OIT - How and when will AI impact the economy - Evidence from China

 ILO Brief 1 How and when will AI impact the economy: Evidence from China  Research Brief 2025 How and when will AI impact the economy: Evidence from China1 Wensheng Peng, Zipeng Zhou, Duwen Xiong, Ekkehard Ernst† and Huilin Zhu†

 This brief contributes to the current literature on AI exposure of jobs by considering cost curves of AI execution to estimate the timeline of AI adoption in China until 2050 using industry reports and sectoral employment and wage data from the National Bureau of Statistics of China.  To analyse the economic impact of AI, it breaks down job tasks into a set of well -defined, complementary, and combinable set of basic tasks, categorized by intelligence types.  For each basic task, it compares the current wage level with the expected evolution of the costs of AI to carry out each task to assess when it is economically profitable to replace a basic task with a machine.  The study reveals that basic tasks with lower AI execution costs, such as strength, auditory, visual, numerical, and text information processing tasks, are likely to be replaced first. Meanwhile, tasks such as solution generation, management, and car e and nursing are less likely to be substituted in the short term, and humans will predominantly engage in jobs involving these basic tasks.  This approach provides policymakers, researchers, and business managers with a quantitative framework to estimate the economic impacts of AI, deepening the understanding of AI advancements and offering references for economic policies, academic research, and business strategies.  Introduction The recent breakthroughs, represented by transformer technologies, have sparked a new wave of enthusiasm for AI , prompting renewed interest in its economic

1 This research brief is based on Chapter 3, "Basic -Tasks and AI Economic Impact Analysis," from the report collection "AI Economics" by China International

 Introduction The recent breakthroughs, represented by transformer technologies, have sparked a new wave of enthusiasm for AI , prompting renewed interest in its economic

1 This research brief is based on Chapter 3, "Basic -Tasks and AI Economic Impact Analysis," from the report collection "AI Economics" by China International Capital Corporation (CICC) Global Institute. The original Chinese report is available at: https://cgi.cicc.com/zh_CN/2024 -ai. Published June 28, 2024. The authors would like to thank Caroline Fredrickson and Sher Verick for helpful comments. CICC Global Institute. Main authors of the initial Chinese version of the report. † ILO, Research Department. Editing, contributors to the methodology and data improvement. implications. Currently, popular estimation methods evaluate AI’s economic impact from a task -based perspective. Autor (2003) was among the first to use this approach to measure the economic impact of computer technology. Acemoglu (2024) applied this task -based analysis framework, to estimate the impact of AI on productivity. By measuring AI’s impact on task Key points ILO Brief 2 How and when will AI impact the economy: Evidence from China productivity, it is possible to estimate AI’s influence on occupational and overall economic productivity. Similar methodologies are employed by institutions like Goldman Sachs (2023) and McKinsey (2023a), which break down production activities into job tasks and business functions to evaluate AI’s impact on production efficiency. One challenge in task -based analysis is identifying an appropriate task classification for studying AI’s economic impact. Traditional automation studies focus on specific tasks within particular scenarios, but researchers could face thousands of tasks, wit h new tasks emerging and old ones disappearing, creating complexity in quantitative analysis. Alternative approaches have been used to identify tasks at the occupational level, including the use of specialized surveys such as ONET 2 (Felten et al., 2018), World

tasks emerging and old ones disappearing, creating complexity in quantitative analysis. Alternative approaches have been used to identify tasks at the occupational level, including the use of specialized surveys such as ONET 2 (Felten et al., 2018), World Bank’s STEP survey 3 (Carbonero et al., 2023) or the description available within labour force surveys (Berg et al., 2023). Although the task categories, such as those available in ONET provide a useful classification, they remain incomplete — working activities may overlap each other. For example, differentiating between "Working with Computers" and "Processing Information" c an sometimes be challenging. Consequently, if we attempt to estimate each activity's contribution to job output, we may encounter difficulties. Hence, for quantitative analysis, developing a classification approach that reduces complexity while still enabling clear estimation of each task's contribution is essential. Transformer-based AI technologies exhibit general capabilities in task execution, fundamentally differing from prior computerization and automation. Unlike traditional systems that require programming for each scenario-specific task, modern AI can effectiv ely

