感谢关注China Economic Review（CER）推送公众号。CER作为一本聚焦中国经济及其与世界经济关系的原创研究成果的期刊，在经济学和金融学类期刊中获得稳步提升的影响力。
CER推送作为作者与读者的沟通平台，将定期介绍发表于CER上的文章。本次推送将介绍陈沁，葛劲峰，谢华庆，徐兴成和杨燕青于2025年发表在CER上的文章Large Language Models at Work in China's Labor Market。
Recent remarkable progress in the field of generative AI and large language models has provoked many pressing questions about the effects of these powerful technologies on the economy. The most significant question surrounding advances in generative AI and LLMs is the impact these technologies will have on the dynamics of the labor market.
近年来，生成式人工智能和大语言模型领域的显著进展引发了人们对这些强大技术的经济影响的广泛关注。其中最为重要的一个问题是生成式人工智能和大语言模型对劳动力市场动态的影响。
A branch of research emphasizing the disruptive labor market impacts of LLMs is emerging rapidly, however, it predominantly focuses attention on the labor market in developed economy, in particularly the U.S. This paper analyzes the potential impacts of LLMs on China’s labor market. We employ three large language models (GPT-4, GLM, InternLM) as classifiers to determine the occupational exposure based on the detailed description for each occupation contained in the general code of occupational classification of the People‘s Republic of China.
当前，已有不少研究开始关注大语言模型对劳动市场的冲击效应，但这些研究大多聚焦于发达经济体，尤其是美国。本文则聚焦于大语言模型对中国劳动力市场的潜在影响。我们利用三种大语言模型（GPT-4、GLM、InternLM）作为职业分类工具，根据《中华人民共和国职业分类大典》中各职业的详细描述，对职业暴露程度进行测定。
The analysis reveals significant heterogeneity in occupational exposure, with more educated, higher paid, white-collar occupations being the most exposed to LLMs. These results align with recent studies on the US labor market, which also found that technological advancements disproportionately impact workers at the higher end of the wage spectrum. Beyond the positive correlation between wage, education, and occupational exposure, our findings indicate a positive correlation between experience premiums and exposure to LLMs, suggesting potential diminishing returns to “learning by doing” in the future. Industry-level analyses reveal that education and healthcare have a higher level of exposure, while manufacturing, agriculture, mining, and construction show lower exposure. In contrast to other developed countries, the uneven age distribution across industries in China amplifies the demographic exposure to LLMs, disproportionately affecting younger workers. Furthermore, LLMs significantly impact labor demand. A positive correlation between vacancy shares and occupational exposure scores suggests that labor demand structures may exacerbate the disruptive impacts of LLMs on China‘s labor market. The positive correlation between the growth rate of vacancy shares and exposure scores further indicates a potential reversal in labor demand trends. Contrary to expectations, China’s economic and labor market structure intensifies rather than mitigates the disruptive effects of LLMs.
研究发现，各职业对大语言模型的暴露程度存在显著异质性：受过更高教育、薪酬较高的白领职业暴露程度最高。这一结果与近期美国劳动市场的相关研究结论一致，即技术进步对高薪职业的冲击程度更大。除了工资和教育水平与职业暴露程度之间存在正相关外，我们还发现经验溢价与暴露程度之间也呈正相关，这表明未来可能出现“干中学”收益递减的趋势。行业层面的分析显示，教育和医疗行业的职业暴露程度较高，而制造业、农业、采矿业及建筑业的暴露程度则相对较低。与其他发达国家不同，中国不同行业的劳动力年龄分布存在较大差异，导致人口结构对大语言模型的暴露效应被进一步放大，尤其是年轻劳动力受到的影响更为显著。此外，大语言模型对劳动力需求也有显著影响。职位空缺比例与职业暴露程度呈正相关，表明劳动力需求结构可能进一步加剧大语言模型对中国劳动力市场的冲击效应。而职位空缺比例增长率与暴露程度的正相关关系则进一步暗示劳动力需求趋势可能发生逆转。与预期相反，中国的经济结构与劳动力市场结构可能会放大而非减轻大语言模型的冲击。
To deepen these insights, we employed advanced language models to evaluate the quintet of occupations: Non-routine Cognitive (Analytical), Non-routine Cognitive (Interpersonal), Routine Cognitive, Routine Manual, and Non-routine Manual Physical. The results indicate that non-routine cognitive skills, particularly analytical ones, are most significantly affected by LLMs. Routine manual skills are also substantially influenced, whereas interpersonal and non-routine manual physical skills remain relatively unaffected. This predominant impact on non-routine cognitive tasks deviates from the “routinization hypothesis”, which posited that ICT primarily automates routine, codifiable tasks performed by middle-wage workers.
