【观点】蔡昉：在变与不变中寻策

变与不变——人工智能的就业影响

蔡 昉

人们习惯于在始终存在、反复出现、一次又一次得到解决的经济史现象面前，表现出盲目的乐观态度。面对某些事件的积累，例如，金融危机之前的非理性行为和泡沫积累，盲目乐观者会宣称“这次不一样了”；在面对另一些事件时，人们往往以“这在以前发生过”作为盲目乐观的理由，以致不能在认识上和行动中做到与时俱进。人工智能（AI）对就业的影响，就属于后一种情形。凡事预则立。面对可能的AI就业冲击，需要澄清一些认识，确立若干政策原则。

一、这一次真的不一样

技术进步对于就业的影响，从工业革命开始就是社会的焦点和研究的关注点。具有“卢德主义”性质的运动和思潮，曾经以各种面貌反复出现。不过，无论是从原因穷究结果，还是从本质看到表象，这一次是真的不一样了。

首先，这一次不再是历史上反反复复出现的“技术性失业”幽灵（spectre），而是可以替代几乎所有职业的终结者（terminator）。从懂科学的企业家马斯克，到关心AI发展的经济学家萨默斯，都认为AI对岗位的替代将是全面的，一旦不久后通用人工智能（artificial general intelligence，AGI）出现，简单的、复杂的、体力的、智力的，无论何种岗位将无一幸免。

其次，AI技术进步的速度之快，越来越具有一日千里、一日三秋的感觉。例如，从“土耳其下棋机器人”（1770年骗局，可将其权且当作这个想法的起点）到图灵1950年论文发表，经过了180年；再到1997年“深蓝”战胜卡斯帕罗夫，又经历47年；再到名为“阿尔法狗”的国际象棋机器人于2016年战胜李世石、于2017年战胜柯洁，也相隔了约20年。而从ChatGPT问世到Sora的出炉，仅仅相隔一年。我们无须用任何复杂的模型来预测，只要看一看这个速度和加速度，即可得出对通用人工智能出现的合理预期。 CXO UNION-CXO联盟（cxounion.cn）

最后，大模型AI的“发展悖论”注定了岗位的大规模丧失几乎是必然的。阵营之间、国家之间、企业之间都认识到占据AI技术和产业的制高点，关乎生死存亡。这导致围绕着AI的发展，形成一种类似冷战时期太空竞赛、军备竞赛、核武器竞赛的竞争。并且，大模型AI的高度耗能、“烧钱”（如ChatGPT-4花了4亿美元，据说ChatGPT-5需要25亿美元）。挖掘模型用途、扩大用户群、提高回报率的必然方向和方式，便是提高劳动生产率，从而减少劳动力和人力资本的使用。

二、哪些东西没变？

然而，只要人类劳动还没有彻底由人工智能替代或者决定，或者说“人机一体”尚未普遍实现之前，就仍有一些东西不会发生变化。而且，这些没变的事物或方面越发弥足珍贵，可以为我们提供一个时间窗口。

最重要的一点就是，人仍然是主导的一方，仍然是“人告诉机器做什么”，这是使我们保持信心的根本。这一点既有技术上的涵义，也有制度上的涵义。也就是说，我们人类应对岗位替代的两条根本出路，迄今尚未发生根本性的变化，虽然也需要与时俱进，不断校正方向。 CXO UNION-CXO联盟（cxounion.cn）

第一，人力资本依然是抵御AI冲击的底气，但是人类需要知道自身的所长和所短，把扬长避短作为AI时代人力资本培养的基本策略。迄今为止，人类智能或自然智能相对于AI，仍然具有优势的方面在于：（1）软技能而非硬技能；（2）非认知能力而非认知能力；（3）情商而非智商；（4）人文的理解力和同理心，而非数理化的解题能力，甚至不是编码技能；（5）隐性的知识（tacit knowledge）而不仅是显示性的技能。

第二，社会福利体系仍然是根本性的托底制度，而且履行此类功能的物质条件日益增强。马克思从早期资本主义的发展看到，一旦劳动力成为商品，从制度上工人便难以摆脱受剥削的命运。北欧在建立福利国家之初，在制度设计中便突出“去商品化”，即弱化劳动力作为纯私人要素的属性，强化劳动者及其家庭的社会权利。在AI的“岗位破坏”日益大于和快于“岗位创造”的条件下，这个理念和做法越来越重要。