2 A database that provides information on occupations in the US. perform tasks, even those not included in the training data. Once AI develops intelligence in a specific area, it can achieve human -level performance across diverse scenarios. Therefore, it is only necessary to classify tasks according to the intelligence of AI, instead of classifying every task in each scen ario separately as before, which is also the biggest difference between our classification methods and those in other studies. Given AI's nature as a data -driven technology, its level of intelligence naturally depends on the volume of training data available in specific domains. For instance, abundant data typically exists in language, numerical information, video, and audio processing, whereas data for movement and physical interactions are relatively limited yet highly feasible for future collection. Conversely, data related to emotional processes and interpersonal relationships developed through longtraining data available in specific domains. For instance, abundant data typically exists in language, numerical information, video, and audio processing, whereas data for movement and physical interactions are relatively limited yet highly feasible for future collection. Conversely, data related to emotional processes and interpersonal relationships developed through longterm interactions may be exc eedingly scarce and challenging to gather. Therefore, AI tasks can be classified based on the varying difficulty levels of information collection and processing across domains. Moreover, to delineate clearly defined task boundaries that facilitate the esti mation of their contributions to output, we have adopted a top -down classification method. This method accurately describes how AI and humans divide labour at different levels of intelligence. It simplifies task classification, making the estimation of AI’s economic impact more feasible. The results also help governments, businesses, and individuals better understand AI’s progress and provide valuable references for economic policies, academic research, and business strategies. The remainder of this brief outlines the key elements of this method and presents estimation results using the data from China as an example. Given the importance of China in the globa l 3 Skills towards Employment and Productivity, a database collected by the World Bank that provides information on skills by occupations in a selected number of developing countries. ILO Brief 3 How and when will AI impact the economy: Evidence from China market for AI investment, the estimates presented in this brief are likely to be relevant for other countries as well (Ernst and Saurabh, 2021).  Methodology: Framework and Data Taxonomy of Basic-Tasks Building on the framework by Fernandez -Macias and Bisello (2022), our taxonomy categorizes task elements based on categories of intelligence, which can be further divided according to the type of information to be processed and the complexity of its handli ng (see figure 1). At the highest level, when an agent performs a task independently, all inputs and outputs of

based on categories of intelligence, which can be further divided according to the type of information to be processed and the complexity of its handli ng (see figure 1). At the highest level, when an agent performs a task independently, all inputs and outputs of information occur solely through the agent itself. However, when social interactions are involved, communication of information between multiple agents may occur. Accordingly, work tasks can be divided into two primary categories: Individual tasks and social tasks. Individual tasks refer to tasks that can be completed independently without interacting with others, while social tasks are those that require interaction. Individual tasks can be further divided into physical tasks and intellectual tasks. This concept aligns with classifications found in the psychological literature, where intelligence can be categorized into concrete intelligence, abstract intelligence, and social intelligence, respectively. These different forms of intelligence correspond to humans’ ability to understand and manipulate p hysical objects, comprehend and operate linguistic and mathematical symbols, and engage in social interactions (Thorndike, 1920). Physical tasks can be further divided based on spatial relationships into strength, dexterity, and spatial navigation tasks. Intellectual tasks can be subdivided , based on the type of thinking required , information processing tasks (fast thinking) and problem -solving tasks (slow thinking). Based on whether the information being processed is encoded, information processing tasks are further categorized into four types: Visual, auditory, text, and numerical information. Problem -solving tasks are further divided into two types: Problem exploration and evaluation, and solution generation and execution . These can be divided further into four categories: Information collection and retrieval tasks, conceptualization and abstraction tasks, solution generation, and planning and execution. For social tasks, based on the complexity of the type of interpersonal interaction, they can be divided into dominant tasks and supportive tasks, which are further categorized into five types: Sales , persuasion and induction; management and supervision; service and

generation, and planning and execution. For social tasks, based on the complexity of the type of interpersonal interaction, they can be divided into dominant tasks and supportive tasks, which are further categorized into five types: Sales , persuasion and induction; management and supervision; service and reception; teaching, training and tutoring; and care and nursing. Splitting basic tasks according to types of intelligence may not be the only approach – other classification methods have the potential to do this as well – but what is important is the idea of classifying the tasks. Its advantage is evident: this classification ensures that the basic tasks do not overlap in functionality. From an economic perspective, we have eliminated the substitutive relationships among the 16 basic tasks while retaining their complementarity, thus forming a set of basis vec tors that ca n combine to represent the entire task space for work. This is significant for understanding how AI integrates into human production activities and its impact on employment. ILO Brief 4 How and when will AI impact the economy: Evidence from China