为了进一步探讨这一问题，我们使用先进的大语言模型评估了以下五类职业：非常规认知（分析型）、非常规认知（人际交往型）、常规认知、常规体力和非常规体力技能。结果表明，LLMs对非常规认知技能的冲击最为显著，尤其是分析类任务。同时，常规体力任务也受到显著影响，而人际交往类和非常规体力任务则相对不受影响。这种对非常规认知任务的主导性冲击与传统的“常规化假设”（Routinization Hypothesis）产生了明显偏离，后者认为信息通信技术（ICT）主要自动化中等薪资劳动者所从事的常规性、可编码任务。
Motivated by the unique occupational exposure of LLMs compared to earlier technologies, we introduce a novel theoretical model to examine why different technologies produce different occupational exposure structures. This model incorporates entropy-based information theory into a stylized task-based framework, which is particularly effective for characterizing comparative advantage. By defining task complexity using entropy and modeling the relative productivity of skills across occupations, we can systematically assess the efficiency of AI technologies, including LLMs, across various tasks. The model also integrates KL-divergence to represent the relative efficiency of AI models across occupations, distinguishing the automation logic of traditional AI from that of deep learning-driven LLMs. This represents a significant advancement in AI technology.
基于大语言模型技术的职业暴露特征显著区别于早期技术的这一独特现象，我们进一步提出了一种新的理论模型，用以解释为何不同的技术会产生不同的职业暴露结构。该模型在任务导向分析框架中融入了基于熵（entropy）的信息理论，尤其适合用于刻画比较优势。通过以熵来定义任务的复杂性，并建模不同职业技能的相对生产力，我们能够系统地评估包括大语言模型在内的AI技术在各类任务上的有效性。此外，该模型还引入了KL散度（KL-divergence）以表达AI模型在不同职业间的相对效率，从而区分了传统人工智能与基于深度学习的大语言模型的自动化逻辑，这代表了人工智能技术的一次重要进步。
The utopian vision of the information age anticipated that computers would democratize information and flatten economic hierarchies. However, this vision has not materialized; instead, the opposite has occurred. Labor income inequality has widened significantly, with computerization facilitating the unprecedented concentration of decision-making power among elite experts, who now have access to abundant and inexpensive information. Our theoretical model predicts that artificial intelligence, by leveraging vast datasets and computational power to synthesize information and rules, could enable a broad range of workers equipped with foundational training to perform high-stakes decision-making tasks. These tasks, traditionally monopolized by elite experts — such as doctors, lawyers, software engineers, and college professors — may become accessible to more individuals. In this regard, our theoretical perspective suggests that artificial intelligence has the potential to reduce income inequality, contrary to the trends observed with earlier technologies.
