文 | 登顶可选
10月23日，国务院办公厅公开发布《关于适应人工智能技术发展助力构建就业友好型发展方式的意见》。
文件提出18项任务，其中包括减缓AI应用的就业替代影响、建立就业风险评估机制，以及调整收入分配和社会保障制度。
一份关于人工智能的政策文件，开始具体讨论如何减少岗位替代、帮助劳动者转岗，以及应对规模性失业风险。
这让一个长期争论的问题变得更加现实——官方出手，那么AI对就业究竟影响有多大？
2025年9月，Salesforce首席执行官马克·贝尼奥夫在播客中透露，公司客服团队已经从9000人缩减至约5000人。
AI承担了越来越多的客户咨询，还帮助公司重新联系过去积压的大量销售线索。贝尼奥夫将这段经历称为自己职业生涯中最令人兴奋的时期之一。
对企业来说，这是一笔漂亮的生意。客服团队缩小了，销售机会却增加了。
当然，4000人的差额并不等于4000人全部遭到裁员。部分员工被重新安排到其他岗位。但有一点已经确定：Salesforce不再需要原来规模的客服团队。
过去两年，AI几乎成了科技公司裁员新闻里的高频词。有人认为机器正在大规模抢走工作，也有人认为，企业只是借AI之名掩盖经营压力。
两种说法都有现实依据，却都无法完整解释这场变化。
因为企业减少人工需求，并不一定要先发出裁员通知。
01｜裁员公告之外，还有一本账
2024年，瑞典金融科技公司Klarna公布了一组数字。
AI客服助手上线一个月，处理了230万次客户对话，相当于700名全职客服人员的工作量。平均问题处理时间从11分钟缩短至不到2分钟。
到了2025年，Klarna向美国证券交易委员会提交的上市文件披露，在截至当年6月的过去12个月里，AI处理了69%的客服聊天。公司估算，仅2024年就节省了约3900万美元成本。
当然，700人不是裁员人数，而是AI承担的等效工作量。3900万美元则说明这项技术已经产生了可以计入经营账本的成本收益。
Klarna后来重新强调人工客服的重要性，调整服务人员的配置。AI能处理大量标准化问题，但复杂投诉、特殊需求和客户体验仍然需要人工参与。
中国企业也在算同一笔账。
国际劳工组织2026年9月发布的一项中国企业研究，访谈了21家企业，调查了1591名从业者。
其中一家保险公司，在300名客服员工规模不变的情况下，借助AI将每日问题处理量从6000个提高至15000个。
一家旅游服务企业估算，AI承担的数据处理工作量相当于5至6名全职员工。
这些数字来自企业访谈，尚不能作为整个行业的平均水平。但它们揭示了一个重要变化：工作量和员工数量，正在逐渐脱钩。
保险公司没有减少300人的客服团队，却能处理过去2.5倍的问题。如果业务继续增长，企业可以用现有团队服务更多客户；如果业务规模保持稳定，新增招聘的必要性就会下降，这就意味着新增就业机会逐渐减少。
这类隐形变化很少被称为AI裁员，公开裁员数据则是另一种口径。
美国裁员咨询机构Challenger, Gray & Christmas的2026年前九个月统计显示，企业宣布的裁员计划中，约12万人被归类为AI相关，占同期全部裁员公告的约21%。
这个数字来自企业公布的裁员理由，并非独立审计确认的AI替代人数。经营不善、组织重组和技术升级往往同时发生，企业给出的解释也有自身的传播动机。
但即使剔除那些无法确定归因的裁员，Salesforce和Klarna的经营数据仍然说明，AI已经改变了部分企业的人工需求。
因此，衡量AI的就业影响，至少需要看三项变化：已经减少的岗位、被AI承担的工作量，以及因此减少的新增招聘。
前两项已经有越来越多的企业案例。第三项更难统计，却直接关系到年轻人的职业前景。
02｜AI先改变的，是年轻人的职业入口
2026年8月，斯坦福大学数字经济实验室更新了一项就业研究。
研究人员利用美国薪资服务商ADP覆盖数百万劳动者的数据发现，在AI暴露程度较高的职业中，22至25岁年轻劳动者的就业表现持续弱于对照群体。
截至2026年6月，两类群体的相对就业差距达到19%。
这不是美国年轻人整体就业下降19%，也不能直接理解为AI消灭了19%的岗位。
研究更重要的发现是：差距主要体现在招聘减少，而非年轻员工被解雇的比例明显增加。
一些年轻人没有被AI裁掉，因为他们根本没能进入那个岗位。
过去，初级岗位不仅承担基础工作，也承担人才培养功能。
程序员从修改简单代码、编写测试开始；市场人员通过搜集资料、制作初稿熟悉业务；咨询和研究岗位的新人，则在整理数据、处理文档的过程中积累判断能力。
如今，这些基础任务恰好是生成式AI能够大量介入的工作。
经验丰富的员工可以让AI生成代码初稿、整理资料，再自行检查、修改和决策。企业减少了重复劳动，也开始重新评估初级岗位的招聘数量。
短期看，这是效率提升；长期看，行业面临一个矛盾：
企业仍然需要经验丰富的人，却在减少让新人获得经验的工作。
如果一家企业减少初级招聘，可以通过更高的效率获得竞争优势。但当越来越多企业采取同样的做法，行业就必须回答下一代资深员工从哪里来？
中国招聘市场也出现了相关信号。
一项分析2019年至2024年约6700万条中国招聘广告的研究发现，AI替代型暴露较高的岗位，常规认知技能需求占比下降；AI增强型暴露较高的岗位，非重复性分析技能需求则有所增加。