三、就业对AI的反应方式及结果

无论是老办法还是新思路，就业对AI替代做出反应的方式，不外以下几种。在概括这些方式之前，我们先给出一个合理的预设前提，即AI的发展终究会以前所未有的幅度提高劳动生产率。在此基础上，劳动者通常并且可以有以下几种出路。 CXO UNION-CXO联盟（cxounion.cn）

第一，转入更高质量的岗位。这是乐观的经济学家始终坚信的一种结果，自从历史上发生“卢德主义运动”以来，也不断被事实所证明。只不过这要求劳动者具有更高的技能与之相适应。换句话说，获得这种新岗位的与失去旧岗位的，通常不是同一批人，很大程度上也不是同一队列的人，甚至不是同一代人。

今后，失去老工作和得到新工作的时间缺口只会更大。包括美国前财长萨默斯在内的许多经济学家，已经从以前对技术进步创造岗位充满信心，转变为如今认为卢德主义自有其道理。鉴于另一位美国前财长姆努钦对AI的就业影响仍然“乐观”，并且如今已经难得找到持这种态度的人了，我们可以称这种似在虚无缥缈之中的岗位为“姆努钦式岗位”。

第二，转到具有“逆库兹涅茨化”特征的岗位上。这是指那些劳动生产率较低，从而报酬也较低的行业。从客观上说，新岗位的正规化程度要低于原来的工作。从主观上说，新岗位的体面程度也要低于原来的工作。总而言之，就业质量被降低。

当美国经济学家罗伯特•索洛提出何以处处可见计算机，生产率却未见提高的疑问时，就触及这种现象的本质。也就是说，当AI必然提高一些领域的生产率时，另一些行业则变得更“卷”，以更低的人均产出为代价吸纳转岗人员。故我们可以称之为“逆库兹涅茨岗位”。

第三，转到具有“鲍莫尔成本病”性质的岗位上。这是指那些人们保持着巨大的需求，却天然具有劳动生产率难以提高特性的行业。经济学家威廉•鲍莫尔把表演艺术作为这种行业的典型例子。无论如何，这种类型的行业和岗位能否继续存在，以及能否得以扩大的核心，在于人们对相应产品和服务的需求及其弹性。显而易见，这类岗位可以被称为“鲍莫尔成本病岗位”。 CXO UNION-CXO联盟（cxounion.cn）

第四，转到由新的消费所诱致出来的岗位上。我们今天的消费内容，在若干年之前可能难以想象，在更早的时候索性就不存在。就业岗位也是如此。未来随着劳动生产率的提高，人们的品味在变化，新事物新观念不断涌现，因而消费的领域不断拓展，职业类型花样翻新。鉴于这类岗位的消费诱因，归根结底由供给侧生产率的提高引起，是一种“供给创造需求”现象，我们可以称之为“萨伊式的岗位”。

第五，转到因重新定义而出现的岗位上。以前不符合就业定义的活动，如今在整体劳动生产率的支撑下，可以被社会承认为“就业”，并以转移支付的方式得到补偿，则可以被认为是就业。例如，如果一个人自认为是“作家”却没有作品出版并获得酬劳，按照失业的调查定义，这种“在过去一周内未从事一小时以上有报酬工作”的状态，则不被算作就业。然而，如果社会负担得起，也完全可以认为这是一种就业。

与此相类似的情形还包括那些并不宣称自己正在“工作”的人，即不再寻求就业的人群。这包括两种情形。一种是当事人有供养来源，例如索性采取啃老等方式“躺倒”。另一种是无需就业，却可以得到普惠性的社会福利支撑。例如，如果实施全民基本收入（universal basic income）制度，就形成一种环境，使受到就业冲击的一些人选择不再参与传统意义上的工作。鉴于这与凯恩斯1930年著名的“我们孙辈的经济可能性”一文提出的命题有关，我们可以称之为“凯恩斯式岗位”。 CXO UNION-CXO联盟（cxounion.cn）

四、在变与不变中寻策

根据经济史上人类长期应对技术替代就业现象的经验，可以提出几个原则性政策建议，即通过制度建设、政策调整、体制改革引导技术发展以及市场主体行为，尽量做到几个“同步”：

