summary
The potential impact of AI on the job market has sparked heated discussions. Some researchers predict that AI will lead to mass job losses, but others believe that the economic boom brought about by AI will create new jobs. Discussing the ultimate impact of AI inevitably falls into the science fiction imagination, but if the question is clearly defined as the impact of AI on employment in the foreseeable future, such as 10 years, economic analysis can still provide valuable clues.
Although AI has demonstrated the ability to surpass ordinary people in some fields, the existing AI technology route is still difficult to surpass the upper limit of human intelligence, nor does it deny the value of human intelligence and the responsibilities that need to be assumed. Until further disruptive technological breakthroughs, the differences and complementarities between AI and human intelligence mean that there is room for collaboration between the two. Based on China's recruitment big data, we have calculated the extent to which various occupations are affected by AI[1]. The substitution and enhancement effects of AI on labor coexist. Occupations such as office administration, transportation and logistics, and data processing are at high risk of being replaced by AI. Careers such as sales, law, and management are more affected by the augmentation. Based on our estimates, AI could lead to a slowdown in overall job growth over the next decade, but not mass unemployment.
We used recruitment data from China to examine the impact of AI on the wage gap. The results show that in the past few years, the wage growth of occupations with high AI substitution effect has been slower, and the gap between the wages of occupations with low substitution effect has widened. Over the next five years, AI could lead to a slight decline in China's share of labor income. We also analysed the impact of AI on human capital, and the results show that the market value of academic qualifications and professional experience may be adjusted by AI. It is worth noting that the mass production of AI-generated works may reduce the market value and return on human capital, which may limit the development of human intelligence.
In the face of AI-brought changes in the job market, policies need to be two-pronged in both the primary and redistribution phases. The primary distribution should focus on vocational training and labor protection for workers, improve the ability to collaborate with AI, and promote employment and increase labor income with the least possible market distortions. The redistribution of workers in favor of workers ensures the Pareto nature of AI development and promotes social support for technological progress. The source of funds for redistribution should be a highly progressive income tax with less distortion; A "robot tax" may discourage investment in technology and should not be used as a first choice. The UBI, which is currently being hotly discussed in developed countries, may become a redistributive policy option in the AI era, but it is costly. What is more practical for China is to improve the existing social security system, especially to make good use of the role of AI in promoting economic development, and improve the help and protection of vulnerable groups.
body
1. Artificial intelligence and human intelligence towards collaboration
Artificial intelligence has shown capabilities that are close to or exceed the average human level in many fields, but the current artificial intelligence is still very different from human intelligence in terms of basic structure, learning methods and working mechanisms, and the two have their own comparative advantages. In some work tasks, artificial intelligence can be completed independently at a high level, thus replacing human labor. However, AI also plays a role in augmenting human labor in most work tasks, so there is still a lot of room for collaboration between the two for the foreseeable future.
(1) The collaborative space between AI and humans
Artificial intelligence, especially large language models (LLMs), has demonstrated powerful capabilities in many fields such as natural language processing and programming. For example, in numerous professional and academic exams, such as the U.S. Bar Exam and the Graduate Record Exam, AI has shown a level comparable to that of humans [2]. But at the same time, AI is still inseparable from human participation and guidance in practical applications. Even in programming, a recognized area where AI is strong, human-AI collaboration is still important to improve outcomes. In 2024, AI expert Andrew Ng demonstrated how a well-designed human-designed workflow framework can improve the programming accuracy of mainstream large language models[3]: Human work can significantly improve weaker AI models such as GPT-3.5 and make them more advanced than GPT-4 (Figure 4.1). The four techniques used by humans in designing workflows (reflection, tool use, planning, and multi-agent) all reflect the unique human thinking system and way of working, and the collaboration between humans and AI models produces greater results than when either party works alone.
Figure 4.1: Various AI models show a higher level of programming in human-designed AI agent workflow frameworks
Note: HumanEval is a benchmark dataset used to evaluate the code generation capabilities of AI models. It consists of a series of programming problems or tasks that the model solves by generating code. The chart shows the performance of various AI models tested on the HumanEval dataset. The black diamond dot is the score of the AI model in zero-shot prompting, which means that the AI is directly asked to complete a task without giving the AI any examples as a reference, which is a kind of prompting method, and other methods such as few-shot prompting are assigned tasks to the AI model and several examples are given by humans at the same time. The colored dots are the scores after the AI model is placed in a frame of different types of "workflows" arranged by humans. The horizontal axis represents the Pass@1 score, which represents the percentage of problems that the model can correctly solve on the first attempt, with higher scores meaning the model is more accurate and efficient at generating the right code. The chart reflects the model score and ranking as of May 27, 2024.
资料来源：Papers with Code，中金研究院
The training and use of large language models themselves also show that AI is still inseparable from the guidance of human intelligence. Chain-of-Thought Prompting is a large language model prompting method, in which the user proposes tasks to the AI model and also demonstrates the thinking process of humans to solve such problems, and guides the AI to imitate the human way of thinking to give answers. Experiments have shown that the chain of thought prompt method can significantly improve the performance of the model on tasks such as arithmetic, common sense, and symbolic reasoning [4]. The effect of solving the problem depends on the student's knowledge reserve (similar to the number of parameters of a large model), but the quality of the solution idea explained by the teacher also affects the final solution effect.
The mathematical reasoning ability of using synthetic data to train large models is another example of how AI can work with human intelligence to improve its level. Synthetic data not only mimics the training set of the target mathematical task, but also increases the complexity and diversity of the questions and answers by rewriting the questions, self-validation, and reverse reasoning. A research project called AlphaGeometry uses 100 million synthetic data points to train neural network models for solving complex geometric problems. The synthetic data allows the model to propose and test different solutions when facing complex problems, making AlphaGeometry's problem-solving ability comparable to that of the gold medalist of the Mathematical Olympiad [5]. This kind of training is similar to a high-level human teacher setting up a tutorial class for the AI, changing the angle of the problem, guiding the AI students to practice and improve the level of problem solving. The ability of AI itself is important, but human teachers who can skillfully solve AI problems are important collaborators in AI problem solving.
Humanity has recognized the importance of AI collaboration in the AI era. According to the 2024 World Worker Survey published by Randstad, a Dutch human resources consulting firm, the top five skills that people want to learn today are: artificial intelligence, information technology and technological literacy, well-being and mindfulness, communication and presentation skills, and management and leadership skills [6]. The first two are directly related to AI technology, while the last three highlight people's understanding of human superiority.
(2) The comparative advantage of AI does not negate the significance of human intelligence
The "miracle of force" of the law of scale is widely regarded as the main reason why large AI models produce human-like intelligence. However, there is still a fierce debate in academia and industry about the boundaries of capabilities that can be achieved by the law of scale. Although AI starts from a "blank sheet of paper", the continuous improvement of data, computing power, and parameters can help AI learn and accumulate Hume-style non-causal empirical knowledge, and finally obtain a real model of the world. On the other hand, the law of scale can push AI to achieve AGI and surpass human intelligence, emphasizing the importance of mental structures for knowledge, similar to the importance of a priori conditions for human cognition pointed out by Kant. Cognitive science tends to believe that the child's mind does not start with a "blank slate", but has a priori intuition about a number of core domains, such as quantity, space, subject, and object, and that the child learns from acquired experience based on some kind of "startup software". There are fierce debates between heavyweight experts and scholars on both sides, but the development and practice of AI technology has begun to integrate the views of both sides, and people have begun to impose some kind of "a priori" structure on AI models to improve their performance, while believing in the law of scale stacking computing power. For example, there is the use of AI model "alignment" processing techniques such as supervised fine-tuning (SFT) and reinforcement learning based on human feedback (RLHF), as well as the introduction of hybrid expert (MoE) architectures in the latest models.