另一项覆盖超过1300万条招聘广告的研究发现，生成式AI普及后，AI相关招聘参与度较高的上市企业，中等工资和教育要求岗位的招聘占比相对下降。
这些研究观察的是招聘结构，不能直接等同于就业人数变化，但它们呈现了企业技能需求的调整方向。
企业越来越重视独立判断、复杂问题解决和AI应用能力，而传统初级岗位承担的基础任务正在被重新分配。
新的AI职业当然也在出现。模型评估、AI产品运营、自动化流程设计等工作，正在创造新的就业需求。
但新旧岗位的技能要求以及数量并不相同，岗位消失和创造也不会同步发生。
对一个刚毕业的年轻人来说，最困难的未必是学会使用AI，而是如何在缺少实际业务经验的情况下，获得企业越来越看重的判断能力。
这已经超出了单纯的裁员问题。
AI正在改变的不只是岗位数量，还有职业成长的路径。
03｜企业赚到了效率，劳动者分到了多少？
Klarna节省了约3900万美元，Salesforce缩小了客服团队，中国的保险公司用同样的人力处理更多业务。
企业获得了可以衡量的生产率收益。但节省下来的成本，并不会自动转化为员工收入。
它可以通过降价让消费者受益，可以用于研发和业务扩张，也可以增加员工报酬，或者成为企业利润。
不同的分配方式，会产生完全不同的就业结果。
一家利用AI扩大市场的企业，可以同时增加收入、利润和就业；另一家主要利用AI压缩人工成本的企业，则可能在员工减少的情况下提高产出。
两家公司都实现了降本增效，劳动者的处境却不一样。
经济学中有一个指标叫劳动收入份额，衡量劳动报酬占创造的增加值的比例。
假设一家企业原本创造1亿元增加值，支付5000万元劳动报酬，劳动收入份额为50%。引入AI后，企业创造的增加值增长至1.5亿元，劳动报酬总额增加到6000万元。劳动者获得的报酬总额上涨20%，收入份额却下降至40%。
这个假设说明：劳动者拿到的钱可以增加，但他们在经济增长中获得的比例仍然下降。
如果企业同时减少基础岗位、提高高技能员工薪酬，平均工资甚至可能上涨，而失去岗位的人已经不在平均工资统计之中。
关于AI与劳动收入份额，中国上市公司的实证研究目前存在不同结果。有的发现AI应用促进了劳动收入份额提高，有的观察到份额下降及收益向高技能员工集中的现象。
研究样本、AI采用程度的衡量方法和统计时期不同，尚不足以得出AI必然降低劳动收入份额的结论。但这恰恰说明，技术进步与劳动者收入增长之间，并不存在自动生效的分配规则。
AI提高生产率之后，谁获得新增价值，取决于市场竞争、企业经营方式、员工技能结构和收入分配制度。
但是，当更多企业能够用更少的人完成更多工作，这个问题就不再局限于企业内部。因为现代社会的工资收入，还连接着税收、社保和公共服务。
04｜当机器接手工作，谁来承担社会成本？
2017年，比尔·盖茨接受Quartz采访时，提出了一个颇有争议的建议：考虑向机器人征税。
2026年8月，他在个人网站再次讨论这一问题，甚至提出研究对AI tokens征税。
盖茨担心的是，当机器替代部分劳动者，原本依靠工资形成的税收和社会保障缴费可能受到影响，而失业保障、职业培训和其他公共服务的需求反而增加。
这不是AI完全不交税的问题。
开发AI的企业、销售AI服务的公司，仍然需要按照适用法律缴纳相应税款。企业支付的工资和社保费用通常也可以按规定税前扣除，不能简单地说雇人是税后成本、买设备才是税前成本。
真正的差异在于，劳动和资本适用不同的税收与缴费规则。
企业雇用员工，需要承担工资及相关社会缴费；投资设备、软件和自动化系统，则按照相应的折旧、摊销及税收政策处理。
例如，中国财政部、税务总局2023年第37号公告规定，2024年至2027年，企业新购进、单位价值不超过500万元的设备和器具，可以按规定一次性计入当期成本费用，在计算企业所得税时扣除。
这是一项普遍的设备投资政策，并非专门鼓励AI替代员工。但它说明，税制会参与企业对人工和资本投资的成本计算。
盖茨提出的机器人税，试图让自动化收益承担更多公共责任。国际货币基金组织却不赞成简单征收专门的AI税或机器人税。
IMF在2024年的财政政策研究中指出，直接对AI征税不仅难以界定征税对象，还可能抑制有益的技术投资。
工厂里的机械臂容易识别，企业用AI审核合同、生成代码、处理财务报表，又该如何计税？如果AI帮助员工提高效率，没有减少就业，是否也要征税？
IMF更倾向于调整现有税制中可能鼓励劳动替代的政策，完善资本收入和企业利润税收，并加强职业培训与社会保障。
盖茨与IMF的路径不同，却共同面对一个问题：
当企业越来越容易在不增加员工的情况下扩大产出，公共财政如何分享生产率收益，并为受到冲击的劳动者提供保障？
税收制度不会因为一家企业部署AI就立即失效。生产率提高也可能带来更高利润、更多消费和新的税收来源。
但如果劳动收入增长长期落后于产出增长，依赖劳动收入的税收与保障安排，就需要作出调整。
这也是为什么，AI就业问题最终会延伸到社会制度。
05｜谁来接手“雇佣”原本承担的功能？
工业革命时期，机器大幅提高了生产效率，劳动者却没有立即分享全部增长成果。
