Still waters, rapid currents: Early labor market transformation under generative AI
Think of a marketing professional using a chatbot to draft ad copy, an IT support specialist troubleshooting a user’s problem with one, or a teacher trying to spot homework written by AI.
想象一下一位营销专业人士使用聊天机器人起草广告文案，一位 IT 支持专家解决用户的问题，或者一位老师试图发现人工智能写的作业。
A few years after ChatGPT’s launch, generative AI has quickly become part of everyday white-collar work.
ChatGPT 推出几年后，生成式人工智能已迅速成为白领日常工作的一部分。
Has it changed how much workers earn, or how many people firms hire?
它是否改变了工人的收入或公司雇用的人数？
Our research provides one of the first comprehensive answers, based on surveys of 25,000 Danish workers across 11 highly exposed occupations, linked to their monthly administrative records through December 2024.
我们的研究提供了第一个全面的答案之一，该研究基于对 11 个高风险职业的 25,000 名丹麦工人的调查，并与他们截至 2024 年 12 月的每月管理记录相关联。
The headline result is a precise null: no effects on earnings or hours, either for workers who use chatbots or for the firms that adopt them.
总体结果是一个精确的零：对于使用聊天机器人的工人或采用聊天机器人的公司来说，收入或工作时间没有受到影响。
Beneath this still surface, however, work is already being substantially reorganized.
然而，在这平静的表面之下，工作已经在大幅重组。
Experts disagree sharply about how generative AI will affect work, both in direction and in size.
对于生成式人工智能将如何影响工作，无论是方向还是规模，专家们意见分歧很大。
Anthropic CEO Dario Amodei (2025) warns that much white-collar work could be displaced within a few years, while Nobel laureate Daron Acemoglu (2025) expects modest productivity gains spread over decades.
Anthropic 首席执行官 Dario Amodei (2025) 警告称，许多白领工作可能会在几年内被取代，而诺贝尔奖获得者 Daron Acemoglu (2025) 预计生产率的小幅增长将持续数十年。
The stakes are especially high for young workers: is this the worst or the best time to start a career in an AI-exposed occupation?
对于年轻工人来说，风险尤其高：现在是在人工智能职业中开始职业生涯的最糟糕还是最好的时机？
In an influential paper, Brynjolfsson, Chandar, and Chen (2025) document that early-career employment has fallen sharply in AI-exposed occupations, which has raised concerns that AI may already be displacing entry-level jobs.
Brynjolfsson、Chandar 和 Chen（2025）在一篇颇具影响力的论文中指出，在与人工智能相关的职业中，早期职业就业人数急剧下降，这引发了人们对人工智能可能已经取代入门级工作的担忧。
By contrast, a more optimistic view, advanced by Autor (2024) , is that generative AI may put expertise that used to be scarce within reach of more workers, opening paths to better-paid careers.
相比之下，Autor (2024) 提出的更乐观的观点是，生成式人工智能可能会让更多工人接触到过去稀缺的专业知识，从而为更高薪的职业开辟道路。
Until now, there has been little direct evidence to weigh these views against each other.
到目前为止，几乎没有直接证据来权衡这些观点。
In this article, we describe how workplaces are absorbing AI chatbots: the employer policies, tools, and training now surrounding the technology, and the new job tasks it is creating.
在本文中，我们描述了工作场所如何吸收人工智能聊天机器人：目前围绕该技术的雇主政策、工具和培训，以及它正在创造的新工作任务。
We then ask whether these changes have surfaced in earnings, hours, and employment, with occupational mobility emerging as the notable exception.
然后我们询问这些变化是否已经体现在收入、工作时间和就业方面，其中职业流动性是一个明显的例外。
We conclude by discussing what our findings mean for policy: how statistical agencies should measure AI’s labor market effects, why policymakers should be careful attributing aggregate trends (such as the decline in entry-level hiring) to AI, and why the occupational mobility channel deserves attention.
最后，我们讨论了我们的研究结果对政策的意义：统计机构应如何衡量人工智能对劳动力市场的影响，为什么政策制定者应谨慎地将总体趋势（例如入门级招聘的下降）归因于人工智能，以及为什么职业流动渠道值得关注。
Denmark provides an unusually clean window into this process.