The obvious superiority of AI over the human brain does not negate the significance of human intelligent activity. The amount of knowledge that the machine brain can store far exceeds that of the human brain, and the calculation speed is also faster than that of humans, and it can summarize the potential laws in massive texts and data, which can undoubtedly efficiently complete a variety of work in human society. However, we need to distinguish the strengths and weaknesses of AI in more detail, so that we can more accurately judge the relationship between AI and humans in the foreseeable future. According to OpenAI researchers, while large language models only "predict the next word", this includes millions of different tasks or dimensions, including grammar, translation, world knowledge, sentiment analysis, lexical semantics, mathematics, spatial analysis, and more [7]. Starting with ChatGPT, users around the world generally feel that the grammar level of large language models is very high, but the ability of other aspects of the model is different. Grammar is the dimension with the strongest regularity, high exhaustibility, and the richest learnable data (Internet corpus) involved in "predicting the next word", which can give full play to the strength of AI empiricist cognition. However, it is not possible to fully learn emotional, mathematical, and physical world knowledge through the Internet corpus, just like a student who memorizes various question types through the sea of questions tactics, although he can win the standardized test, is not really good at math. So, even though AI has outperformed humans on standardized tests, it doesn't mean that human learning has lost its meaning.
The cognitive starting point of AI is more open than that of humans, and the empirical learning method can "see" the correlation and possibility between huge amounts of data, but in the field where there is no massive sample sample for it to learn, its conclusions are less robust due to the lack of common sense guidance and the ability to construct theories. It's worth pointing out that AI doesn't just include GPT-style large language models. In fact, AI systems with higher accuracy tend to have more initial constraints and a more defined learning path or working mechanism. For example, AlphaGo, which defeated Go world champion Lee Sedol, is based on the Monte Carlo Tree Search (MCTS) mechanism, and the earlier convolutional neural networks (CNNs/ConvNets) for image recognition are based on architectures suitable for extracting image features and supervised learning using labeled data. The bias brought about by these constraints makes these AI models more accurate and reliable for specific tasks, although they are not as generalized and comprehensive as large language models. Regardless of the existing AI model, human intelligence has more general initial rules, common sense intuition, and the ability to construct theories. At the same time, the brain adopts more sparse and selective information processing strategies, which makes human thinking more efficient (e.g., decision-making speed) and cognitive robustness and adaptability (Figure 4.2).
Figure 4.2: Comparison of cognitive acquisition mechanisms between different types of AI and humans
Note: This is not an exhaustive list of AI types that are representative of AI with different architectures and working principles, such as AlphaFold and Sora, which are based on diffusion models or even more complex architectures. Current large language model (LLM) AI has the fewest preset structures, while human preset structures (children's "startup software") are probably the most complex, and the full picture is not yet known in existing scientific research.
资料来源：OpenAI, Google DeepMind, Lake et al. (2017), Wellman and Gelman (1992)[8], 中金研究院
The performance of AI in different types of exams is good and bad, which also reflects to a certain extent that it has its own strengths compared with human intelligence. Taking GPT-4 as an example, its score on the GRE Verbal test is close to 100%, but the score on the high school physics and chemistry test is only about 60%, and the score on the Codeforces Competitive Programming, English Literature and Writing for American College Students, and the Mathematics Competition for American Students in Grades 10 and Below [9]. In a simple image inference game, the human benchmark score is 95%, while GPT-4 scores 69% and GPT-4V only 25% [10]. After human intervention and targeted training, the performance of AI models has improved significantly, for example, in benchmarks for tasks such as natural language reasoning and image classification, AI models have shown a level close to or better than human benchmark performance (Figure 4.3), but AI models are still much inferior in more complex tasks such as trip planning. For example, the Gemini 1.5 Pro only achieves a maximum accuracy of 42% when a human gives 100 examples, while the GPT-4 Turbo performs only 31%. In the case of only one example, the accuracy rate of the two is only about 10% [11]. The platform that tests AI capabilities has published the performance of AI models on tens of thousands of benchmarks across thousands of tasks[12], and the results show that the capabilities of AI models on these tasks vary widely, reflecting both the strengths of AI in specific tasks and the space for human engagement and continuous improvement of AI.
Based on the current level and development route of AI technology, we should objectively look at the capabilities and weaknesses of AI. Since everyone's skills are different, advances in AI will inevitably have a heterogeneous impact on individual humans. For humans, AI is both a tool and a competitor. As for the relationship between humans and AI, we should neither be blindly optimistic nor arrogant, but should play the role of humans and AI working together. In reality, regardless of AI capabilities, although AI can assist humans in many fields such as medicine, the final decision-making and responsibility must still be borne by humans [13], and humans should not give up their ultimate responsibility for their work [14]. In the long run, humanity needs to maintain its dominant position in the development of more powerful AI and even AGI, to ensure that the development of AI and human intelligence is a healthy competition, and to ensure that the improvement of social productivity is consistent with the direction of human well-being.
Figure 4.3: AI models of all types meet or approach human benchmark performance across a variety of tasks
Note: Multi-task language understanding, or MTLU for short, refers to the simultaneous completion of multiple natural language processing tasks through a single model. The AI performance shown in the figure does not refer to the performance of a specific AI model, nor is it a large language model, but refers to the score of the AI model that performs the best on this task, and there are many AIs that have specialized human intervention and training for a certain task. For example, models such as ViLBERT and GPT-4RoI perform better in visual common sense reasoning tasks, and models such as ANNA perform better in primary reading comprehension.
资料来源：Stanford AI Index Report 2024，中金研究院
2. How does AI affect employment?
The impact of AI on employment is a growing concern around the world. According to a 2023 survey conducted by market research firm YouGov[15], about three-fifths of respondents (57%) globally are concerned about jobs being replaced by AI. In Asia, a significant proportion of respondents are concerned about jobs being replaced by AI. A whopping 76 per cent of respondents in India are worried, while about seven in ten respondents in the UAE and Indonesia are worried. In contrast, respondents in Chinese mainland and Hong Kong were relatively calm, with 55 percent expressing concern. A 2023 survey of 31 countries conducted by Ipsos Global Advisor also showed that an average of 57% of workers worldwide expect AI to change the way they currently work, and 36% expect AI to replace their current jobs. An earlier Pew Research Center survey of 10 countries [17] showed that when people look to a longer term in the future, such as 50 years, the vast majority of people believe that robots and computers may take over many of the jobs that are now performed by humans, with 91% in Greece, 89% in Japan, and 84% in Canada.
Do people really need to be so worried? To judge the overall impact of AI on the job market, it is necessary to analyze and understand it from multiple perspectives. We will first explore the impact of AI on different occupations in the Chinese workplace, and then summarize the possible impact of AI on China's overall job market. Our research finds that so far too much concern about the substitution role of AI for human labor has neglected the augmenting effect; While the penetration of AI into the economy will inevitably lead to a slowdown in job growth and structural unemployment, there is no prospect of AI-induced mass unemployment in the foreseeable future.
(1) The substitution and enhancement effects of AI on different occupations: a study based on China's recruitment big data
Existing research employs a variety of methods to assess the impact of AI on different professions. The basic idea of these methods is to assess the overlap between AI and human activities in a job task, and the specific evaluation methods are broadly divided into three categories: machine learning taxonomy based on expert scoring, text analysis methods based on patent text and job descriptions, and direct evaluation using large language models such as GPT-4. Regardless of the methodology, the above research is inseparable from high-quality data on occupational characteristics. Most of the current studies have used the American Occupational Traits Database (O*NET) [18] or similar data from other countries [19]. O*NET currently includes 923 SOC occupational classifications, as well as descriptions of occupations involving a wide range of dimensions such as tasks, abilities, knowledge, education, skills, interests, work activities, work styles, wages, and employment trends. O*NET data is collected from working individuals or career experts and is regularly revised to keep up with the changing occupational environment, with its most recent revision being in 2019 [20].
The classification of machine learning algorithms based on expert scores comes from Frey and Osborne (2013) pioneering research in this field [21]. Frey and Osborne (2013) made an expert judgment on the degree to which 70 occupations can be automated by AI-related technologies such as machine learning and mobile robots, combined with the U.S. O*NET database to determine the "automation bottlenecks" of each occupation, such as perception and operation, creativity, and social intelligence, and finally used the labeled occupations and task data to calculate the automation probability of all 702 occupations. This method is a scientific method for assessing the impact of AI careers, but there is also a significant problem: the results are quite subjective because they rely on expert labeling and evaluation. With the development of natural language processing technology, some studies have begun to use text analytics methods [22] to measure the degree of exposure of occupations to AI by analyzing the similarities between occupational task descriptions and technological advancements that have occurred, such as patents or research papers. Recent research is beginning to use large language models, such as GPT-4, to assess how well occupational tasks match AI capabilities, with the advantage that AI's assessment of occupational exposure is more accurate, timely, and cheaper. Based on AI's large knowledge base, its evaluation accuracy is no less than that of human experts, and its high efficiency makes research based on massive amounts of data possible [23].