经济史研究中有一个概念叫“恩格斯停顿”，用来描述英国工业革命早期，生产率增长与劳动者实际工资增长长期脱节的现象。
不同研究对其持续时间和工资变化幅度存在争议，但它提出的问题至今仍然成立——生产力进步，并不保证普通劳动者同步受益。
2026年10月27日，国家发展改革委网站刊发了国家信息中心研究人员关于防范“恩格斯停顿”的专家文章，讨论如何避免技术进步与民生改善脱节。
历史不会简单重演，今天的劳动市场、社会保障和技术扩散速度，也与十九世纪完全不同。
但AI正在让一个长期被视为理所当然的关系受到挑战。
过去，企业扩大业务通常需要增加相应的人力投入。雇佣关系不仅提供工资，也让新人积累经验，并将劳动者连接到社保和公共财政体系。
今天，企业已经能够在部分业务中，通过AI提高产出而不同比例增加员工。
这意味着，雇佣关系原本承担的几项社会功能，开始面临不同程度的压力。
基础工作被AI接手，职业培养需要新的路径；企业人工需求减少，劳动收入的分配方式需要重新审视；工资增长与生产率增长脱钩，公共保障也需要寻找更稳定的支持机制。
这些变化不会在同一天发生，也不会同时出现在所有行业。但它们已经超出了单个企业如何裁员的范围。
10月23日公布的国务院办公厅《意见》，恰好回应了这些问题。
文件明确提出减缓AI应用的就业替代影响，鼓励企业探索失业缓冲期制度，并要求加强转岗培训、就业风险评估、收入分配和社会保障制度建设。
尤其值得注意的是，文件提出开展AI就业影响专项调查，持续跟踪岗位创造与替代、技能需求变化，并将AI应用就业风险纳入规模性失业风险防范机制。
这意味着，AI对就业的影响不仅需要被讨论，还需要被持续测量和管理。政策已经开始回应，接下来考验的是落实速度。
企业可以在一个季度里完成AI系统部署，在下一轮预算中减少招聘计划；但一个劳动者从失去工作到掌握新技能、重新获得稳定收入，往往需要更长时间。
这才是AI就业转型中最现实的时间差。
尾声
回到最初的问题：AI裁员潮，到底有多少是真的？
目前没有一份统计能够完整回答。
裁员公告记录了部分岗位消失，企业运营数据记录了AI接手多少工作，招聘研究则开始捕捉那些从未出现的就业机会。
三者共同说明，AI对劳动需求的改变已经发生，但影响远不止公开裁员人数。
我们也不必因此认定，人类工作终将被机器全面取代。技术可以创造新需求、新职业和更高收入，历史上的生产率进步也确实扩大过社会财富。
真正悬而未决的，是收益和成本如何分配，以及制度能否跟上企业改变用工方式的速度。
企业已经能够按季度计算AI节省的人工成本，劳动者却必须按月支付生活账单。
如果企业能够用更少的人创造更多价值，社会就需要回答：那些不再被企业雇用的人，如何继续获得收入、积累能力，并拥有基本保障？
AI能省下多少人工，企业已经开始算清。省下人工之后，人该怎么办，这笔账才刚刚开始。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Man, you got a choice.
On 23 October, the Office of the State Council made public publish ' s Opinions on Adapting to Artificial and Intelligent Technology Development to Build Employment Friendly Development.
The paper identified 18 tasks, including mitigating the effects of employment substitution in AI applications, establishing an employment risk assessment mechanism and adjusting income distribution and social security systems.
A policy paper on artificial intelligence began to discuss, in particular, ways of reducing job substitution, helping workers to move to work, and addressing the risk of unemployment on a scale.
This makes a long-debated issue even more realistic — the official hand-in-hand, so what is the impact of AI on employment?
In September 2025, Chief Executive Officer of Salesforce, Mark Beniov, revealed in a podcast that the company ' s customer service team had been reduced from 9,000 to about 5,000.