 Table 1: Basic tasks by type of intelligence Dimension Basic-tasks Explanation AI tools Individual tasks Physical tasks Fixed site Strength Refers to tasks that are accomplished purely by muscular force, e.g., lifting heavy objects, swinging a hammer, etc. Robots or Machines driven by general AI. Dexterity Involves precise movements of the hands or fingers, using the hands and arms to manipulate materials or objects, and, as technology advances, also includes the manipulation of various types of equipment to accomplish tasks with materials and objects, such as welding, carving, spinning, weaving, cooking, cleaning, etc. Robots or Machines driven by general AI, especially humanoid robots. Position movement Spatial navigation Includes moving objects or one's position in unstructured or changing space and finding appropriate routes in the presence of other objects or buildings, e.g., walking, driving, transporting

Robots or Machines driven by general AI, especially humanoid robots. Position movement Spatial navigation Includes moving objects or one's position in unstructured or changing space and finding appropriate routes in the presence of other objects or buildings, e.g., walking, driving, transporting objects, etc. Robots or vehicles driven by autopilot. Intellectual tasks Information processing Noncoded information Processing visual information Acquiring and understanding noncoded information through vision, and making judgements and outputs, e.g., face recognition, judging object types, judging image and video content, tracing images, drawing images, etc. Tools like OpenCV, DALL·E and Midjourney. Processing auditory information Acquiring and understanding non-codified information through hearing, and making judgements and outputs, e.g., determining the source and type of sound, imitating vocalizations, etc. Tools like OpenAI's Whisper, ElevenLabs and 15.ai. Encoded information Processing text information Processing of verbal and written information, generally comprehending the content of the input information or outputting it according to the given content requirements, e.g., basic reading, writing, listening, or reading aloud Tools like ChatGPT. Processing numerical information Be able to use mathematical tools such as existing formulas, theorems, algorithms, or statistical methods to compute, record, or present numbers, e.g., prepare graphs and charts of numbers, use mathematical formulas to compute data, use statistical tools to determine trends and distributions of data, etc. Tools like IBM Watson Analytics. Problem solving Problem exploration and evaluation Information collection and retrieval Finding, locating and accessing information needed to make decisions or solve problems (mainly refers to the thinking process), e.g., questionnaires, searching databases, reviewing

Problem solving Problem exploration and evaluation Information collection and retrieval Finding, locating and accessing information needed to make decisions or solve problems (mainly refers to the thinking process), e.g., questionnaires, searching databases, reviewing literature, retrieving audio and video, etc. Techniques related to slow thinking, such as Chain-ofThought. OpenAI's Deep Research could be a prototype. Conceptualization and abstraction Use of acquired information, combined with the ability to conceptualize, learn, and abstract, to make assessments based on certain norms and values, e.g., problem categorization, attribution analysis, etc. Solution generation and execution Solution generation Generate new ideas and methods with a rational structure and model based on the problem to be solved, e.g., artistic design, new product ideas, marketing strategies, etc. Techniques related to specific expertise and slow-thinking. Producing deliverable results without process monitoring. AI agent could be a prototype. Planning and execution Involves the steps required for an organization to effectively implement a solution to a problem, with a high degree of logic and big picture thinking, e.g., developing a strategic plan, writing an academic article, proving a theorem or law, producing an engineering guide, planning an itinerary, etc. Social tasks Dominant tasks Sales, persuasion, and induction Inducing others to do something or buy goods or services, as well as various types of negotiations and debates, such as: selling products or services, negotiating contract terms, persuading customers to pay, persuading stakeholders to cooperate, conducting seminars, etc. Requiring the establishment of long-term relationships and more complex emotional interactions, which may be difficult for AI to achieve. Management and supervision