Previous studies once generally considered that occupations that involve routine tasks, such as office work, production, and sales, face a higher risk of AI automation. Occupations that require more creativity, social skills, and emotional investment, such as education, healthcare, and artistic creation, are less affected by AI [24]. However, recent research has found that as AI advances in language capabilities, high-skilled occupations such as law, education, and creativity may also face a higher risk of AI substitution [25]. There are some common shortcomings in existing studies. First, due to the availability of data, research has focused too much on developed countries and less on developing countries. Second, most of the existing studies use O*NET data from the United States to describe occupational characteristics, and there may be differences in occupational characteristics in different countries. Third, most studies do not distinguish between the substitution and enhancement effects of AI, and the mechanisms of their impact on employment and wages are different. Finally, O*NET was last fully updated in 2019, resulting in a lag in existing research data that fails to reflect the rapid development of AI in the past few years.
Using the big data of recruitment advertisements on China's Liepin and Zhaopin.com [26], we used locally deployed large language models to construct a database of occupations-tasks/job responsibilities in China, and measured and compared the AI exposure of different occupations. To ensure comparability with other studies, we categorized job openings in China according to a six-digit hierarchy of the American Standard Classification of Occupations (SOC-2018). For each SOC-6 occupation category, we randomly selected about 400 ads, totaling more than 200,000 ads, and obtained data on tasks and job responsibilities covering 544 SOC-6 occupations. By extracting the top 10 core tasks for each occupation and calculating the importance weights based on how often these tasks appear in all job advertisements, we obtained a quantitative and comprehensive job description system for Chinese occupations. As far as we know, this is the first attempt in China to build an occupational task database based on real job market data and an internationally comparable occupational classification system, rather than expert judgment or indirect mapping. With this database of occupational tasks, we can use AI technology to measure the level of AI exposure in each occupation. Since we use the latest data as of April 2024, the results of the study can reflect the latest changes in AI technology and the content of career tasks. Unlike existing studies, which rely primarily on U.S. O*NET data, our occupation data is based on Chinese recruitment data, which reflects the micro characteristics of China's job market.
Based on the above database, we asked large language models such as GPT-4 to score the substitution effect and enhancement effect of each task under each occupation, and summed up the AI exposure measurement of each occupation according to the importance of the task in the occupation description. As mentioned earlier, the feasibility and effect of allowing AI to evaluate the AI exposure of each occupational feature task makes full use of the text analysis capabilities of large language models, and has been confirmed in similar studies. Not only do we save a lot of time and money by using this approach, but the results are more objective and neutral than human expert assessments. To make up for the shortcomings of existing research, our study expands the dimension of similar research. First, we distinguish between the substitution (automation) effect and the augmentation effect of AI, and evaluate their impact on employment and wages, respectively. Second, we consider the impact of three different types of AI on human occupations: AI in the general sense, large language model AI, and large language model plus humanoid robot. Third, we also consider the cost factor of AI, i.e., whether the cost of AI can be reduced to a level that is acceptable to the enterprise for a limited period of time.
Regarding the substitution and enhancement effects of AI, the literature shows that they have different mechanisms and outcomes on employment and wages [27]. The substitution effect reduces the demand for related employment, but the enhancement effect may increase the demand for related employment and promote wage growth due to higher labor productivity and inter-industry spillover effects. Our calculations show that the impact of AI on different occupations varies significantly (Chart 4.4). Some of the most repetitive occupational groups include: office and administrative support, transportation and material handling, computing and mathematics, and life, physics and social sciences. Because GPT-4 believes that there is a high probability of autonomous driving in the next decade, many occupations under the transportation and material handling category have high substitution scores [28]. Research work such as information and data collection and processing, and report writing are also considered to be the dominant areas of large-scale model AI, and the substitution role scores are high.
In our results, sales, legal, and management were assessed by GPT-4 as difficult to replace because they "contain complex decisions, nuanced differences, ambiguous interpretations, or context-specific information," while all three occupations were assessed as having high scores for enhancement, especially sales. This result may be related to the fact that most of the job advertisements for the core tasks of these occupations are results-oriented, rather than simply independent actions. Taking telemarketers as an example, existing studies generally believe that AI has a high degree of substitution, but our measurement has a low substitution score. This is not due to the cost factor, on the contrary, 100% of the ten core tasks sold have passed the cost test of applying AI; Instead, the job description includes tasks that require personalized solutions, empathy, and nuanced observation of human behavior, such as "sales target achievement", "new customer development", "sales strategy execution", and "customer satisfaction improvement", which GPT-4 believes are difficult for AI to complete independently. In contrast, the O*NET database describes the tasks of telemarketers mainly as "contacting customers by phone, obtaining customer information, and explaining products and services". Community and social services also have higher enhancement scores, mainly because occupations such as "education, guidance and career counseling counselors", "marriage therapists", and "mental health counselors" include both AI expertise such as data collection and Q&A, as well as sales attributes.
Figure 4.4: Exposure to AI technology by occupation in China
Note: Greater than zero means that the occupation is more affected by AI than the mean of each occupation; The squares represent the median, the vertical lines represent the P25 and P75 data points, and the ends of the lines represent the P10 and P90 data points. The straight line shows the distribution of AI exposure scores for more than 30 SOC-6 specific occupations under each SOC-2 occupational category.
Source: Liepin, Zhaopin, CICC Research Institute
This leads to an interesting question, which we briefly touched on in the first section of this chapter: who is responsible for the results of work and profession, humans or AI? If humans do not give up their responsibility for their work, this willingness in itself will lead to a different assessment of the substitution of AI for human work. So should humanity abdicate the responsibilities that are still in our hands? Does work, as a basic component of human social activities, inherently contain the premise that "human beings cannot shirk their responsibilities"? It's hard to answer that. Interestingly, however, large language models, which are AI, do not seem to be willing to take responsibility for the results of their work as blame for the results of their work after reading and understanding a large number of job advertisements.
We find that there is noteworthy heterogeneity in the impact of AI on subdivided occupations under some occupational categories. The occupations with the strongest substitution heterogeneity include: office and administrative support, transportation and material handling, agriculture, forestry, animal husbandry and fishery, security, production, etc., which is manifested by the large length of substitution lines in Figure 4.4. Taking agriculture, forestry, animal husbandry and fishery as an example, it includes both "farm workers and laborers, crops, nurseries and greenhouses", which have only 1/10 of the physical work, and "agricultural product grading and classification workers", which have an AI replacement degree of up to 80%, so it brings great heterogeneity. Similar characteristics are shown in areas such as security. It is worth mentioning that if one takes into account the three laws of respect for humans specified by Asimov for robots [29], some of the tasks of security work may never be performed by AI. This heterogeneity of the impact of AI within occupations suggests that we need to be more cautious and meticulous when analyzing the impact of AI on employment.
Interestingly, when we limit AI to pure large language models, the proportion of tasks that AI can completely replace humans is only 2% of the total tasks on average, which makes the substitution effect of AI very small, even without considering the cost feasibility dimension. AI-enhanced tasks account for up to 90% of the total, and it has high enhancement scores across all occupational categories. It's important to note that this is an assessment made by the large language model itself. Unlike human concerns, large language models seem to be more inclined to position themselves as human tools. When humanoid robots are taken into account, the proportion of tasks that can be replaced by AI increases slightly compared to AI in a general sense, from about 16% to about 23%, and occupations with significant increases in substitution include: community and social services, food preparation and services (mainly catering), personal care and services, office and administrative support, sales, etc. Estimating a profession's AI exposure is a dynamic process. Our results only reflect the judgment of large language models on the development of AI in the next decade, and the content behind this judgment is a black box that we do not know. In addition, occupational exposure is only one of many factors that affect employment and wages, and other factors such as the speed of technology adoption, structural changes in the economy, and policy responses also play an important role. Therefore, our estimates can be seen as a reference and starting point, rather than a decisive conclusion on the direction of labor demand for each occupation.