AI has undertaken an increasing number of client consultations and has also helped companies reconnect with the backlog of past sales leads. Beniov described this experience as one of the most exciting periods of his career.
It's a beautiful business for a business. The customer service team has been reduced and sales opportunities have increased.
Of course, the difference of 4,000 is not equal to that of all 4,000 people being laid off. Some employees were reassigned to other positions. But one thing is certain: Salesforce no longer needs a customer service team of its original size.
In the past two years, AI has become almost a high-frequency word in the news of tech-company layoffs. There are those who believe that machines are taking jobs on a large scale, as well as those who believe that businesses are simply using the name of AI to hide their operating pressures.
There is a realistic basis for both statements, but neither can fully explain the change.
Due to the reduced labour demand of enterprises, layoffs do not have to be notified first.
There's a book in addition to the staff announcement.
In 2024, Klarna, a Swedish financial technology company, published a set of figures.
AI's passenger service assistant was on line for one month and handled 2.3 million client dialogues, corresponding to the workload of 700 full-time visitors. The average problem time has been reduced from 11 minutes to less than 2 minutes.
By 2025, Klarna had disclosed to the United States Securities and Exchange Commission that, over the past 12 months up to June of that year, AI had processed 69 per cent of the passenger service chats. The company estimates that approximately $39 million was saved in 2024 alone.
Of course, 700 are not staff reductions, but equivalent AI workloads. The amount of $390 million indicates that the technology has generated cost benefits that can be recorded in the operating books.
Klarna subsequently re-emphasized the importance of manual passenger service and adjusted the staffing of the service. AI can handle a large number of standardization issues, but complex complaints, special needs and client experience still require manual involvement.