customers to pay, persuading stakeholders to cooperate, conducting seminars, etc. Requiring the establishment of long-term relationships and more complex emotional interactions, which may be difficult for AI to achieve. Management and supervision Managing or supervising the behaviour of others, e.g., monitoring project progress, assigning tasks to team members, reviewing employee performance and providing feedback, coordinating inter-team or inter-individual cooperation, resolving employee conflicts, etc. Supportive tasks Service and reception Responding to or fulfilling requests from customers, leaders, and coworkers, e.g., answering customer inquiries, ordering and serving food, handling customer complaints, guiding tour groups, reporting on work, etc. The requirements for interpersonal relationship building and emotional interaction are weaker than the dominant task. Tools like customer service chatbots, Duolingo and Character.ai could be a prototype. Teaching, training, and tutoring Transferring knowledge or helping others to improve their skills, e.g., job skills training, coaching team members to improve sales skills, general education, vocational training, etc. Care and nursing Providing personal assistance, medical care, emotional support, etc., to others, such as co-workers, customers, or patients, to meet the welfare needs of others, e.g., assisting an elderly customer with purchasing goods, ensuring the health of a patient, organizing a charity or philanthropic event, etc.

Data source: Fernández-Macías & Bisello (2022), CICC Global Institute ILO Brief 5

How and when will AI impact the economy: Evidence from China Quantitative Analysis Framework The total human working time equates to the time spent completing various basic tasks. The improvement in current AI capabilities means AI can perform an increasing number of basic -tasks at human -equivalent levels. Therefore, knowing how many basic tasks AI can

Quantitative Analysis Framework The total human working time equates to the time spent completing various basic tasks. The improvement in current AI capabilities means AI can perform an increasing number of basic -tasks at human -equivalent levels. Therefore, knowing how many basic tasks AI can execute allows us to estimate the economic and social implications of AI on labour demand and on the structure of production across geographies, sectors and occupations. Similar to previous general -purpose technologies (Jovanovic et al., 2005), AI influences the economy through three steps: R&D, absorption, and growth. Under the basic task framework, these steps can be described as follows: First, AI R&D departments advance technology to expand the range of executable basic-tasks while reducing AI execution costs. Second, when the cost of using AI to perform a particular basic task falls significantly below human labour costs, businesses adopt AI. Finally, as AI costs conti nue to decline, cheaper AI solutions are deployed to execute more basic tasks. This allows investment in more production factors under constant total costs, resulting in higher outputs and economic growth. The quantitative model is based on profit -maximizing firms, using AI cost changes for each basic -task to estimate AI’s annual economic impact. The basic tasks, in complementary form, constitute labour as a factor of production, which combines with capital to generate

4 For each industry, we use the CES function to estimate the relationship between basic-tasks and total labour factor inputs, where the substitution elasticity coefficient between basic-tasks is set to 0.5, according to Acemoglu (2024). For capital and labour factors, we use the Cobb-Douglas function to estimate the relationship between labour factors, capital factors, and industry output. Considering the possibility of “Baumol's cost disease”, f or the long-term estimation of industry output to GDP, we use the CES function to estimate the long-term relationship between industry output and GDP from the perspective of the aggregate demand function, where the elasticity of substitution between

output. Considering the possibility of “Baumol's cost disease”, f or the long-term estimation of industry output to GDP, we use the CES function to estimate the long-term relationship between industry output and GDP from the perspective of the aggregate demand function, where the elasticity of substitution between industries is set to 0.7 (Nordhaus, 2008). industry outputs. The outputs of various industries collectively form GDP (Nordhaus, 2008; Acemoglu, 2024).4 The decision on whether a basic task should be executed by AI or human labour depends on the costs. Data For the estimation model, data on value -added, employment, and capital inputs, was obtained from yearbooks or statistical agencies’ websites. Two key datasets that critically affect estimation results are: (1) The future cost trends of AI for each basic ta sk, and (2) the distribution of basic tasks across industries.5 Regarding the first dataset, we referenced AI forecast published by institutions such as Stanford HAI (2024) and the World Economic Forum (2023), as well as a survey by Grace et al. (2024), which collected predictions from thousands of AI experts about the future development of AI technology. Using these references, we estimated the approximate number of years until which AI will achieve technical feasibility for each basic tas k. 6 Based on trends in AI costs, computational power, and energy prices, we projected the costs for all technically feasible basic tasks up to 2050.7 While AI may achieve significantly higher output per unit of time, we evaluate AI costs based on the equivalent cost for AI to complete the same workload that a human worker would accomplish. Furthermore, AI may encounter the "Solow Productivity Paradox" whereby productivity gains only show up much later 5 As Acemoglu (2024) noted, obtaining precise estimates for these datasets is impossible, but we endeavor to produce robust approximations. 6 Based on the reported performance trends of different AI tools or

