank accounts, and other services on the platform (Start a business, n.d.).
From the descriptions above, we find that the two branches of AI have been used in this SEA case. First, machine learning is applied in facial recognition and electronic signatures; both fall under the field of image recognition. Then, ES is used to examine the form-filling and approval-making. Finally, the business license can be delivered through express services, or it can be printed using a specialized printer in the office. This combination is very interesting because it provides efficiency, but also good levels of transparency and legitimacy, since expert system can be easily audited in terms of the specific rules they use to make decisions.

CoMBINING MACHINE LEARNING ANd EXPERT SySTEMS To PRoVIdE GoVERNMENT SERVICES: THE LoGIC BEHINd SMART EXAMINATIoN ANd APPRoVAL
In China, few argue when SEA is applied in the provision of a growing number of government services. The news about SEA is mostly positive, and it is common sense that SEA is an innovation that will benefit citizens. The rationality of SEA in China lies in the suitable usage of AI technologies, which can avoid or at least ameliorate algorithm bias. As we have shown, the fundamental aim of SEA is to write computer programs capable of making decisions with a quality similar to human decisionmaking. There are two main components of this process (Fig. 1). The first is the input module, and the second is the judgment module. SEA is able to turn the traditional analog method into one that is fully digital.

Use of Machine Learning for Recognition and Verification
To implement ADM, the materials need to be digitized and structured to be machine readable. The procedure involves two main components: who is applying for the service, and what are they submitting? Identity verification is frequently the very first step of any government service delivery; the government agency needs to verify the identity of the applicant to avoid the possibility of fraud. In traditional service delivery, a government employee in China would verify a customer's identity by carefully comparing the photo on the ID card to the face of the applicant (in person) to tell whether or not it is the same individual. In contrast, when using SEA, the facial recognition program can observe the characteristics of the face of an individual and compare them with the photo stored in the service system to verify the applicant's identity. Two conditions allow the success of identity recognition. First, there is a national population information database in China that stores data on 1.399 billion people, including 13 data items, and it can be utilized in the provision of digital public services (Liu & Cheng2017). Second, the AI algorithm for facial recognition is accurate enough to accomplish the verification process. As an example, some studies show that facial recognition systems in Asia are capable of providing data precisely and timely on the number of people at an exhibition site, as well as their age, gender, and time of stay (Chien et al., 2019); the accuracy rate of this process could be as high as 99% (Zhi & Liu, 2019). Furthermore, the two famous Chinese Internet companies, Alibaba ( Sun, 2018) and Tencent (Xu, 2019), have already supported some local governments in the delivery of mobile services through facial recognition technology.
The second step is to recognize the content of the documents submitted by the applicant. In traditional service delivery, a government employee in China would examine the documents one by one to verify their authenticity. The key point of such examination is to assess the genuineness of the red seal, which is a symbol of the administration. In contrast, when using SEA, the materials will be scanned and recognized by an Optical Character Recognition (OCR) program so that the content will be machine-readable (Feng et al., 2019;Wang et al., 2020) and ready to be utilized as one of the factors considered in the decision-making process.
Then, there could be two possible paths to test the genuineness of these documents. First, the data-matching approach requires that the document submitted when applying for a service is entered as an inquiry to match the existing data in a database and produce a response of either success or failure. Second, it could be the case that no documents are required because all the necessary data about citizens have already been collected and merged to a large data warehouse, or the data from various departments and agencies can be exchanged through an electronic city-wide agreement. Therefore, when the requirement of specific data for a certain service is submitted, the data can be extracted from the central database or exchanged from a database from another government agency automatically. Thus, through a central database or data sharing between government agencies, SEA can help to minimize or even omit many of the required documents.