一是保持岗位破坏速度与岗位创造速度的同步性，特别是在数量上最大程度使转岗具有可行性。

二是保持各行业生产率提高速度的同步性，避免索洛悖论情形的发生。 CXO UNION-CXO联盟（cxounion.cn）

三是保持AI替代劳动力的速度与培训劳动者能力的速度同步性，尽可能缩短再就业摩擦期。

四是保持生产率提高与生产率分享的同步性，这也是公平与效率统一的要求和体现。

从政府职能的角度保障落实以上原则，可以从若干既重要且紧迫的应对之策入手。

首先，加快建设中国式福利国家。对此应该强调几点：

第一，以只争朝夕的精神，或者说以摩尔定律的速度加快完善社会福利体系。

第二，按照普惠的原则设计和完善福利制度。这意味着改变以往严格识别社会福利受益对象的理念，因为在岗位的加速流失时代，已经越来越无法区分一个人是否“躺倒”，而且AI驱逐劳动者本身具有强烈的外部性。

第三，用社会共济、社会保护和权益保障，抵消非正规就业的蔓延趋势及其对劳动者的不利影响。

其次，大幅度延长义务教育或免费教育年限。与AI竞争的需要，一方面对人力资本提出越来越高的要求，另一方面需要更偏重非认知能力的培养。哈佛大学儿童发展中心的研究显示，在人生的最初几年，大脑每秒钟能够建立超过100万个神经元连接，这在此后任何生命阶段都无法重现。非认知能力的最佳培养时间在三岁和四岁，最理想的举措是把义务教育延长到这个学前教育年龄。何况，劳动生产率的预期大幅度提高，可以显著扩大教育公共资源，足以支撑更长的儿童在校时间。 CXO UNION-CXO联盟（cxounion.cn）

最后，消除在儿童发展、教育与培训、流动与就业、社会保障，以及其他基本公共服务方面存在的制度性障碍。其中特别是消除流动儿童和留守儿童存在的制度原因。研究表明，对于孩子的人力资本培养，特别是对于非认知能力的获得，以及他们终生的社会流动机会，父母的养育和照护具有学校和社会均无法替代的作用。因此，推动户籍制度改革，解决留守儿童和流动儿童面临的此类问题，应该置于最高的优先序。

翻译：

CAI Fang: Seeking solutions between change and invariance

Change and unchange-the employment impact of artificial intelligence

CAI Fang

People are accustomed to blind optimism in the face of ever-present, recurring, and time-and-again solved phenomena of economic history. Faced with the build-up of events, such as the build-up of irrationality and bubbles that preceded the financial crisis, blind optimists declare that “this time is different”; In the face of other events, people often use “this has happened before” as a reason for blind optimism, so that they fail to keep pace with The Times in understanding and action. The impact of artificial intelligence (AI) on employment falls into the latter category. Foreknowledge ensures success. In the face of the possible impact of AI employment, some understanding needs to be clarified and several policy principles need to be established. CXO UNION-CXO联盟（cxounion.cn）

1. This time is really different

The impact of technological progress on employment has been the focus of society and research since the Industrial Revolution. Movements and trends of thought with the nature of “Luddism” have repeatedly appeared in various forms. However, whether it is from the cause to investigate the result, or from the essence to see the appearance, this time is really different.

First, this time it is not the spectre of technological unemployment, which has recurred throughout history, but the terminator, which could replace almost every occupation. From Musk, an entrepreneur who understands science, to Summers, an economist who cares about the development of AI, they all believe that AI’s replacement of jobs will be comprehensive, and once artificial general intelligence (AGI) appears soon, it is simple, complex, physical, and intellectual. No job will be spared.

Secondly, the speed of AI technology progress is fast, and more and more have the feeling of thousands of miles a day and three days a day. For example, 180 years elapsed between the “Turk chess-playing robot” (a hoax of 1770 that may be considered the starting point for the idea) and the publication of Turing’s paper in 1950; Another 47 years passed until Deep Blue defeated Kasparov in 1997. It is also about 20 years since a chess robot named Alpha Dog defeated Lee Sedol in 2016 and Ke Jie in 2017. From ChatGPT to Sora came out, only one year apart. We don’t need to use any complex model to predict, just look at this speed and acceleration to come up with a reasonable expectation of the emergence of general artificial intelligence. CXO UNION-CXO联盟（cxounion.cn）

Finally, the “development paradox” of large model AI predestined the large-scale loss of jobs to be almost inevitable. Between camps, between countries, and between enterprises, they all realize that occupying the commanding heights of AI technology and industry is a matter of life and death. This has led to the development of AI, creating a competition similar to the Cold War space race, arms race, and nuclear weapons race. Moreover, large model AI is highly energy-intensive and “money burning” (for example, ChatGPT-4 cost $400 million, and ChatGPT-5 is said to cost $2.5 billion). The inevitable direction and way to explore the use of models, expand the user base, and improve the rate of return is to increase labor productivity, thereby reducing the use of labor and human capital.