Exhibit 4.5: Responses to the questionnaire on which jobs are likely to be replaced by AI
Note: The chart shows the results of CICC's April 2024 survey of listed companies, with the specific question "Which jobs are likely to be replaced by AI in the next three years?" ”。
Source: CICC Research Institute
Figure 4.6: Responses to the questionnaire on AI substitution or job creation
Note: The chart shows the results of CICC's questionnaire survey of listed companies in April 2024. The red bar indicates that AI is expected to replace existing workers, and the gray bar indicates that AI is expected to create new jobs. For better comparison, the answer "new jobs will be created, but I don't know what they will be" are grouped into the "uncertain" category in the diagram.
Source: CICC Research Institute
We also conducted a questionnaire survey of some listed companies in China and received 112 valid responses covering most industries, of which about 30% were manufacturing companies. The results of the questionnaire are similar to those of our previous studies. For example, positions such as customer service, human resources, and administration, as well as production positions, are more likely to be replaced by AI in the next three years (Figure 4.5). Faced with the question "Will AI replace existing employees in your company in the next three years?" 39.3% of respondents said they were "not sure". When asked whether the use of AI will create new jobs in your company in the next three years, 53.6% of respondents said that new jobs will be created, but I don't know what they will be (Figure 4.6).
(2) AI may lead to a slowdown in job growth over the next decade, but it will not lead to mass unemployment
Regarding the overall impact of AI on employment, there are conflicting views in existing studies. Some studies suggest that AI and automation technologies contribute to higher unemployment rates, with multiple studies estimating that 14-56% of existing jobs are at high risk of being automated [30]. However, other studies have argued that AI technology will lead to job growth, because AI can create jobs by improving efficiency and promoting industrial upgrading, and its job-creating effect is enough to offset the effect of replacing labor [31]. According to the World Economic Forum's Business Survey, about 50% of businesses expect AI to create jobs, while only 25% expect them to be reduced [32]. Overall, the existing research literature shows that the employment impact of technological change depends on the combined effect of mechanisms such as the substitution effect and the creative effect on human work, and the overall effect is highly uncertain [33].
Exhibit 4.7: Estimates of the impact of AI on employment growth rates by occupational group over the next decade
Source: Zhaopin, CICC Research Institute
We estimate the potential impact of AI on China's employment growth based on the AI exposure index measured in the previous section and the existing research literature. Kogan et al. (2023) correlated the AI exposure index with employment growth over the next decade: a 40 percentage point increase in AI exposure for substitution corresponds to a 4.4 percentage point decrease in the cumulative employment growth rate over the next decade, while a 40 percentage point increase in enhanced AI exposure corresponds to a 9.1 percentage point increase in employment growth over the next decade [34]. Based on its estimates, we calculated the change in employment growth rate per unit of AI exposure change, and applied this coefficient to the Chinese SOC-6 occupations to calculate the net employment impact for each occupation. We find that AI has a different impact on job growth across different occupations in China (Chart 4.7). The impact of AI on total employment can be obtained by weighting the net employment impact of each occupation as a proportion of total employment in China [35]. Our estimates suggest that AI could lead to China's cumulative employment growth over the next decade being 1.8 percentage points lower than the baseline level, with an average annual growth rate of about 0.18 percentage points lower. Compared with the huge number of employed people and labor force on the mainland, the impact of this figure is not significant. This means that AI itself will not cause mass unemployment in the next decade.
There are, of course, uncertainties in the above estimates. First, the change in employment growth here takes into account the productivity effect at the sector-wide level and its positive spillover effect, i.e., AI-enhanced industries expand and spill over to other industries, creating new labor demand. If we only consider the impact on existing employees, the probability of unemployment will also increase due to the introduction of new technologies due to the outdated skills of existing workers and the decline of human capital (skills loss effect). This reminds us to be cautious in interpreting changes in the headline unemployment rate, as a more stable overall picture may mask structural changes. Second, we use the micro parameters of the labor market based on US data, which we are currently unable to estimate due to the lack of micro data. Despite these limitations, our estimates provide a basis for assessing the potential impact of AI on employment in China.
3. Increasing differentiation is a feature of income distribution in the AI era
(1) AI may lead to a widening of the wage gap
Existing studies employ a variety of methods and data to assess the impact of AI and automation on wages. A major consensus is that these technological changes are likely to exacerbate wage inequality, but different studies have different opinions on the mechanism and extent of the impact. Empirical studies have found that in regions with high levels of information technology adoption and automation, there is wage polarization in the labor market [36], with fewer middle-wage jobs and more jobs at both ends [37]. The International Labour Organization (ILO) reports that while technological advances create new jobs, they are likely to exacerbate inequality, with wage losses being the most severe among low-wage workers, women, and those in informal employment [38]. It has also been suggested that AI may accelerate automation, which may disproportionately affect low-skilled workers [39]. In contrast, highly skilled workers and less affected occupations may experience wage increases, increasing the wage gap [40].
The impact of AI on the wage gap can be more complex and nuanced than existing information and automation technologies. Some studies have distinguished between the labor substitution effect and the labor enhancement effect of AI, and found that they have different effects on wages and employment. The connotation of labor-saving technology is to improve the quality of capital goods that can replace workers in routine tasks (or reduce prices while the quality remains the same), and promote the substitution of labor by capital. In contrast, labor-enhanced technologies can increase worker productivity and benefit workers who master new technologies, but they can have a negative impact on workers who are skilled in old technologies and unable to adapt to them, mainly older, more educated, and with higher relative wages. Overall, the substitution effect is associated with a decrease in the share of labor, and the enhancing effect is associated with a slight increase in the share of labor [41]. Another group of arguments emphasizes that AI may have a positive impact on employment and wages by increasing productivity and creating new tasks, but exacerbates income disparities between different groups because they benefit unevenly. For example, R&D-intensive innovative firms pay higher wages, which exacerbates the wage gap between firms [42]. Analysis and literature review based on Chinese data also show that AI has increased the urban wage premium of the non-conventional labor force, but has had a negative impact on the conventional labor force, especially certain groups (women, high-skilled), exacerbating income disparities between regions, industries, and groups [43].
We estimate the potential impact of AI on workers' wage gaps using the wage distribution data of each occupational sub-category in China's online recruitment data, combined with the AI exposure index of each occupation in China that we measured. Consistent with the literature projections, occupations with strong AI substitution effects experienced slower cumulative wage growth in 2018-23 (Figure 4.8). At the same time, the Gini coefficient of wages within occupations with less AI augmentation increased slightly, but to a lesser extent (Figure 4.9). The literature points out that there are two opposite forces affecting wages for AI-enhanced occupations, namely, the intra-occupational wage polarization caused by skill-biased technological progress, and the skill loss effect on the narrowing of the wage gap between old and new employees. According to Chinese data, there is no effect that has shown an overwhelming impact in occupations that have been greatly enhanced by AI. Further differentiating between the two effects requires more microscopic worker-level data. However, caution should also be exercised when interpreting the impact of AI on income disparity. Given that large language models will only be widely used from 2023 onwards, their impact on the wage gap still needs to be accurately assessed by observing the data.
Figure 4.8: Occupations with strong AI substitution have slower wage growth
Note: The data shows the growth rate in 2023 compared to 2018.
Source: Zhaopin, CICC Research Institute
Figure 4.9: Intra-occupation wage gaps with strong AI augmentation have not widened
Source: Zhaopin, CICC Research Institute
(2) AI may lead to a decline in the distributed share of labor in GDP
Existing studies have used a variety of methods and data to analyze the determinants and trends of the share of labor income. A major consensus is that the share of labour income has been declining over the past few decades. Some studies based on U.S. data show a declining share of labor income in manufacturing. As for the driving factors, earlier studies have emphasized that capital deepening is a key factor leading to the decline of labor share [44]; Recent studies have begun to use industry data to analyze trends in labor share, such as a 59-country study in the Quarterly Journal of Economics, which found a significant decline in labor share globally, about half of which can be explained by declines in the relative prices of investment goods [45]. The study, based on U.S. industry data, found that the decline in labor share occurred mainly within industries, especially in manufacturing and trade; Therefore, outsourcing labor-intensive supply chains may also be a key reason for the decline in the share of labor in the United States [46]. Studies based on European industry data have also found that capital deepening and changes in the structure of sectoral employment are the main factors contributing to the decline in the share of labor in Europe [47]. Recent studies have begun to focus on the impact of skill-biased technical changes (SBTCs) on labor share, suggesting that it may explain the 20% decline in labor share in the U.S. manufacturing sector between 1970 and 2010 [48]. The above study describes and explains the downward trend in the share of labor income over the past few decades, especially in sectors such as manufacturing (Chart 4.10). However, if we take a longer view, the change in the share of labour income over the past two hundred years has been more cyclical than there is a clear linear upward or downward trend (Figure 4.11). This reminds us that we should have a comprehensive and dynamic perspective when analyzing changes in the share of labor income, and although the decline in the share of labor income over the past few decades is a real problem, it does not lead to the extreme conclusion that the trend will continue in the long run, let alone directly lead to the extreme conclusion that factors such as technological progress will cause the share of labor income to fall to zero.