Use of Expert Systems for decision-Making
Decision-making in the public sector is based on pre-established rules. In traditional service delivery, government employees make decisions according to rules derived from laws and regulations. These rules are clear and formal and frequently exist on paper. This allows public servants to decide whether the applicant has the right to access a certain service or not. In contrast, when using SEA, an ES is built to have all rules coded and embedded in the program. Thus, the system will make a judgment according to the consistency between the rules and the digital materials automatically and almost immediately. Unlike machine learning's black-box approach (Wachter et al., 2017), ES are a sort of "white box" with a high level of transparency in terms of the rules they use, and they enhance the abilities of civil servants by helping them to make rules-based decisions (Suksi, 2020).
When some atypical or very complicated situation occurs, a different method should be used. In traditional service delivery, public servants would probably use their own expertise and previous experience on the issue or communicate with their colleagues, perhaps even reporting to a supervisor to make a suitable judgment. Similarly, when using SEA, special cases or new circumstances are left to human judgment because there are not always specific rules that can be applied to them. Therefore, it is common for SEA systems to have a mechanism to deal with these unusual situations via human intervention. Because data on all cases are stored in the SEA system, government employees are able to compare existing cases and new cases when dealing with these unusual situations. Furthermore, a strategy could be that new automated rules are derived from unusual, but relatively common, cases.
Additionally, in China there is one published policy that strengthens and standardizes each transaction (The State Council, 2019), and another policy to entitle the citizen to directly evaluate the service (General office of the State Council, 2019). If a SEA program has made a wrong decision in a specific service, these policies could help the applicant to appeal to the supervisory department. For example, in the case of registering a company, the two main challenges are the address standardization and business scope standardization. This is because there are many rules that regulate the buildings and overall requirements for some types of businesses. In fact, one of the most important factors to approve the registration of a company is to assess whether the building of the new company meets all the legal requirements of its specific business type. With data sharing between a real estate registration agency and the Market Supervision and Administration Bureau (the department in charge of the registration), the rules coded in the system can make the assessment quickly and accurately. Moreover, there are some special occasions when some pre-examination-and-approval procedures (Zhao, 2017) are needed, because the rules are too complicated to be coded, or the situation is too rare or unique to fit any regular rules.
In conclusion, an ES is used to deal with services in typical situations when rules are clear and can be precisely coded into a transparent algorithm. When a special or complicated situation arises, human intervention is acknowledged, and manual processing is used. Finally, the mechanisms for evaluating the service help applicants to assert their rights when a mistake is made. Thus, SEA combines machine learning and ES for citizens to complete the services automatically. Will it be possible for all government services to be dealt with using this approach? We argue that the tentative answer to this question is no, not always, as the main requirement for SEA is to have structured data and structured approval rules. Therefore, in the next section, we propose a preliminary typology of services based on how much automation and transparency are needed. We acknowledge that this is a preliminary proposal and do not argue that it is exhaustive, but a starting point for a necessary discussion about the potential of different automated decision-making approaches to different government services.

dISCUSSIoN ANd IMPLICATIoNS FoR THEoRy ANd PRACTICE
Based on our analysis of the SEA structure and logic, it seems clear that some services are more suitable for traditional face-to-face delivery, others could be offered online, but still require some human intervention, and a few could be offered online and fully automated. In our view, the difference is related to whether the necessary input and assessment can be successfully done by an AI algorithm (or another automated decision-making application) and how transparent the decision for the case of a single individual needs to be. In addition, it is important to assess whether the necessary materials and applicable rules can both be represented and entered into the system as structured and machinereadable data. Government services are expected to be provided fairly to all individuals in society, and when the rules are transparent, it is easier for individuals to assess this fairness and understand whether or not biases exist. Therefore, combining machine learning and expert systems could generate services that are more efficient, but also more transparent and, potentially, fair. This ultimately could contribute to making the government (and whoever designs and implements the AI system) more accountable. In addition, experts' systems will always make decisions following the same rules and the same algorithmic decision process, which will ensure the same treatment for any service applicant. In contrast, AI algorithms will learn from the data and the decisions they make every day and, therefore, the rules they use today to make a decision could be different from the rules they will use in two years or in five years for the same decision. Again, this is important in terms of fairness and accountability in the delivery of public services and that is why combining different approaches to automated decision-making could help to obtain some of the advantages and avoid some of the challenges/issues of each of them. Following, we present some implications for theory and practice.