2. What hasn’t changed?

However, as long as human labor has not been completely replaced or determined by artificial intelligence, or “man-machine integration” has not been widely realized, there are still some things that will not change. Moreover, these things or aspects that have not changed become more precious and can provide us with a window of time. CXO UNION-CXO联盟（cxounion.cn）

The most important point is that people are still the dominant party, and it is still “man telling the machine what to do”, which is fundamental to keeping our confidence. This has both technical and institutional implications. In other words, the two fundamental ways for us humans to cope with job replacement have not fundamentally changed so far, although they also need to keep pace with The Times and constantly correct the direction.

First, human capital is still the base to resist the impact of AI, but human beings need to know their own strengths and weaknesses, and take promoting strengths and avoiding weaknesses as the basic strategy of human capital training in the AI era. So far, human intelligence or natural intelligence still has an advantage over AI in the following aspects: (1) soft skills rather than hard skills; (2) non-cognitive abilities rather than cognitive abilities; (3) EQ rather than IQ; (4) Human understanding and empathy, not mathematical problem-solving skills, not even coding skills; (5) tacit knowledge is not only a tacit skill. CXO UNION-CXO联盟（cxounion.cn）

Second, the social welfare system remains a fundamental underpinning system, and the material conditions for fulfilling such functions are increasingly enhanced. From the development of early capitalism, Marx saw that once labor force became a commodity, it was difficult for workers to escape from the fate of exploitation. At the beginning of the establishment of the welfare state in Northern Europe, “de-commodification” was highlighted in the system design, that is, weakening the attribute of labor force as a purely private element and strengthening the social rights of workers and their families. In the context of AI’s “job destruction” increasingly greater than and faster than “job creation” conditions, this concept and practice are increasingly important.

3. How employment responds to AI and its results

Whether old or new, there are several ways in which employment will respond to AI substitution. Before summarizing these approaches, let’s make a reasonable presupposition that the development of AI will eventually increase labor productivity by an unprecedented degree. On this basis, workers usually have and can have the following ways out.

First, move into higher-quality jobs. This is an outcome that optimistic economists have always believed in, and which has been repeatedly proved by facts since the “Luddite movement” occurred in history. It just requires workers to be more skilled to match it. In other words, the people who get these new jobs and the people who lose the old ones are usually not the same people, not in the same cohort for the most part, not even in the same generation. CXO UNION-CXO联盟（cxounion.cn）

In the future, the gap between losing an old job and getting a new one will only widen. Many economists, including former US Treasury Secretary Larry Summers, have moved from their faith in the job creation of technological progress to the view that Luddism has its own merit. Given that Mnuchin, another former US Treasury secretary, is still “optimistic” about the employment impact of AI, and that it is now hard to find anyone with such an attitude, we can call this position in the air “Mnuchin position.”

Second, move to positions characterized by “reverse Kuznets.” This refers to those industries where labor productivity is low and therefore pay is low. Objectively speaking, the degree of formalization of the new position is lower than that of the original job. Subjectively speaking, the new job is also less respectable than the old one. In short, the quality of employment is reduced.

The US economist Robert Solow got to the essence of this phenomenon when he asked why computers are everywhere but productivity is not rising. That is, while AI is bound to increase productivity in some areas, other industries are becoming more “curly,” absorbing transfer workers at the cost of lower output per capita. So we can call it the “anti-Kuznets post.”

Third, move to a position with the nature of “Baumol cost disease”. This refers to those industries where people maintain a huge demand, but it is naturally difficult to improve labor productivity. The economist William Baumol used the performing arts as a prime example of such an industry. In any case, the survival and expansion of these types of industries and jobs lies at the heart of the demand and elasticity of the products and services. Obviously, this type of job can be called “Baumol cost disease job.”

Fourth, move to jobs induced by new consumption. The content we consume today might have been unimaginable a few years ago, but simply did not exist earlier. The same is true for jobs. In the future, with the improvement of labor productivity, people’s tastes are changing, new things and new ideas continue to emerge, so the field of consumption is constantly expanding, and the type of occupation is renovated. In view of the consumption incentives of such jobs, in the final analysis, caused by the improvement of productivity on the supply side, is a phenomenon of “supply creates demand”, we can call it “Say type of jobs”. CXO UNION-CXO联盟（cxounion.cn）

Fifth, move on to roles that have emerged as a result of redefinition. Activities that did not previously meet the definition of employment but can now be recognized as “employment” by society on the back of overall labor productivity and compensated in the form of transfer payments can be considered employment. For example, if a person identifies as a “writer” but is not paid for his or her work, the status of “not engaged in more than one hour of paid work in the past week” is not counted as employment, according to the survey definition of unemployment. However, if society can afford it, it can also be considered a form of employment.