AI may encounter the "Solow Productivity Paradox" whereby productivity gains only show up much later 5 As Acemoglu (2024) noted, obtaining precise estimates for these datasets is impossible, but we endeavor to produce robust approximations. 6 Based on the reported performance trends of different AI tools or applications, we match them with the basic-tasks to estimate roughly the years that are technically feasible. 7 For physical tasks, which involve physical interaction, costs were primarily benchmarked against humanoid robots, while intellectual and social tasks were benchmarked against AI software services. Sources from Epoch AI (2024), Artificial Analysis (2024), Ark -invest (2019), McKinsey (2023b), Mecademic (2024) and Change Discussion (2024). See endnote for detailed citation information. ILO Brief 6 How and when will AI impact the economy: Evidence from China than when new technologies are being introduced. We estimate that a lag of approximately 2 –4 years could occur between the successful research and development of AI technology and its full deployment. Additionally, AI deployment may require complementary e xpenditures, such as initial investments in specialized equipment (e.g., servers), as well as subsequent costs associated with maintenance, personnel training, data processing, and application adaptation. In our estimation of AI basic -task cost curves, we have incorporated deployment costs and deployment time lags to calculate realistic cost curves across different basic tasks at the point of actual deployment (see Figure 1).  Figure 1: Expected cost curves for AI performing various basic-tasks in deployment

Note: The figure shows relative costs, that is, considering one hour of human workload and converting it into the cost for AI to perform that workload. The cost curves are based on global expert estimates, with values converted into RMB per hour.

Data source: Epoch AI (2024), Artificial Analysis (2024), Ark-invest (2019), McKinsey (2023b), Mecademic (2024), Change Discussion (2024), CICC Global Institute

on global expert estimates, with values converted into RMB per hour.

Data source: Epoch AI (2024), Artificial Analysis (2024), Ark-invest (2019), McKinsey (2023b), Mecademic (2024), Change Discussion (2024), CICC Global Institute

Basic tasks represent a new classification framework, hence no existing statistical data is available. Therefore, to construct our second dataset, we used a survey method to obtain industry-level data for China. The survey was designed in a scoring format, with a total of 100 points distributed according to the proportion of time spent on each basic task within a given industry’s total working hours. The questionnaire was sent to industry analysts at CICC, covering hundreds of listed companies across 19 industries. Table 2 illustrates the distribution of the time required to complete basic tasks across industries, providing insights into the production characteristics of each sector. 8

8 The industry classification in this article uses the major industry categories in the "Industrial Classification for National Economic Activities (GB/T 4754-2017)". For ease of presentation and understanding, some industry names have been adjusted or Each number represents the share of total working hours spent on each of the basic tasks, adding up to the total hours work in each industry horizontally. Note that attribution across basic tasks is based on its predominance in each occupation. For instance, in health care, only 6.5 per cent of time is spent on actual care as defined in table

1. This has to do with the fact that actual nursing staff represents only between a third and two fifth of total employment and often spend a considerable amount of their time on non -care related activities such as planning, or compliance. simplified. The manufacturing industry is split into light manufacturing, resource processing, and equipment manufacturing.

0 50 100 150 200

employment and often spend a considerable amount of their time on non -care related activities such as planning, or compliance. simplified. The manufacturing industry is split into light manufacturing, resource processing, and equipment manufacturing. 0 50 100 150 200 2024 2026 2028 2030 2032 2034 2036 2038 2040 2042 2044 2046 2048 2050 Strength Dexterity Spatial navigation Processing visual information Processing auditory information Processing text information Processing numerical information Information collection and retrieval Teaching, training, and tutoring (Rmb/hour) ILO Brief 7 How and when will AI impact the economy: Evidence from China  Table 2: Time share of basic-tasks in production activities of various industries Industry Strength Numerical information processing Visual information processing Dexterity Spatial navigation Auditory information processing Text information processing Information collection and retrieval Teaching, training, and tutoring Service and reception Planning and execution Conceptualization and abstraction Solution generation Sales, persuasion, and induction Management and supervision Care and nursing Financial 1.7 3.8 3.1 5.8 7.3 0.7 3.9 4.1 9.2 9.9 5.3 5.7 7.6 17.4 9.4 5.1