Chinese enterprises are also accounting for the same amount.
In September 2026, the International Labour Organization conducted a study of Chinese enterprises in publish, which interviewed 21 enterprises and surveyed 1591 practitioners.
One of the insurance companies, with a constant size of 300 passenger staff, increased the number of problems handled daily from 6,000 to 15,000 through AI.
A tourism service enterprise estimates that AI has a data-processing workload equivalent to five to six full-time employees.
These figures are from business interviews and do not yet constitute an industry-wide average. However, they reveal an important change: the workload and the number of staff are gradually decoupling.
The insurance company did not reduce the customer service team by 300, but was able to deal with the past 2.5 times. If business continues to grow, businesses can use existing teams to serve more clients; if business size is stabilized, the need for additional recruitment will decrease, which means that new jobs will gradually decrease.
Such invisible changes are rarely referred to as AI layoffs, and open cuts are another calibre.
Statistics from the first nine months of 2026 of the United States Advisory Body on Retrenchment, Challenger, Gray & Christmas, show that about 120,000 people are classified as AI-related in the company ' s announced retrenchment scheme, accounting for about 21 per cent of all redundancies over the same period.
This figure is derived from the reason for the reduction of staff published by the enterprise and is not the number of AI replacements confirmed by the independent audit. Poor operations, organizational reorganization and technological upgrading often occur simultaneously, and enterprises give their own motivation for dissemination.
However, even with the removal of staff reductions for which attribution cannot be established, the operating data of Salesforce and Klarna indicate that AI has changed the labour requirements of some businesses.
Thus, the employment impact of AI needs to be measured by at least three changes: the reduced jobs, the workload of AI, and the resulting reduction in new recruitment.
The first two already have a growing number of business cases. The third, more difficult to measure, is directly related to the career prospects of young people.
It's a young man's career entrance.
In August 2026, the Digital Economy Laboratory of Stanford University updated an employment study.
Using data from the United States pay service provider ADP covering millions of workers, researchers found that young workers between 22 and 25 years of age continued to have weaker employment performance than the comparator group in occupations with higher AI exposure.
By June 2026, the relative employment gap between the two groups had reached 19 per cent.
This is not a 19 per cent decline in the overall employment of young people in the United States, nor can it be directly understood that AI has eliminated 19 per cent of jobs.
More importantly, the study found that the gap was mainly reflected in the decrease in recruitment and a marked increase in the percentage of dismissals of non-young employees.
Some young people were not cut off by AI because they had no access to that position.
In the past, junior positions were not only basic jobs but also human development functions.
Programmers begin by changing simple codes and preparing tests; marketers become familiar with business by collecting information and producing first drafts; and new persons in advisory and research positions accumulate judgement in the process of collating data and processing files.
Today, these basic tasks happen to be those in which generating AI can be heavily involved.
Experienced staff can allow AI to generate first draft code, collate information, and then examine, modify and make decisions. Enterprises have reduced duplication of work and have begun to reassess the number of recruitments for junior positions.
In the short term, this is an increase in efficiency; in the long run, the industry faces a contradiction:
Businesses still need experienced people but are reducing the number of jobs in which new people can gain experience.
If an enterprise reduces primary recruitment, it can obtain a competitive advantage through greater efficiency. But when more and more enterprises do the same, the industry has to answer where the next generation of senior employees comes from.
The Chinese recruitment market also showed signs.
A study analysing about 67 million Chinese job advertisements between 2019 and 2024 found that substitution-type AI exposed high-level jobs with declining demand for conventional cognitive skills; and enhanced AI exposed high-level jobs with increased demand for non-repeated analytical skills.
Another study, covering over 13 million job advertisements, found a relative decline in the percentage of recruitment to middle-paying and education-required jobs in publicly listed enterprises with a high level of participation in AI-based AI.
These studies look at recruitment structures, which cannot be directly equated with changes in employment numbers, but they provide a reorientation of the skill needs of enterprises.
Enterprises are increasingly focusing on independent judgement, complex problem resolution and AI application capabilities, while the basic tasks of traditional junior positions are being redistributed.