Contribution to Theory: Proposing a Typology of AI Used in Government Services
As our main contribution to theory, we propose a preliminary typology of government services based on the appropriate level of automation and process transparency. It is important to clarify that services categorized in each type could be different in different national realities, since the needs depend on the culture and citizens' willingness to accept certain conditions or limitations for specific services. What is acceptable is relative to the culture of specific countries and localities. Acknowledging this limitation to the generalizability of our initial proposal, we still believe this typology could help to start a conversation about the balance between automation and transparency when providing public services.
First, there are traditional government services at physical offices. These are services that should be provided with human intervention in a face-to-face interaction. In these services, verification and examination can only be completed by human intervention and cannot be executed by a computer program. Driver's license applications are a typical example of these services. After an individual passes the examination on traffic code knowledge and takes the 5-hour safe-driving course, they make an appointment with the local Department of Motor Vehicles (DMV) to take the road test. In the road test, the examiner decides whether or not the individual passes the test and receives their driver's license. These services need to be provided in this way because they require physical presence and activity that cannot be performed digitally, as well as the judgment of that activity by an individual. Face-to-face processes are also necessary, for example, when biometric information for the individual, such as fingerprints, need to be collected by the government.
Second, there are government services that can be provided online, but still require human intervention. The required materials can be submitted through a website, and, after a period of time, the approval decision is made by a government employee. A registered residence application is a typical example of these services. The most difficult application for registering as a permanent resident in China is found in the capital city of Beijing. Applicants complete the online forms and submit many documents. Then they obtain a decision after about three months of verification and analysis of the application. These services need to be provided in this way because the verification of the materials or the examination according to the rules has to be done by a government employee. In this case, there are many requirements to be considered: holding a residence permit, not exceeding the legal retirement age, consecutive payment of social insurance in Beijing for seven years or more, legal and stable employment, legal and stable residence, educational background, work and residence area, innovation and entrepreneurship, honors and awards, obeying the law, and paying taxes, among others. However, all these requirements can be submitted online and do not require the physical presence of the applicant at any moment in the process.
Third, there are government services based on automated decision-making, including AI. These are services that can be done completely online and approved automatically, without direct human intervention. In these services, the required materials can be submitted through a website or an App, and the logic of the examination rules is clear and can be coded into a system so that the approval will be made promptly within a few seconds. In this paper, we have discussed one cases of this type. The reason that these services can be automatically completed by ES is that the data (materials) and the rules are structured to be machine readable. In addition, as discussed before, in this type of service, machine-learning-based AI is not always suitable because it is not transparent enough and the potential biases cannot be adequately assessed. Future research should include additional cases from other national contexts and test the validity of this initial typology. Understanding the benefits and challenges of simultaneously using different approaches to automated decision making in the same public service, similar to what happen with SEA, is also another area for future research about this topic. Finally, assessing the similarities and differences in terms of people's acceptance of AI-based and other automated decision-making tools for different services or stages within a single service could also be a next step in this research agenda.

Implications for Practice
The fact that rules-based ADM is more transparent seems to reduce some of the skepticism and criticism toward applying AI in government and in China, where SEA is spreading at an amazing speed. According to the analysis of the search results from the most popular search engine, Baidu, SEA has been used in 31 out of 34 provinces and municipalities. There are also 129 out of 348 cities with at least one SEA case. The initiative's use in different regions of China varies, and the provinces with better economic conditions have been able to implement SEA for more services faster. Given the improvements in transparency and the potential reduction in corruption, systems like SEA could be important alternatives in many countries around the world, particularly the ones that have experience high corruption indices and lack of trust in government by citizens.
By examining a typical SEA case (registering a company), we have been able to better understand the potential of this approach. A government service could be completely automated by carefully and purposefully combining two types of automated decision-making. In the SEA case, machine learning is used to do identity recognition and material verification, and an expert system is used for specific decisions on individual cases based on clear and transparent pre-established rules, which are applied consistently to everybody requiring a government service. This combination makes SEA more transparent and, although it might still be biased, many of the biases could be identified and corrected, if needed. We argue that having a more transparent black box helps to mitigate the problems derived from algorithm bias and, with this approach, more governments might be willing to try automated decision making in their provision of public services.