Information services 2.4 6.3 6.1 8.5 4.6 1.5 6.1 4.9 4.1 12.8 5.9 3.7 8.7 10.8 11.3 2.4 Mining 22.0 2.5 2.6 20.3 14.6 0.5 1.2 2.7 1.7 2.8 3.0 2.4 2.5 11.6 9.0 0.7 Real estate 2.4 6.0 4.0 8.0 2.4 4.0 6.0 6.0 4.0 7.4 9.4 6.0 9.4 7.0 10.0 8.0 Leasing and business services 1.1 4.9 5.1 5.1 4.2 4.7 4.9 4.9 5.2 9.9 5.2 5.0 5.0 19.9 10.3 4.9 Water, electricity, heating, and gas supply 3.1 9.4 4.1 11.6 7.5 0.6 8.8 6.1 6.0 6.8 5.3 6.7 5.6 8.4 5.9 4.2 Medical and health 2.2 6.5 6.5 6.5 5.4 1.6 6.5 7.6 7.6 8.6 7.6 5.4 6.5 7.6 7.6 6.5 Resource processing 24.9 8.4 5.5 14.9 5.8 1.4 3.4 2.7 3.2 3.7 4.3 1.6 2.3 11.2 5.5 1.3

Cultural, sports, and entertainment 5.5 2.8 2.8 6.7 8.4 2.1 3.8 2.5 7.1 17.6 2.8 2.8 2.2 17.4 10.1 5.7 Education 3.0 5.0 4.0 10.0 7.0 1.0 5.0 6.0 15.0 5.0 4.0 7.0 8.0 7.0 8.0 5.0 Equipment manufacturing 5.3 6.5 3.2 12.8 6.6 1.5 4.8 5.4 5.7 9.3 7.1 2.7 6.0 10.9 7.5 4.8 Transportation and logistics 14.2 0.1 2.3 1.4 61.8 2.0 1.5 0.5 1.1 1.1 4.8 0.3 0.2 2.3 5.4 1.0 Construction 21.7 0.9 0.6 14.0 0.4 0.6 0.9 0.9 13.4 1.1 14.2 0.9 5.6 1.0 22.8 1.2 Agriculture, forestry, animal husbandry, and fishery 27.5 0.6 1.9 45.1 12.1 0.9 0.8 0.6 3.8 1.1 0.7 0.8 0.7 1.0 1.2 0.9

husbandry, and fishery 27.5 0.6 1.9 45.1 12.1 0.9 0.8 0.6 3.8 1.1 0.7 0.8 0.7 1.0 1.2 0.9 Wholesale and retail 8.7 1.7 1.1 3.6 6.4 0.8 1.3 1.3 3.6 12.8 2.4 1.8 1.0 40.8 10.0 2.8 Residential services 12.5 0.2 1.6 21.1 3.8 0.5 0.4 0.9 5.1 27.1 5.5 1.9 1.5 12.3 4.0 1.6 Light manufacturing 8.3 1.5 2.0 28.8 7.5 1.0 2.8 2.0 2.4 7.2 2.7 2.5 4.0 16.6 8.1 2.4 Water conservancy and environmental protection 3.6 5.8 3.9 21.6 9.4 0.6 7.7 4.3 6.4 3.5 3.3 10.2 5.0 4.8 6.7 3.2 Accommodation and catering 6.7 0.5 1.1 10.0 5.6 2.1 1.1 1.1 5.0 37.7 1.2 1.1 3.0 13.9 5.0 5.0

Note: The industries on the left side of the chart are arranged from high to low according to the average hourly labour cost in 20239, and the basic-tasks on the top of the chart are ranked by the year of technically feasibility from the earliest to the latest. The data units in the chart are percentage s, which are estimated based on the actual situation in

2024.

Data source: CICC Global Institute, CICC Research Department, National Bureau of Statistics of Chin a Findings and possible applications Based on t