Of course, a new AI career also emerged. Model assessments, AI product operations, and automated process design are creating new employment needs.
However, the skill requirements and number of new and old jobs vary, and the loss and creation of jobs do not occur simultaneously.
The most difficult for a young recently graduated person is not necessarily to learn to use AI, but rather to gain the ability of the firm to judge more and more in the absence of practical business experience.
This goes beyond the mere issue of redundancy.
AI is changing not only the number of jobs but also the path to professional growth.
How much did the workers get?
Klarna saved about $39 million, Salesforce reduced the customer service team, and Chinese insurance companies used the same manpower to handle more business.
Enterprises receive measurable productivity gains. However, cost savings do not automatically translate into staff income.
It can benefit consumers through price reductions, can be used for R & D and business expansion, can increase employee remuneration, or can become business profits.
Different types of distribution produce completely different employment outcomes.
One enterprise using AI to expand the market can increase income, profits and employment at the same time, while another, which mainly uses AI to compress labour costs, may increase output with fewer employees.
Both companies achieved efficiency gains, while workers were in a different situation.
One indicator in economics is the share of labour income, which measures the share of labour remuneration in the added value created.
A business would have created a value added of 100 million yuan and paid 50 million yuan for work, with a share of 50 per cent. With the introduction of AI, the added value of enterprise creation increased to 150 million yuan and the total amount of labour remuneration increased to 60 million yuan. The total remuneration received by workers increased by 20 per cent, while the share of income fell to 40 per cent.
This assumption suggests that workers can get more money, but their share of economic growth continues to decline.
If an enterprise simultaneously reduces basic jobs and increases the remuneration of highly skilled workers, the average wage may even rise, and those who lose their jobs are no longer included in the average wage statistics.
With regard to AI and the share of labour income, empirical studies of Chinese listed companies now have different results. Some found that AI applications have contributed to an increase in the share of labour income, while others have observed a decrease in the share and a concentration of income towards highly skilled workers.
The study of sample, the different measurement methods and statistical periods used by AI, is not sufficient to conclude that AI necessarily reduces the share of labour income. But this is precisely an indication that there is no automatic distribution rule between technological advances and the growth of workers ' income.
After AI increases productivity, who gets added value depends on market competition, business practices, employee skill structures and income distribution systems.
However, when more enterprises are able to do more with fewer people, the problem is no longer confined to the enterprise itself. Because of the wage income of modern society, it also links taxes, social security and public services.
Who will bear the social costs when machines take over?
In 2017, in an interview with Quartz, Bill Gates made a controversial proposal to consider taxing robots.
In August 2026, he revisited the issue on his personal website and even proposed a study to tax AI tokens.
Gates is concerned that when machines replace some workers, taxes and social security contributions that would otherwise be based on wages may be affected, while the need for unemployment security, vocational training and other public services increases.
It's not about AI not paying taxes at all.
Companies developing AI, companies selling AI services, are still required to pay corresponding taxes in accordance with applicable law. Wages and social security costs paid by enterprises can normally also be deducted prior to the imposition of taxes, and it cannot simply be said that the employee is the post-tax cost and that the purchase of equipment is the pre-tax cost.
The real difference is that labour and capital are subject to different tax and contribution rules.
Enterprises employ employees who bear wages and related social contributions; investments in equipment, software and automated systems are treated in accordance with the corresponding depreciation, amortization and tax policies.
For example, Circular No. 37 of 2023 of the Ministry of Finance and the General Tax Administration of China provides that equipment and appliances purchased by enterprises with a unit value of up to $5 million may be credited to the current cost in a one-time manner between 2024 and 2027 and deducted in the calculation of the income tax of enterprises.
This is a general equipment investment policy and is not specifically designed to encourage AI replacement staff. It indicates, however, that the tax system is involved in the costing of enterprise investment in labour and capital.
The robotic tax put forward by Gates is an attempt to increase public responsibility for automation gains. The International Monetary Fund is not in favour of a simple AI or robotic tax.
In its 2024 fiscal policy study, the IMF noted that direct taxation of AI not only makes it difficult to define what is taxed, but may also discourage beneficial investment in technology.