Exhibit 4.10: U.S. Manufacturing Declines in Labor Income Share
Note: The chart shows three measures of the share of labor income in the U.S. manufacturing industry.
资料来源：Autor et al. (2020)[49]，中金研究院
Figure 4.11: The share of labor income is relatively stable from an ultra-long-term perspective
Note: The chart shows the share of labour and capital in national income in the UK between 1770 and 2010, with similar trends in other countries, such as France.
资料来源：Piketty and Goldhammer (2014)[50]，中金研究院
The theoretical literature makes some predictions about the change in the share of labor income in the era of AI. Korineck and Stiglitz (2017) argue that this is not necessarily a problem even if AI eventually completely replaces human labor; Because in this case, although the economic output is mainly created by machines, the absolute level of income of workers does not fall [51]. Aghion et al. (2017) argue that even in a future where AI technology is highly developed, the share of labor income will tend to stabilize rather than zero; Because the actual degree of automation is endogenously determined, it is constrained by the substitution between products [52]. Acemoglu and Restrepo (2018) proposed the concept of the self-stabilizing effect of automation and the balanced growth path. They argue that when the rate of automation exceeds the rate of new task creation, automation reduces the cost of using labor, which inhibits further automation and creates new tasks that are more suitable for labor to complete. This self-stabilizing effect allows automation and the creation of new tasks to advance at an equal rate, thus maintaining a stable share of labor income. In response to Wassily Leontief's pessimistic prediction in the 1980s about machines replacing people: "The new technologies that emerged in the early 20th century made horses redundant... Labor will become less and less important", Acemoglu and Restrepo (2018) respond that "the difference between human labor and horses is that humans have a comparative advantage in more complex new tasks (resulting from the application of new technologies), while horses do not" [53]. However, due to the early adherence of AI technology, it is difficult for current empirical studies to give a definitive answer to the impact of AI on the share of labor income.
We estimate the potential impact of AI on the share of labor income using the parameters of AI exposure and literature estimates for each occupation in China [54] measured above. The results show that the share of labor income is likely to decline by 0.73% cumulatively over the next five years compared to the baseline scenario, mainly in office and administrative support, agriculture, forestry, animal husbandry and fishery, production, transportation and material handling, and sales (estimated employment growth but wage decline). Education, management, law, community and social services, and health care support were among the few occupations that saw an increase in their share of labour income, but by less than 0.1 per cent. The results of our survey of Chinese listed companies also reflect expectations that AI may lead to a decline in the share of labor income. When asked, "Do you expect AI to reduce the proportion of your company's labor costs as a percentage of your total operating costs in the next three years?" 45.5% of respondents answered "very likely", 47.3% answered "not surely", and only 7.2% answered "not possible".
Fourth, AI will impact the existing human capital accumulation model
Traditional human capital theories hold that education can lead to higher productivity and income returns by improving workers' knowledge and skills [55]. However, the development of AI may affect human capital accumulation activities in two ways. On the one hand, the partial replacement of human labor by AI may reduce the value of certain conventional skills, thereby weakening the marginal rate of return on investment in education [56]. On the other hand, the impact of AI on different skills is uneven: workers who master key skills to collaborate with AI may have an increased value of their human capital [57].
Using information from the 2018-23 recruitment big data, we found that over the past five years, the education requirements for those with high AI enhancement increased more than those with low AI enhancement (Chart 4.12), while the number of years of work experience required increased more slowly (Chart 4.13). At the same time, our survey of listed companies in China also shows that the AI era does not reduce the importance of education in the eyes of employers. When asked, "If AI becomes more and more widely used in your company, will your company still pay attention to the candidate's academic qualifications or school when hiring in the future?" The proportion of respondents who answered "yes" was as high as 77.7%, 19.6% answered "not sure", and only 2.7% answered "no". Employers' emphasis on academic qualifications over work experience may reflect the value of state-of-the-art skills and the depreciation of human capital with "generational characteristics". The value of human capital is affected by the age or era in which it was acquired, which is the "vintage-specific human capital" of human capital. The expertise or knowledge of a past era may seem obsolete in the face of a new wave of technological revolution. In the era of AI, the key to the accumulation of human capital is not academic qualifications or working years, but the ability to continuously learn and keep skills updated.
Exhibit 4.12: Occupations with strong AI augmentation have increased their educational requirements more
Note: The data shows the growth rate in 2023 compared to 2018.
Source: Zhaopin, CICC Research Institute
Figure 4.13: Occupations with strong AI augmentation have a slow increase in experience requirements
Note: The data shows the growth rate in 2023 compared to 2018.
Source: Zhaopin, CICC Research Institute
The impact of AI on the accumulation of human capital or even the development of human intelligence may not stop there. The progress of human intelligence is often based on a large number of ordinary creations. In the fields of art, science, and scholarship, the vast majority of works cannot reach the pinnacle of human intelligence, but their existence provides the soil for the emergence of geniuses and the birth of masterpieces. Many great scientists and artists have come out of the ordinary, and their growth is inseparable from the accumulation of general work. Even if it is a genius achievement, the flash of inspiration is based on thinking and trying day after day, not innate. However, AI has the potential to fundamentally change this landscape of human intelligence: AI's ability to massively generate works could significantly reduce the market value of ordinary works created by humans. This may lead to the withdrawal of human beings from learning and creation in related fields, and also reduce the likelihood of great works. For example, the efficiency of AI in translation may make fewer and fewer people willing to devote a lot of time to learning foreign languages professionally or engaging in translation work full-time, and it is the long-term, professional practice and experience that have created outstanding literary translators such as Ge Baoquan and Zhu Shenghao. From this perspective, AI may not only fail to surpass human intelligence, but also reduce the probability of great results.
In order to cope with this dilemma, society may need to take precautions and be cautious about the development and application of AI. One idea is to reserve space for human creation, preventing the mass market from being completely dominated by AI. This may require the proper boundaries of AI in key areas to protect the space for human intelligence to develop. Another way of thinking is to give special value to human creations and highlight their uniqueness, such as labeling and certifying "original human creations". If the market can respond effectively and spontaneously, such as by establishing a segregated equilibrium between humans and AI, this may not be an imminent risk in the foreseeable future, and no policy intervention will be required. However, it is important to keep a close eye on the changes in the value of AI and human works in the mass market. In the era of AI, individuals should focus on continuous learning to improve their capabilities, and society also needs to reserve market space for human development and protect the systems in which human intelligence can be trained.
5. Flexible employment and social security in the AI era
In the AI era, flexible employment is showing a trend of further expansion. The International Labour Organization (ILO) defines "non-standard employment" as temporary work, part-time work, and multi-party employment relationships (including the platform "gig economy" and "on-demand economy")[58], which is collectively referred to as "flexible employment" in this chapter. The current round of artificial intelligence technology development, represented by large language models, may further increase the proportion of flexible employment. Large language models can facilitate more natural and efficient communication between employees and the company, providing real-time translations for smooth collaboration across geographic boundaries. It also provides easy-to-use tools for tasks like content creation, customer support, and market research, lowering the barrier to entry for becoming an entrepreneur and freelancer. According to a recent survey, nearly seven in ten respondents believe generative AI will increase their likelihood of becoming a freelancer. In particular, high-skilled independent workers (among the top 2% in their respective fields) have been enhanced by AI technologies and have increased productivity, with nearly half of the respondents already building generative AI solutions for various businesses [59].