丹麦为了解这一过程提供了一个异常清晰的窗口。
Few workforces have taken up generative AI faster than Denmark’s, where adoption rates are on par with the United States (Humlum and Vestergaard 2025).
很少有劳动力比丹麦更快地采用生成式人工智能，丹麦的采用率与美国相当（Humlum 和 Vestergaard 2025）。
Its labor market is also flexible: hiring and firing costs are low, much as in the U.S., so firms and workers face few barriers to adjusting employment as technology changes.
中国的劳动力市场也很灵活：招聘和解雇成本都很低，与美国相似，因此随着技术的变化，企业和工人在调整就业方面几乎没有障碍。
Critically, Denmark is unusually well equipped to measure who is actually using a new technology: each resident has an official digital mailbox through which Statistics Denmark can send survey invitations.
至关重要的是，丹麦非常有能力衡量谁实际上在使用新技术：每个居民都有一个官方数字邮箱，丹麦统计局可以通过该邮箱发送调查邀请。
Working with Statistics Denmark, we fielded two rounds of representative surveys, in late 2023 and late 2024, each inviting 115,000 workers to report on their use of AI chatbots and on changes in how their work is organized.
我们与丹麦统计局合作，于 2023 年末和 2024 年末进行了两轮代表性调查，每轮邀请 115,000 名员工报告他们对人工智能聊天机器人的使用情况以及他们工作组织方式的变化。
The surveys target 11 occupations that are highly exposed to AI chatbots: software developers, IT support specialists, customer support specialists, office clerks, accountants, financial advisors, HR professionals, legal professionals, marketing professionals, journalists, and teachers.
该调查针对高度接触人工智能聊天机器人的 11 种职业：软件开发人员、IT 支持专家、客户支持专家、办公室文员、会计师、财务顾问、人力资源专业人员、法律专业人员、营销专业人员、记者和教师。
We chose these occupations because they involve tasks where AI chatbots can save workers time, based on the exposure assessments of Eloundou, Manning, Mishkin, and Rock (2024).
我们选择这些职业是因为根据 Eloundou、Manning、Mishkin 和 Rock（2024）的暴露评估，它们涉及人工智能聊天机器人可以节省工人时间的任务。
Together, they make up the parts of the labor market where any effects of AI chatbots are likely to show up first and be largest.
它们共同构成了劳动力市场的各个部分，人工智能聊天机器人的影响可能首先显现出来，并且影响最大。
The 2024 round, the primary basis for our findings, received about 25,000 valid and complete responses across 7,000 workplaces and measured workers’ use of AI chatbots, employer initiatives such as usage policies, enterprise tools, and training, and changes in the organization of work.
2024 年的调查是我们调查结果的主要基础，收到了 7,000 个工作场所的约 25,000 份有效且完整的回复，并衡量了员工对人工智能聊天机器人的使用情况、雇主举措（例如使用政策、企业工具和培训）以及工作组织的变化。
The samples are representative of the underlying workforce on characteristics such as age, gender, experience, earnings, and wealth, and survey responses check out against variables also recorded in the administrative registers.
这些样本代表了底层劳动力的年龄、性别、经验、收入和财富等特征，调查结果与行政登记册中记录的变量进行了核对。
Because respondents are drawn from the registers, every response can be linked to monthly administrative data on earnings, hours worked, and occupations—for the respondents themselves, and for the roughly 3 million other workers at Danish workplaces.
由于受访者是从登记册中抽取的，因此每个回答都可以与受访者本人以及丹麦工作场所约 300 万其他工人的收入、工作时间和职业等月度管理数据相关联。
The link gives us both sides of the picture: the outcomes that registers record, namely earnings and hours, and the adjustments they miss, from shifting job tasks and work organization to the complementary investments that firms make around new technologies (Brynjolfsson, Rock, and Syverson 2021).
这种联系为我们提供了图景的两面：记录记录的结果，即收入和工作时间，以及他们错过的调整，从改变工作任务和工作组织到公司围绕新技术进行的补充投资（Brynjolfsson、Rock 和 Syverson 2021）。
Our analysis then asks one simple question: have chatbot adopters fared differently in the labor market than comparable non-adopters?