信息时代时曾催生过一种理想化愿景，人们期望计算机能实现信息民主化、削平经济等级。然而，这一愿景并未实现，反而产生了相反的结果。劳动收入不平等显著扩大，计算机化使精英专家获得前所未有的决策权力，这些精英专家掌握了丰富而廉价的信息资源。我们的理论模型预测，人工智能技术利用庞大的数据集与计算能力合成信息和规则，有望使大量接受过基础训练的普通劳动者能够执行高风险的决策任务。这类任务以往被医生、律师、软件工程师和大学教授等精英专家所垄断，而如今可能变得更广泛可及。从这一理论视角看，人工智能有望缓解收入不平等，这与此前技术发展的趋势相反。
In terms of policy implications, our theoretical findings for artificial intelligence should be viewed as scenario analyses rather than definitive predictions. The impact of any technology depends not only on its inherent features but also on institutional responses. The first policy implication of our theoretical model is the disruptive impact of advanced AI technologies, such as LLMs, on expert labor. AI has the potential to reduce scarcity by empowering more workers to perform expert-level tasks. However, this potential will not be realized automatically. Thus, a key policy orientation for the future is to develop training programs that enable workers to effectively utilize AI tools. The second policy implication concerns the substantial uncertainties surrounding the technological impacts on labor markets, as evidenced by the unrealized utopian vision of the information age. To design effective policy tools, it is crucial to systematically track labor market demand conditions. Unlike the U.S., which has a system like ONET, China currently lacks a comprehensive statistical framework to monitor occupational demand. Establishing such a system should be a priority, given the impending disruptive waves of AI technologies.
在政策含义方面，我们的理论结论应被视作情景分析而非明确预测。任何技术的影响不仅取决于技术本身，还取决于制度性应对措施。我们的理论模型首先提示了先进AI技术（如LLMs）对专业型劳动力的冲击，指出AI技术有可能通过赋能更多劳动者执行专家级任务来降低专业型人才的稀缺性。然而，这种潜力不会自动实现。因此，未来的关键政策导向之一是发展相关培训项目，帮助劳动者有效地使用AI工具。第二个政策含义涉及技术对劳动力市场影响的不确定性，这一点在信息时代未能实现的乌托邦愿景中已有体现。为了制定有效的政策工具，系统性地追踪劳动力市场的需求条件至关重要。与拥有ONET职业信息系统的美国不同，中国目前尚未建立全面的职业需求统计框架。鉴于人工智能技术的迅猛发展，建立类似体系应成为政策制定的优先事项。
As a general-purpose technology, the introduction and proliferation of large language models constitute a substantial technological upheaval with significant implications for the overall economy. This paper employs measures of occupational exposures to LLMs, in conjunction with aggregate assessments of occupational composition, to evaluate the potential impact of LLMs on labor market within China’s economy. Furthermore, we develop a theoretical model to provide a deeper understanding of the reasons behind occupational exposure to LLMs, which diverges from the prediction of the routinization hypothesis.
作为一种通用目的技术（general-purpose technology），大语言模型的出现和推广势必带来显著的技术冲击，并对整个经济产生深远影响。本文运用职业暴露测度指标，结合对职业构成的总体评估，考察了大语言模型对中国劳动力市场的潜在影响。此外，我们构建的理论模型深化了对大语言模型的职业暴露原因的理解，这种理解不同于传统常规化假设的预测。
作者简介：
陈沁：上海脉策数据科技有限公司首席经济学家，主要研究方向为大模型在经济学研究中的各类应用。
葛劲峰（通讯作者）：上海海事大学经济管理学院副教授。主要研究方向为中国宏观经济、创新与经济增长，宏观劳动力市场。
谢华庆：上海人工智能实验室研究员，主要研究方向为AI for Science，人工智能和创新经济学。
徐兴成（通讯作者）：上海人工智能实验室研究员，主要研究方向为人工智能，大模型后训练、对齐与复杂推理，及其应用。
杨燕青（通讯作者）：上海科学智能研究院首席战略官，复旦大学兼职教授，上海市统计学会常务副会长，中国计算机学会计算经济学专业组执委，中国国际金融学会常务理事。获得复旦大学经济学博士学位。研究领域包括人工智能经济学、人工智能战略及应用、数字经济和全球治理。