The mechanical arm of the factory is easily identifiable, and the company uses AI to audit contracts, generate codes, process financial statements, and how do it be taxed? Should AI help staff to become more efficient and not reduce employment, would taxes also be levied?
The IMF prefers to adjust policies in the existing tax system that may encourage labour substitution, improve capital income and tax on enterprise profits, and strengthen vocational training and social security.
Gates has a different path than the IMF, but is facing a problem together:
How can public finances share productivity gains and provide security for workers affected by shocks when enterprises are becoming increasingly easy to expand without increasing their workforce?
The tax system does not immediately lapse because an enterprise deploys AI. Increased productivity may also lead to higher profits, more consumption and new tax sources.
However, if growth in labour income lags behind growth in output in the long run and tax and security arrangements relying on labour income require adjustment.
This is why the issue of AI employment ultimately extends to the social system.
Who will take over the functions originally performed by “employment”?
During the period of the industrial revolution, the machine had significantly improved its productivity, and workers had not immediately shared the full fruits of growth.
One of the concepts in the economic history study, called the “Ngos pause”, was used to describe the long-term disconnect between productivity growth and real wage growth of workers in the early years of the British industrial revolution.
Different studies dispute their duration and the extent of wage changes, but the issues it raises are still valid to date - productivity progress does not guarantee that ordinary workers benefit simultaneously.
On 27 October 2026, the website of the National Commission for Development and Reform published an expert article by researchers of the National Information Centre on prevention of the “Ngos pause” to discuss ways to avoid a disconnect between technological advances and improvements in human life.
History will not simply repeat itself, and today ' s labour market, social security and technology are spreading at a different pace than in the nineteenth century.
But AI is challenging a relationship that has long been taken for granted.
In the past, the expansion of business typically required a corresponding increase in human investment. The employment relationship not only provides wages but also allows new entrants to accumulate experience and connect workers to the social security and public finance systems.
Today, businesses have been able to increase their workforce in different proportions in part of their operations by increasing output through AI.
This means that the employment relationship is beginning to face varying degrees of pressure for several social functions that it was supposed to assume.
Basic work is taken over by AI and new paths are needed for vocational development; labour demand in enterprises is reduced and the distribution of labour income needs to be revisited; wage growth is delinked from productivity growth and public security needs to find more stable support mechanisms.
These changes will not occur on the same day or at the same time in all industries. However, they have gone beyond the scope of how individual enterprises can be laid off.
The Views of the Office of the State Council, published on 23 October, respond to these concerns.
The paper clearly proposes to mitigate the employment substitution effects of AI applications, encourages enterprises to explore the unemployment buffer period system, and calls for enhanced training on transfer, job risk assessment, income distribution and social security systems.
Of particular note is the proposal to conduct a special survey on the employment impact of AI, to keep track of job creation and substitution, changes in skills needs, and to incorporate AI into the scale unemployment risk prevention mechanism.
This means that the employment impact of AI needs not only to be discussed but also continuously measured and managed. Policy has begun to respond, and the next test is the pace of implementation.
An enterprise can complete the AI system deployment in one quarter and reduce its recruitment plan in the next budget cycle; however, it often takes longer for a worker to lose his or her job to acquire new skills and regain stable income.
This is the most realistic time lag in the AI employment transition.
End
Back to the original question: how many of the AI layoffs are real?
At present, there is no complete statistical response.
Retrenchment announcements recorded the disappearance of some jobs, business operations data recorded the number of jobs taken over by AI and recruitment studies began to capture jobs that never occurred.
The three shared the view that changes in AI ' s demand for labour had taken place, but that the impact was far more than on the number of public layoffs.
Nor do we have to assume that human work will eventually be completely replaced by machinery. Technology can create new demand, new jobs and higher incomes, and historical productivity gains have indeed increased social wealth.
What remains to be done is how benefits and costs are distributed and whether the system can keep pace with changes in the way enterprises work.
Enterprises have been able to calculate AI ' s labour savings on a quarterly basis, while workers have to pay monthly subsistence bills.
If an enterprise can create more value with fewer people, society needs to answer: How can those who are no longer employed by enterprises continue to earn income, build capacity and have basic security?
How much labour can AI save and the enterprise has started to clear up. This account is only just beginning, after saving labour and what people do.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.