然后，我们的分析提出了一个简单的问题：聊天机器人采用者在劳动力市场上的表现是否与类似的非采用者不同？
We compare workers who use AI chatbots to otherwise similar workers in the same occupation who do not, and workplaces with chatbot initiatives to similar workplaces without them, tracking outcomes month by month around the launch of ChatGPT in November 2022.
我们将使用人工智能聊天机器人的员工与同一职业中不使用人工智能聊天机器人的类似员工进行比较，并将采用聊天机器人举措的工作场所与不采用人工智能聊天机器人的类似工作场所进行比较，并在 2022 年 11 月推出 ChatGPT 前后逐月跟踪结果。
Because the records extend back years before ChatGPT existed, we can check that adopters and non-adopters followed similar labor market trajectories beforehand, so that differences emerging afterward can more plausibly be attributed to the technology.
由于记录可以追溯到 ChatGPT 存在之前的几年，我们可以检查采用者和未采用者事先遵循相似的劳动力市场轨迹，因此随后出现的差异可以更合理地归因于该技术。
Employers have rapidly embraced AI chatbots.
雇主已迅速接受人工智能聊天机器人。
Our 2024 survey asked workers about their employer’s policy on AI chatbot use, with five possible answers: chatbot use is encouraged, use is allowed, use is not allowed, the workplace has no policy, or the worker does not know of one.
我们 2024 年的调查询问了员工雇主对人工智能聊天机器人使用的政策，有五个可能的答案：鼓励使用聊天机器人、允许使用、不允许使用、工作场所没有政策或员工不知道有政策。
As Figure 1 shows, 43% of workers in our 11 occupations work for employers that explicitly encourage chatbot use, while only 6% are prohibited from using them.
如图 1 所示，在我们的 11 个职业中，43% 的员工为明确鼓励使用聊天机器人的雇主工作，而只有 6% 的员工被禁止使用聊天机器人。
Among workers in encouraging workplaces, 61% have access to enterprise chatbots and 39% have received training in using them.
在鼓励工作场所的员工中，61% 的人可以使用企业聊天机器人，39% 的人接受过使用培训。
The rapid adoption of AI chatbots has been bottom-up: even in workplaces that take no active steps, about 41% of workers have used chatbots at work.
人工智能聊天机器人的快速采用是自下而上的：即使在没有采取积极措施的工作场所，也有约 41% 的员工在工作中使用了聊天机器人。
Yet, despite this high worker-driven baseline, adoption more than doubles in workplaces that combine encouragement with enterprise tools and training: 93% of workers in such settings have used AI chatbots at work, 28% use them daily, and 19% report saving more than one hour per day.
然而，尽管工人驱动的基线很高，但在将鼓励与企业工具和培训相结合的工作场所，采用率增加了一倍多：在这种环境中，93% 的工人在工作中使用人工智能聊天机器人，28% 每天使用它们，19% 的人报告每天节省超过 1 小时。
This pattern, in which adoption and reported benefits are strongest where employers pair the tools with organizational support, echoes a long line of research showing that new technologies pay off most when firms invest in complementary organizational practices, including how they manage people (e.g., Brynjolfsson and Hitt 2000; Bloom, Sadun, and Van Reenen 2012).
这种模式在雇主将工具与组织支持结合起来时采用和报告的好处是最强的，这与一系列研究相呼应，这些研究表明，当公司投资于互补的组织实践（包括他们如何管理人员）时，新技术的回报最大（例如，Brynjolfsson和Hitt 2000；Bloom，Sadun和Van Reenen 2012）。
This widespread embrace of AI marks a shift from the early days of ChatGPT, when many employers restricted its use over concerns about data confidentiality and output accuracy.
人工智能的广泛接受标志着 ChatGPT 早期的转变，当时许多雇主出于对数据保密性和输出准确性的担忧而限制其使用。
Outright bans are now rare, persisting mainly in occupations that handle sensitive data, such as financial advisors, or that require high factual accuracy, such as legal professionals.
彻底的禁令现在很少见，主要持续存在于处理敏感数据的职业，例如财务顾问，或需要高度事实准确性的职业，例如法律专业人士。
AI chatbots do more than speed up existing work; they also create new tasks, shown in Figure 2.