In the era of the digital economy, flexible employment has become an important part of total global employment, promoting full employment, but there are also many controversies related to labor protection and employment quality. As of 2021, there are about 200 million people in flexible employment in China, about 4.5 million people in the UK regularly find work through online platforms, and the new digital economy policies in Japan and South Korea have also promoted "squawning people" to become "digital nomads" [60]. Digital platforms and flexible employment have improved the mobility and matching efficiency of the labor force, but at the same time, flexible employment also has many labor security problems. First, flexible employment lacks a formal labor contract relationship, with data from 2019 showing that only about 8% of platform workers in China have established a formal employment relationship with a platform [61]. Second, the platform's high bargaining power may lead to practitioners being in a relatively weak position. For example, studies have shown that groups such as food delivery workers and ride-hailing drivers work longer hours and have lower net income levels in some cities [62]. Third, flexible employment often suffers from the problem of insufficient unemployment insurance protection. As a whole, the proportion of unemployment insurance recipients in the total unemployed population is relatively low, and unemployment insurance coverage needs to be further improved. In 2023, the G20 countries jointly called for adequate and sustainable social protection for workers in the gig and platform economies[63].
The expansion of AI for flexible employment also highlights the urgency of improving the social security system in the AI era. The development of AI is likely to exacerbate the challenges faced by flexible workers, such as the substitution effect that may increase job instability. Studies have shown that after the launch of ChatGPT, the number of freelance job postings with a high substitution effect decreased by 21%[64]; Entry-level freelancers, such as those who write formulaic SEO articles or HTML code, do basic data analysis, and graphic design, are at greater risk [65]. In the era of digital economy, it is difficult to incorporate gig social security into the "three-in-one" model, which combines social pooling and personal accounts, and is shared by government, enterprises, and individuals according to a certain proportion [66]. The platform and the employer believe that they should not bear social security liability. Flexible workers also have a lower ability and willingness to pay, and even if they are enrolled, they often choose a plan with a lower level of protection, which leads to a lower level of pension after retirement [67].
In order to meet this challenge, the social security system needs to be improved from multiple perspectives. Social security has two functions: compulsory savings and redistribution. In the traditional employment relationship, the enterprise assumes the responsibility of paying social insurance for employees and helps employees complete compulsory savings to avoid irrational inter-period consumption behavior. However, under the new situation, it is difficult for enterprises to determine whether a flexible employee is their own employee, and therefore cannot perform the "compulsory savings" function for them. Starting from the function of compulsory savings, it is necessary to strengthen the certainty of enterprise contributions, which is also conducive to the survival of the social security system. In order to achieve this, it is necessary to strengthen the legal relationship and rights and obligations of enterprises and workers from the signing of labor contracts and other links, and determine the responsibility for payment.
From the perspective of the function of redistribution, social security is an important system for income redistribution in mainland China, and the degree of redistribution mainly depends on the scale of social security relative to GDP. The ratio of social insurance expenditure to GDP in mainland China has gradually increased from 1% in 1989 to about 11% in 2019, indicating that the level of social security is constantly improving with economic development, but there is still a certain gap compared with the average level of OECD countries. The government could consider providing financial subsidies for low-income people to participate in the program, and should also incentivize those in flexible employment to participate in the insurance. The International Labour Organization (ILO) has recommended a more flexible and portable social security scheme for non-standard workers [68]. Migrant workers face relatively many challenges in the job market, and a certain proportion of migrant workers are engaged in occupations affected by AI substitution, such as production and logistics [69], and there is room for improvement in the social security of rural elderly people who are closely related to their economic decision-making [70]. While ensuring the steady growth of social insurance expenditure, it is of great significance to improve the protection of vulnerable groups in the era of AI.
6. Reflection and enlightenment
The previous analysis shows that AI is profoundly affecting the labor market, and occupations with routine tasks are at a higher risk of AI substitution, while more occupations are mainly affected by AI augmentation. After taking into account the positive effects of AI on productivity and job creation, the negative impact on total employment is limited. However, all types of occupations are facing disruptions from the introduction of AI technology, and even in occupations with high AI enhancement effects, self-employed workers face higher employment uncertainty. At the same time, AI could exacerbate the wage gap and could lead to a small decline in the share of labor income. In response to the negative impact of AI on the job market, policies can play a role in both the primary and redistribution phases.
(1) Initial distribution stage: vocational training and labor protection
In the field of primary distribution, the main focus of policy is to promote the acquisition of new skills for workers to adapt to the AI era, and vocational training is a realistic policy starting point, which can also promote the employment and income of workers with the least distortion of the labor market [71]. In order to improve the relevance and effectiveness of vocational training, the government can explore the creation of an enabling environment that supports lifelong learning and individualized training [72], and transform it into an "enabler" role, allowing workers to choose and participate in retraining through the provision of training vouchers, scholarships, etc., to better meet the specific needs of different workers. The government should provide these incentives in particular to workers working in small and medium-sized enterprises (SMEs) and gig platforms that lack training mechanisms [73]. One of the challenges of skills training is that the future of AI is uncertain and requires forward-looking training programs. To do this, training institutions need to update their approaches to align teaching and learning management processes with technological advances, leverage contemporary technologies and innovations, and adopt experiential-based education to develop students' resilience and lifelong learning capabilities [74]. On the other hand, enterprises should actively provide on-the-job training and career development opportunities that are compatible with new technologies. A McKinsey survey of business executives in developed countries in Europe and the United States shows that companies themselves are willing to do the same, with 64% of US private sector executives and 59% of European private sector executives believing that companies should take the lead in closing the skills gap, rather than waiting for governments, universities, or individuals to do so[75].
While strengthening skills training and education, governments should pay attention to the challenges posed by AI to labor protection, especially the use of AI in workplace monitoring and recruitment and evaluation. Applying AI to implement monitoring and algorithmic management of employees provides companies with more control over the work process, which can put stress on employees and reduce their autonomy. The National Labor Relations Board (NLRB) has launched an investigation and warning against a range of AI-based surveillance practices in the workplace. For example, Amazon uses AI to track the movement of warehouse workers and automatically generate performance goals, which results in: every step an employee takes in the workplace, every conversation, and every toilet visit is recorded; Employees who talk to each other for more than 30 minutes will receive a written warning. According to a Pew Research Center survey conducted in the United States, more than half of respondents disagree with companies using AI to track employees' movements, time spent at their desks, and their activities on work computers (Chart 4.14). In addition, AI is being used for hiring, evaluation, and firing decisions. According to the survey, 83% of U.S. companies currently use AI in recruitment and selection [76]; This may seem to reduce human bias, but if the algorithm is trained on historical data that reflects past discrimination, there is a risk of perpetuating established injustices.
Figure 4.14: Percentage of Americans view companies using AI to monitor different types of work-related behaviors
Note: The survey was conducted from December 12 to 18, 2022.
资料来源:Pew Research Center,中金研究院
(2) Redistribution phase: progressive taxation and transfer payments
The government's policies at the primary distribution stage can help workers acquire skills and keep jobs, but in reality, many people find it difficult to re-enter the workforce through vocational training. Therefore, in addition to providing training opportunities, the government also needs to consider redistributive policies to provide basic livelihood support and support to the affected population. First, redistribution is not only feasible in the age of AI, but it can also deepen the political foundation for technological progress. Korinek and Stiglitz (2017) cite the example of the 19th-century Luddite movement of textile workers in England to destroy machines, arguing that without appropriate redistributive policies, "preventing innovation" can become a natural response for workers in a deteriorating situation. In a worker-dominated country, visionary innovators should support redistribution to ensure that workers' conditions are not exacerbated by technological progress. Innovation expands the boundaries of production possibilities, and appropriate redistributive policies can enable workers and entrepreneurs (innovators) to share the fruits of technological progress. Conversely, without proper redistribution, the new economic equilibrium is more likely to be at E1, i.e., entrepreneurs benefit and workers worsen, which may destabilize the social foundations that support innovation and hinder the continuous advancement of technological progress [77].