Limitations and Future Research
There are a few limitations in this research. First, the proposed typology is based on evidence from a few cases using SEA in China and it is not necessarily applicable to all kinds of public services. Therefore, future research can focus on specific aspects of this typology and the combination of different types of AI-based and other automated decision-making techniques that allow approaches like SEA in China. It could be the case that different combinations of service-delivery modes and different AI techniques are more suitable for certain kinds of services or for specific activities and processes within a complex service. Second, this study did not analyze the variables affecting SEA success in detail. Therefore, understanding the potential benefits, challenges, and unintended consequences of applying AI-based and other automated decision-making techniques in government, particularly SEA in Chinese provinces, could be another avenue for future research. For this, a qualitative approach, including interviews with government officials from Shenzhen City and Hangzhou City, where SEA has grown rapidly, could be a sensible strategy.
Finally, the results of this research could help not only local governments in China, but also local governments throughout the world to better understand the use and potential of AI-based and other automated decision-making techniques for public services. Therefore, as another avenue for future research, researchers could propose similar studies in different countries. Particularly, we hope our study help local government managers to better understand the potential and limitations of ADM systems such as SEA, as well as their similarities and differences when compare with new AI techniques that are opaquer in terms of how they actually make decisions. Future research could also explore how combinations of expert systems and some AI techniques, such as Machine Learning, could produce better results than pure AI-based systems at different levels of government and in countries with very different characteristics.

CoNCLUSIoN
From our review of the current literature, it is clear that the old branch of AI, called rule-based or knowledge-based ES, has been applied for decades in government settings. These practices are less likely to be questioned by society and analysts because, to a certain extent, they can be considered white-box algorithms. This is because it is much easier to understand and assess their way of working (making decisions) based on pre-coded rules. In addition, many of these ES are not designed to make decisions about specific individuals. Rather, they are usually applied to help government employees in broader ways to solve complex and difficult problems by using expertise at a relatively lower cost. In contrast, some new AI techniques are less transparent, and it would be difficult to identify their potential biases (both related to their specific algorithms and/or the data they use). In addition, as mentioned before, AI-based systems learn overtime and the criteria and specific processes they use to make decisions could be different in the present when compared to the past and the future, which could have implications in terms of accountability and fairness in society.
Our study provides evidence about the potential benefits from the combination of expert systems and advanced AI techniques, such as Machine Learning, to provide better services. For example, this combination has the advantage of improving efficiency and effectiveness, but still keeping a good level of transparency in terms of how decisions are actually made. We argue that this is very important in government services, particularly when decisions are directly affecting specific individuals, since transparency is highly valued in the public sector and is expected to lead to better accountability. In addition, in contrast to purely AI-based decision making that can apply certain criteria today and, after learning from the data, apply a different set of criteria in a few years from now, expert systems would always apply the same criteria and, therefore, fulfill, the values of equal treatment and fairness in the provision of government services. Thus, a service provision combining expert systems and AI could obtain some of the benefits of AI, but also preserve some values that are essential in the public sector.
This research was partially funded by the Shanghai Planning Office of Philosophy and Social Science under grant number 2017BTQ004, titled "Research on Evaluation of Government Information Sharing in Government to Citizen Affairs." The views and opinions are those of the authors and do not necessarily reflect the official positions of the Shanghai Planning Office of Philosophy and Social Science.