人工智能聊天机器人的作用不仅仅是加快现有工作的速度；他们还创建新任务，如图 2 所示。
Workers who reported new tasks described them in their own words, and we classified these free-text descriptions into six categories: AI ideation, AI content drafting, AI data insights, AI quality review, AI ethics and compliance, and AI integration.
报告新任务的工作人员用自己的话进行描述，我们将这些自由文本描述分为六类：人工智能构思、人工智能内容起草、人工智能数据洞察、人工智能质量审查、人工智能道德与合规性以及人工智能集成。
Roughly four in ten new tasks involve generating content with AI: the ideation, drafting, and data insight categories.
大约十分之四的新任务涉及使用人工智能生成内容：构思、起草和数据洞察类别。
About a third involve overseeing AI, combining quality review of its outputs with ethics and compliance work, such as checking legal accuracy or detecting AI-generated homework.
大约三分之一涉及监督人工智能，将其输出的质量审查与道德和合规工作结合起来，例如检查法律准确性或检测人工智能生成的作业。
And the largest single category, about a quarter of new tasks, is integration: fine-tuning AI assistants, writing usage policies, and building chatbots into everyday workflows.
最大的单一类别（约四分之一的新任务）是集成：微调人工智能助手、编写使用策略以及将聊天机器人构建到日常工作流程中。
Even in workplaces with no AI policy, about 8% of users have taken on entirely new tasks, and this share doubles where employers actively support adoption.
即使在没有人工智能政策的工作场所，也有大约 8% 的用户承担了全新的任务，而在雇主积极支持采用的情况下，这一比例会翻倍。
Most users (85%) spend the time they save on other job tasks; far fewer do more of the same tasks or take more leisure.
大多数用户 (85%) 将节省下来的时间花在其他工作任务上；做更多相同任务或花更多闲暇时间的人要少得多。
This is consistent with work on “job transformation” arguing that technologies reshape what jobs consist of long before they eliminate them (Freund and Mann 2026; Autor and Thompson 2025).
这与“工作转型”的研究一致，认为技术在消除工作之前很久就重塑了工作的内容（Freund 和 Mann 2026；Autor 和 Thompson 2025）。
The examples, drawn from workers’ own descriptions, are concrete.
这些例子取自工人自己的描述，很具体。
Marketing professionals report “prompting and iterating with AI to produce marketing copy, social media posts, and product descriptions.” Software developers report “fine-tuning AI coding assistants by providing feedback and project-specific examples.” Legal professionals describe “developing organizational AI usage policies and guidelines.” And a large share of teachers report new work related to “detecting AI-generated homework.”
营销专业人士报告称，“利用人工智能进行提示和迭代，以制作营销文案、社交媒体帖子和产品描述。”软件开发人员报告“通过提供反馈和特定于项目的示例来微调人工智能编码助手。”法律专业人士描述了“制定组织人工智能使用政策和指南”。很大一部分教师报告了与“检测人工智能生成的作业”相关的新工作。
Notably, the new work extends even to workers who have never used AI chatbots.
值得注意的是，这项新工作甚至扩展到从未使用过人工智能聊天机器人的工人。
About 4% of non-users report new workloads resulting from the tools, and the share is highest in workplaces with chatbot initiatives.
大约 4% 的非用户表示这些工具带来了新的工作负载，并且在采用聊天机器人计划的工作场所中这一比例最高。
Among teachers who have never used the tools, 10% report new AI-related work, mostly keeping tabs on, and responding to, students’ AI use.
在从未使用过这些工具的教师中，10% 的教师报告了与人工智能相关的新工作，主要是密切关注并回应学生对人工智能的使用。
These spillovers to non-users signal a broader workplace transformation, consistent with task-based theories in which new technologies reinstate labor demand by creating new work (Acemoglu and Restrepo 2018).
这些对非用户的溢出效应标志着更广泛的工作场所转型，这与基于任务的理论相一致，在该理论中，新技术通过创造新的工作来恢复劳动力需求（Acemoglu 和 Restrepo 2018）。
The emergence of new work is also consistent with the historical record.
新作品的出现也与历史记录相符。
Research by David Autor and coauthors shows that 60% of U.S. employment today is in job titles that did not exist in 1940 (Autor, Chin, Salomons, and Seegmiller 2024)—the creation of new work is how labor markets have absorbed new technologies over time.