The redistribution programme involves two questions: how to finance and how to distribute, i.e. "where does the money come from" and "where does it go"? Each of these issues affects the efficiency and feasibility of redistribution. From a financing point of view, theoretically the least distorting tax scheme should be chosen. The "robot tax" supported by American political and business celebrities such as Bill Gates, Elon Musk, and Bernie Sanders may not be the best solution[78]. Schaefer and Schneider (2024) attempt to demonstrate that taxing labor income is a less distorting option than a robot tax, using a macroeconomic model that includes generational overlap [79]; Because a robot tax reduces investment and economic growth and does not have a redistributive insurance function, it is not a policy preference for rapid and inclusive growth. At the same time, attention should be paid to the graded and progressive labor income tax, because AI will increase labor income inequality, and progressive tax rates can play a redistributive role. Since the income from land appreciation is mainly derived from social progress rather than individual efforts, taxation of land is conducive to promoting social equity; A similar line of thinking suggests taxing factors such as capital, that receive "windfalls" as a result of AI applications. However, this requires a clear distinction between the existing capital of "windfalls" and the new capital that invests directly in AI technology [80], otherwise it may undermine AI technological progress.
Figure 4.15: Technological progress expands the boundaries of production possibilities, and appropriate redistribution leads to a new equilibrium of win-win for workers and entrepreneurs
Note: The northeast direction of E0 represents the area where the equilibrium point with Pareto improvement should be located.
Source: Korinek and Stiglitz (2017), CICC Research Institute
In terms of distribution, universal basic income (UBI) is receiving increasing attention in Western countries as a potential policy tool to deal with the risk of skilled unemployment [81]. Existing research uses a variety of methods and data to assess UBI, and one main consensus is that while UBI may help provide income security and reduce inequality, there are significant economic, political, and ethical challenges to its implementation. Proponents of UBI are currently advocating for two different options: one to keep most of the existing benefit programs and add modest UBI, or to drastically reduce or eliminate benefit programs and fund UBI by reallocating those funds. Advocates of the latter option argue that this can reduce the size of the government and improve market efficiency, while simplifying procedures also reduces the cost of implementation. Some studies have suggested that when comparing UBI and private insurance protection mechanisms, UBI is a better option only if the verifiability of the causal relationship of unemployment due to AI is low or the probability of unemployment is highly sensitive to technological advances, because there is no suitable private insurance mechanism to cover all people in need of protection [82].
Critics of UBI argue that meaningful UBI is extremely expensive, and that in the case of the United States, it would cost more than the entire current U.S. federal budget. They fear that UBI could lead to the elimination of existing targeted welfare programs, ultimately harming more vulnerable populations. According to estimates, if each adult in the United States receives $1,000 per month, the cost of UBI will be 20% of the GDP of the United States in 2023 [83]. How much does it cost to implement UBI in China? Based on the data from the Seventh Population Census, the cost of UBI would be 15.4% of China's GDP in 2023, assuming that the working-age population aged 15 and above will receive 1,400 yuan per person per month, which is roughly one-fifth of the US standard (similar to the per capita GDP of both countries).
A more pragmatic approach would be to gradually expand the coverage and benefits of existing social security systems as AI brings productivity gains and economic growth. For example, government insurance programs may achieve better insurance outcomes than UBI or private insurance [84], and targeted unemployment benefits or retraining subsidies may be better able to help groups most affected by AI substitution, given the concentration of AI impact in the occupational distribution. For China, the existing social security system is still imperfect. For example, China's youth face vulnerabilities in the field of employment, and the current unemployment insurance system provides relatively low protection for them. Migrant workers in production, logistics and other occupations are vulnerable to the impact of AI, and society can further improve the help and protection of similar groups. Expanding the coverage of social security in China and improving the level and fairness of social security protection is even more important in the AI era.
[1] In this chapter, we would like to express our sincere gratitude for using the big data of Chinese job advertisements provided by Dr. Chen Qin, Chief Economist of Pulse Technology, in the process of measuring the exposure of Chinese occupations to AI. Researchers are also welcome to use our Chinese occupation-task database based on this big data and locally deployed AI models.
[2] OpenAI, GPT-4 Technical Report, 2023.
[3] News source: https://new.qq.com/rain/a/20240329A041XC00?. The four workflow techniques mentioned below are: Reflection: Allowing Agents to review and revise their own outputs; Tool use: let the large language model call APIs for actual operations; Planning: Break down complex tasks and let the agent execute as planned. Multi-agent collaboration: Multiple agents play different roles to complete tasks.
[4] Wei et al., Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, 2023.
[5] Trinh et al., Solving Olympiad Geometry Without Human Demonstrations, Nature, 2024.
Liu et al., Best Practices and Lessons Learned on Synthetic Data for Language Models, 2024.
[6]https://www.fastcompany.com/91011036/the-5-skills-workers-value-the-most-in-2024-according-to-new-research
[7] Jason Wei (OpenAI), Intuitions on language models, Stanford CS25 2024 Guest Lecture.
[8] OpenAI, Training language models to follow instructions with human feedback, 2022.
Google DeepMind, AlphaGo: Mastering the ancient game of Go with Machine Learning, 2016.
Lake et al., Building machines that learn and think like people, 2017.
Wellman and Gelman, Cognitive development: Foundational theories of core domains, 1992.
[9] OpenAI, GPT-4 Technical Report, 2023.
[10]资料来源：Stanford AI Index Report 2024.
[11] Google DeepMind, Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context, 2024.
[12] Papers With Code, a mainstream AI and machine learning research and evaluation platform, as an example, covers 10,888 benchmarks and 4,886 tasks as of the end of May 2024.
[13] Harvard Business Review, Managing AI decision-making tools, 2021.
[14] Autonomy in moral and political philosophy. Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/autonomy-moral/
[15]https://business.yougov.com/content/46597-more-than-half-of-global-public-now-worried-about-ai-replacing-jobs
[16]https://www.ipsos.com/en-ph/ai-making-world-and-most-asian-markets-nervous-about-job-security-ipsos-global-advisor-survey
[17]https://www.pewresearch.org/global/2018/09/13/in-advanced-and-emerging-economies-alike-worries-about-job-automation/
[18] Frey and Osborne, The future of employment: How susceptible are jobs to computerisation?, 2017. (Working paper version 2013).
Felten et al., A method to link advances in artificial intelligence to occupational abilities, 2018.
[19] Georgieff and Hyee, Artificial Intelligence and Employment: New Cross-Country Evidence, 2022.
Cazzaniga et al., Gen-AI: Artificial Intelligence and the Future of Work, IMF, 2024.
[20]https://www.onetcenter.org/overview.html
[21] Frey and Osborne, The future of employment: How susceptible are jobs to computerisation?, 2017. (Working paper version 2013).
[22] Michael Webb, The impact of artificial intelligence on the labor market, 2019.
Kogan et al., Technology and Labor Displacement: Evidence from Linking Patents with Worker-Level Data, 2023.
Sytsma and Sousa, Artificial Intelligence and the Labor Force: A Data-Driven Approach to Identifying Exposed Occupations, RAND 2023.
[23] Eloundou et al., GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models, OpenAI working paper 2023.
Gmyrek et al., Generative AI and jobs: A global analysis of potential effects on job quantity and quality, ILO working paper, 2023.
Joint Research Group of the National School of Development of Peking University and Zhaopin.com: Research Report on the Potential Impact of AI Large Models on the Labor Market in Mainland China, 2023.
[24] Autor and Dorn, The growth of low-skill service jobs and the polarization of the US Labor Market. American Economic Review, 2013.
Ellingrud et al., Generative AI and the future of work in America., MGI, 2023.
[25] Felten et al.,How will language modelers like chatgpt affect occupations and industries?, 2023.
Eloundou et al., GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models, OpenAI working paper 2023.
[26] Liepin data as of March 22, 2024, and Zhaopin's data as of April 11, 2024.
[27] Kogan et al., Technology and labor displacement: Evidence from linking patents with worker-level data, 2023.
Autor et al., New frontiers: The origins and content of new work, 1940–2018, QJE, 2024.
Acemoglu and Restrepo, The race between man and machine: Implications of technology for growth, factor shares, and employment, 2019.
World Economic Forum (WEF), Jobs of Tomorrow: Large Language Models and Jobs, 2023.
Pizzinelli et al., Labor market exposure to AI: Cross-country differences and distributional implications, 2023.
Gmyrek et al., Generative AI and jobs: A global analysis of potential effects on job quantity and quality, ILO, 2023.
[28] This article only lists the rating results of GPT-4 turbo (2024-04-09). We also scored using Claude 3, which is generally more conservative in its evaluation of the potential for AI alternatives than GPT-4.