The same is true of those who do not claim to be “working,” that is, those who no longer seek employment. There are two scenarios. One is that the parties have sources of support, such as simply “lying down” by eating the elderly. The other is that you do not need to work, but you can get inclusive social welfare support. For example, if a universal basic income is implemented, it creates an environment in which some people who are hit by employment choose not to work in the traditional sense. Given that this relates to Keynes’s proposition in his famous 1930 essay “The economic possibilities of Our grandchildren,” we might call it a “Keynesian job.” CXO UNION-CXO联盟（cxounion.cn）

4. Find solutions in change and invariance

According to the long-term experience of human beings in dealing with the phenomenon of technology replacing employment in economic history, several principle policy suggestions can be put forward, that is, to guide technological development and the behavior of market players through system construction, policy adjustment and institutional reform, and try to achieve several “synchronization” :

The first is to maintain the synchronization of the rate of job destruction and the rate of job creation, especially in terms of the number of jobs to maximize the feasibility of job transfer.

The second is to maintain the synchronization of productivity improvement rate of various industries to avoid the occurrence of Solow paradox. CXO UNION-CXO联盟（cxounion.cn）

The third is to maintain the speed of AI labor replacement and the speed of training workers, and shorten the period of re-employment friction as much as possible.

The fourth is to maintain the synchronization of productivity improvement and productivity sharing, which is also the requirement and embodiment of the unity of fairness and efficiency.

To ensure the implementation of the above principles from the perspective of government functions, we can start with several important and urgent countermeasures.

First, accelerate the construction of a Chinese-style welfare state. Several points should be emphasized in this regard:

First, accelerate the improvement of the social welfare system in the spirit of seizing the day and night, or in the speed of Moore’s Law. CXO UNION-CXO联盟（cxounion.cn）

Second, we should design and improve the welfare system in accordance with the principle of universal benefits. This means changing the past concept of strictly identifying who benefits from social welfare, because in the era of accelerating job loss, it is increasingly impossible to distinguish whether a person is “lying down”, and AI expulsion of workers itself has strong externalities.

Third, social solidarity, social protection and protection of rights and interests should be used to offset the spread of informal employment and its negative impact on workers.

Second, significantly extend the period of compulsory or free education. The need to compete with AI, on the one hand, puts higher and higher demands on human capital, on the other hand, requires more emphasis on the cultivation of non-cognitive abilities. Research from Harvard University’s Center for Child Development shows that in the first few years of life, the brain is able to make more than 1 million neuronal connections per second, which cannot be repeated at any later stage of life. Non-cognitive skills are best developed at the ages of three and four, and the ideal move would be to extend compulsory education to this pre-school age. Moreover, a significant increase in expected labor productivity could significantly expand public resources for education, enough to support longer school days for children.

Finally, remove institutional barriers to child development, education and training, mobility and employment, social protection, and other basic public services. In particular, the institutional reasons for the existence of migrant children and left-behind children should be eliminated. Studies have shown that the nurturing and care of parents play an irreplaceable role in the cultivation of children’s human capital, especially in the acquisition of non-cognitive abilities, as well as their lifelong social mobility opportunities. Therefore, promoting the reform of the household registration system to address such problems faced by left-behind children and migrant children should be given the highest priority.

本文由CXO UNION-CXO联盟（cxounion.cn）转载而成，来源于第一财经·莱维特；编辑/翻译：CXO UNIONCXO联盟小U。

如需加入CXO UNION（CXO联盟）高管社群，请联系社群小伙伴哦~

免责声明: 本网站(http://www.cxounion.cn/)内容主要来自原创、合作媒体供稿和第三方投稿，凡在本网站出现的信息，均仅供参考。本网站将尽力确保所提供信息的准确性及可靠性，但不保证有关资料的准确性及可靠性，读者在使用前请进一步核实，并对任何自主决定的行为负责。本网站对有关资料所引致的错误、不确或遗漏，概不负任何法律责任。

本网站刊载的所有内容(包括但不仅限文字、图片、LOGO、音频、视频、软件、程序等) 版权归原作者所有。任何单位或个人认为本网站中的内容可能涉嫌侵犯其知识产权或存在不实内容时，请及时通知本站，予以删除。