David Autor 和合著者的研究表明，如今美国 60% 的就业岗位在 1940 年并不存在（Autor, Chin, Salomons, and Seegmiller 2024）——随着时间的推移，新工作的创造是劳动力市场吸收新技术的方式。
What is unusual is catching the process in the act: our data capture the emergence of AI oversight and integration tasks within two years of the technology’s arrival, before any of them appear in official occupational classifications.
不同寻常的是捕捉实际过程：我们的数据捕获了人工智能监督和集成任务在该技术到来后两年内出现的情况，然后它们出现在官方职业分类中。
Have the rapid currents of adoption and reorganization surfaced in what workers are paid?
采用和重组的快速潮流是否已经体现在工人的工资上？
Figure 3 says no. In levels, Panel (a) shows that adopters do earn more than comparable non-adopters (about 9% more where employers encourage use and 4% where they do not) and follow rising earnings trends.
图 3 说不。从水平上看，面板 (a) 表明，采用者的收入确实比非采用者高（雇主鼓励使用的情况下大约高出 9%，不鼓励使用的情况下则高出 4%），并且收入趋势呈上升趋势。
But these gaps were already there before ChatGPT.
但这些差距在 ChatGPT 之前就已经存在了。
Once we index to ChatGPT’s launch, the difference-in-differences in Panel (b) is flat, and its confidence intervals rule out any differential earnings change larger than 2%, for both workers and workplaces.
一旦我们对 ChatGPT 的推出进行索引，面板 (b) 中的差异就持平，并且其置信区间排除了工人和工作场所任何大于 2% 的收入差异变化。
The contrast between the two panels echoes the debate over computers and wages in the 1990s (Krueger 1993; DiNardo and Pischke 1997): users of a new technology often do earn a sizable premium, yet such premia can say more about who picks up the tools than what the tools do for them.
两个小组之间的对比呼应了 20 世纪 90 年代关于计算机和工资的争论（Krueger 1993；DiNardo 和 Pischke 1997）：新技术的用户通常确实赚取了相当大的溢价，但这种溢价更多地说明了谁使用了这些工具，而不是这些工具为他们做了什么。
These null results hold even for daily chatbot users, for workers who save more than an hour a day, for those who took on new AI tasks, and for those whose employers actively support adoption.
即使对于日常聊天机器人用户、每天节省超过一个小时的工人、承担新人工智能任务的人以及雇主积极支持采用人工智能的人来说，这些无效结果也成立。
They also hold for early adopters, already using chatbots when we surveyed in 2023, who have had longer for any effect to appear.
它们也适用于早期采用者，当我们在 2023 年进行调查时，他们已经在使用聊天机器人，但他们需要更长的时间才能看到效果。
And they hold in each of the 11 occupations, software development and marketing included, where decentralized wage setting should let pay track individual productivity most readily.
他们在包括软件开发和营销在内的 11 个职业中都占有一席之地，在这些职业中，分散的工资设定应该让工资最容易跟踪个人生产力。
Firm-level outcomes are just as still.
公司层面的结果也同样如此。
Workplaces that encourage chatbot use, including those that provide enterprise chatbots and training, show no differential changes in headcounts or wage bills, job creation or destruction, or the composition of hires or separations.
鼓励使用聊天机器人的工作场所，包括那些提供企业聊天机器人和培训的工作场所，在员工人数或工资账单、就业机会的创造或破坏、雇用或离职的构成方面没有显示出差异变化。
Incumbent workers at adopting workplaces have not been pushed out of their jobs or their occupations.
收养工作场所的现任工人并未被赶出他们的工作或职业。
The workplace estimates are as precise as the worker-level ones, ruling out effects larger than 2%.
工作场所的估计与工人层面的估计一样精确，排除了大于 2% 的影响。
Workers themselves agree on this still surface: asked directly, 98% of adopters, and virtually all non-adopters, report that AI chatbots have not affected their earnings, suggesting that spillovers to workers who do not use the tools have also been limited.
工人们自己也同意这一表面现象：直接询问时，98% 的采用者和几乎所有非采用者都表示，人工智能聊天机器人并未影响他们的收入，这表明对不使用这些工具的工人的溢出效应也很有限。