[29] The three laws of robotics: 1. Robots must not harm humans, or deliberately fail to act so that humans are harmed, 2. Robots must obey commands given by humans, unless such orders would conflict with the first law, and 3. Robots must do their best to preserve themselves as long as they do not conflict with the first or second law. (See Yang Qiong, "Exploring the Ethical Principles of Robots", China Social Science Daily, No. 2196, 2021)
[30] Frey and Osborne, The future of employment: How susceptible are jobs to computerisation?, 2017. (Working paper version 2013).
Nedelkoska and Quintini, Automation, skills use and training, OECD 2018.
Council of Europe, Artificial intelligence and labour markets: friend or foe?, 2020.
[31] Gregory et al., Racing with or Against the Machine? Evidence on the Role of Trade in Europe, 2022.
[32] World Economic Forum, Future of Jobs Report 2023.
[33] Acemoglu and Restrepo, Automation and new tasks: How technology displaces and reinstates labor, 2019.
[34] Kogan et al., Technology and labor displacement: Evidence from linking patents with worker-level data, 2023.
[35] We first averaged the net employment impact of the SOC-6 level to the SOC-2 level, and used a large language model to map the occupational categories of the SOC-2 level with the occupational categories of the seventh national census to obtain the employment proportion of each SOC-2 occupational group.
[36] Autor and Dorn, The growth of low-skill service jobs and the polarization of the US labor market, 2013.
Frey and Osborne, The future of employment: How susceptible are jobs to computerisation?. Technological forecasting and social change, 2017.
[37] Autor et al., The polarization of the US labor market, American Economic Review, 2006.
Böhm, M. J., The price of polarization: Estimating task prices under routine-biased technical change, 2020.
Böhm, et al., Occupation growth, skill prices, and wage inequality, 2019.
[38] International Labour Organization, Work for a brighter future–Global Commission on the future of work, 2019.
International Labour Organization, Global Wage Report 2022-23: The impact of inflation and COVID-19 on wages and purchasing power, 2022.
[39] Brynjolfsson and Unger, Artificial intelligence, IMF F&D, 2023.
Zarifhonarvar, Ali, Economics of chatgpt: A labor market view on the occupational impact of artificial intelligence, 2023.
[40] Ellingrud et al., Generative AI and the future of work in America, MGI 2023.
Felten et al., How will language modelers like chatgpt affect occupations and industries?, 2023.
[41] Kogan et al., Technology and labor displacement: Evidence from linking patents with worker-level data, 2023.
[42] Aghion et al., A theory of falling growth and rising rents, 2019.
[43] Li Jing et al., "Artificial Intelligence, Labor Task Types and City-Scale Wage Premium", Journal of Finance and Economics, No. 12, 2023. He Qin and Liu Mingze, "The Impact of Artificial Intelligence on Employment Scale and Labor Income", Journal of Capital University of Economics and Business, No. 4, 2023.
[44] Krusell et al., Capital‐skill complementarity and inequality: A macroeconomic analysis, Econometrica, 2000.
[45] Karabarbounis and Neiman, The global decline of the labor share, QJE, 2014.
[46] Elsby et al., The decline of the US labor share, Brookings papers on economic activity, 2013.
[47] Arpaia et al., Understanding labour income share dynamics in Europe, 2009.
[48] Oberfield et al., Micro data and macro technology, Econometrica, 2021.
[49] Autor et al., The Fall of the Labor Share and the Rise of Superstar Firms, QJE, 2020.
[50] Piketty and Goldhammer, The Capital- Labor Split in the Twenty- First Century, Harvard University Press, 2014.
[51] Korinek and Stiglitz, Artificial Intelligence and Its Implications for Income Distribution and Unemployment. NBER, 2017.
[52] Aghion et al., Artificial Intelligence and Economic Growth, NBER, 2017.
[53] Acemoglu and Restrepo, The race between man and machine: Implications of technology for growth, factor shares, and employment, 2018.
[54] The impact of labor-saving technologies will reduce the share of labor force by 2.5% cumulatively over the next five years, while the impact of labor-enhancing technologies will increase by 0.75% (but since it is not statistically significant, the next step is calculated as 0).
Kogan et al., Technology and labor displacement: Evidence from linking patents with worker-level data, 2023.
[55] Gary S Becker, Human Capital: A Theoretical and Empirical Analysis with Special Reference to Education, NBER, 1994.
[56] Acemoglu and Restrepo, Artificial intelligence, automation and work, NBER, 2019.
[57] MGI, Skill shift: Automation and the future of the workforce, 2018.
[58] ILO, Non-standard employment around the world: Understanding challenges, shaping prospects, 2016.
[59] A.team, Survey: How AI Boosts the Productivity and Earnings of Top Tech Freelancers, 2024/05/06.
[60]https://www.gov.cn/xinwen/2021-05/20/content_5609599.htm
FTChinese: "Gig Economy Workforce in England and Wales Reaches 4.5 Million", 5 November 2021.
Wei Shangjin, "Let the "New Employment Form" Illuminate the Future", Fudan Financial Review, No. 16.
[61] Zhou Chang, "China's Digital Labor Platform and the Protection of Workers' Rights and Interests", ILO Work Report, November 2020.
[62] Jia Donglan and Zhu Huilin, "Actively Promoting the Protection of Workers' Remuneration Rights and Interests in New Employment Forms", Chinese Human Resources and Social Security, No. 2, 2024.
[63] G20, Providing adequate and sustainable social protection for workers in the gig and platform economy, 2023.
[64] Demirci et al., Who is AI Replacing? The Impact of Generative AI on Online Freelancing Platforms, SSRN, 2023.
[65] A.team, Survey: How AI Boosts the Productivity and Earnings of Top Tech Freelancers, 2024/05/06.
[66] Wang Yong, "New Employment Forms: From High Efficiency to Long Effect", Fudan Financial Review, vol. 16, 2023.
[67] Cai Jiming, "The Path Choice of Getting Out of the Dilemma of Flexible Employment and Social Security", Social Security Review, No. 1, 2024.
[68] ILO, Non-standard employment around the world: Understanding challenges, shaping prospects, 2016.
[69] National Bureau of Statistics, "2023 Migrant Worker Monitoring Survey Report," April 30, 2024.
[70] CAFF50: China Rural Ageing Finance Survey Report 2022, December 2022.
[71] In contrast, the impact of minimum wages, trade unions, and collective bargaining on employment and income, as well as the trade-offs, are more controversial in academic circles.
[72] WEF, The Future of Jobs, 2016.
[73] OECD Skills Strategy 2019, Chapter 4. Developing relevant skills over the life course.
[74] Padmaja and Mukul, Upskilling and reskilling in the digital age: the way forward for higher educational institutions, 2021.
[75] Illanes et al., Retraining and reskilling workers in the age of automation, 2018.
[76]https://www.npr.org/2023/01/31/1152652093/ai-artificial-intelligence-bot-hiring-eeoc-discrimination; https://www.shrm.org/topics-tools/news/employers-embrace-artificial-intelligence-hr
[77] Korinek and Stiglitz, Artificial Intelligence and Its Implications for Income Distribution and Unemployment, NBER, 2017.
[78] Lewis Silkin, Robot tax: the pros and cons of taxing robotic technology in the workplace, 2018.
[79] Schaefer and Schneider, Public Policy Responses to AI, Graz Economics Papers, 2024.
[80] Korinek and Stiglitz, Artificial Intelligence and Its Implications for Income Distribution and Unemployment, NBER, 2017.
[81]https://www.futureofworkhub.info/comment/2019/12/4/robot-tax-the-pros-and-cons-of-taxing-robotic-technology-in-the-workplace
[82] Schaefer and Schneider, Public Policy Responses to AI, Graz Economics Papers, 2024.
[83] Ocampo, José Antonio, and Joseph E. Stiglitz, eds. The welfare state revisited. Columbia University Press, 2018.
[84] Schaefer and Schneider, Public Policy Responses to AI, Graz Economics Papers, 2024.
Article source:
This article is excerpted from: "Chapter 4 Substitution and Enhancement, Transforming the Job Market", which was published on June 28, 2024
赵扬 分析员 SAC 执证编号：S0080521080006 SFC CE Ref：AZX409
吴晓慧 分析员 SAC 执证编号：S0080524010007 SFC CE Ref：BTK904
Legal Notices