
德意志银行(Deutsche Bank)称之为“让AI变得糟糕的夏天”。事实上,今年夏天英伟达(Nvidia)首席执行官黄仁勋与Anthropic首席执行官达里奥·阿莫迪就后者的预测展开激烈争论。阿莫迪曾预测,人工智能将取代一半的白领岗位。整个夏天,两位高管都在频频交锋,互相抛出关于AI在“第四次工业革命”中会让多少岗位消失的预言。
但从劳动节的视角来看,情况已大不相同。8月下旬,市场意外受到冲击,原因竟是一份来自麻省理工学院(MIT)的调查:95%的大型企业生成式AI试点项目都以失败告终。这一结果引发了科技股抛售,并引起市场关于AI是否正在形成股市泡沫的讨论。而另一项最新数据也浮出水面:美国人口普查局(Census Bureau)发现,大型企业的AI采用率正在开始下降。经历两年快速试验和备受瞩目的试点项目后,许多公司似乎正在重新评估将AI纳入长期运营的实际价值。
凯莉·莫纳汉是Upwork研究院的常务董事,她掌握着来自自由职业市场的大量数据。今年9月,Upwork发布了一份关于招聘趋势和最紧缺技能的新报告。莫纳汉对《财富》杂志表示:“我看到的情况是,人类正重新回到关键环节中。我们确实看到,人类技能正变得格外稀缺和重要。人们正在意识到,即便是最先进的AI模型,依旧有10%到12%的误差。我们还无法彻底克服这一统计学问题……如今人们使用AI生成内容时发现,他们需要进行事实核查。而唯有人类才能提供事实核实。”

这一下降趋势的数据来自美国人口普查局每两周开展一次的《商业趋势与前景调查》(BTOS)。该调查覆盖超过120万家企业,能够提供不同规模企业的最新技术采用情况。最新数据显示,以六期调查的移动平均值计算,大型企业(即员工人数超过250人的公司)的AI采用率已从今年年初的峰值14%下降至2025年夏末的12%。
这一逆转是在此前几个季度的快速攀升之后出现的。大型企业的AI采用率从2023年9月的3.7%升至2024年12月的5.7%,并在2025年第二季度达到9.2%。中型企业的采用率依然较低,最高约为4.8%;而最小型企业,尤其是员工在1-4人的企业,则保持在5.5%左右,呈现温和而稳定的增长。
阿波罗全球管理公司(Apollo Global Management)首席经济学家托尔斯滕·斯洛克指出,调查中有一项问题是:企业在过去两周内是否使用了机器学习、自然语言处理、虚拟代理或语音识别等AI工具来协助生产商品或服务。斯洛克在9月7日的《Daily Spark》简报中写道:“核心结论是,美国人口普查局的双周数据已经开始显示,大型企业的AI采用进程正在放缓。”
对整体经济的影响
这一转变具有特别重要的意义,因为大型企业在塑造供应链与劳动力市场的技术趋势中发挥着巨大作用。就在一年前,“AI淘金热”还推动企业争相引入生成式AI和自动化解决方案,因为它们承诺能够带来显著的成本节约和生产率飞跃。而如今,相关表述变得更为谨慎,甚至带有怀疑的色彩。
人们同样对AI对就业市场的影响充满担忧,尤其是在入门级岗位方面。斯坦福大学经济学家进行的一项开创性研究显示,AI的显著影响才刚刚开始显现。由AI领域思想领袖埃里克·布林约尔弗森主导的研究发现,自2022年以来,22至25岁、处于职业早期的劳动者,在受AI冲击最大的岗位上就业人数相对下降了13%。美国银行研究院(Bank of America Institute)指出,从同一时间点开始,应届毕业生的失业率就已高于整体失业率。
不过,莫纳汉在随后向《财富》提供的另一份澄清声明中强调,人类技能的崛起并不仅仅是对AI采用率下滑的反应。她表示:“每一次重大技术转型都会重塑技能结构,把人们推向价值链更高端的位置。”她举例说,自动取款机并没有取代银行柜员,而是让他们转向更高价值的服务,这些服务更需要人际交往和解决问题的能力。她指出,AI也遵循着类似路径:“随着自动化承担部分任务,市场对特定人类技能的需求正在上升,包括判断力、沟通能力和情境理解能力。”
教育领域正在应对AI与人类技能的冲击
斯坦福大学(Stanford)计算机科学教授、AI初创公司联合创始人尤尔·莱斯科维克对《财富》杂志表示,几年前GPT-3的出现曾在校园里引发一场“存在危机”,但“人类的专业知识如今比以往任何时候都更重要”。他的学生似乎也理解这一点:他们主动要求将居家开卷考试改回到手写、人工批改的形式,以检验人类知识水平,而不依赖AI工具。莱斯科维克表示,AI是一种极其强大却“非常不完美”的工具,无论是学生还是专业人士,都必须学会如何使用。“我们必须既能考察人类使用工具的能力,也能考察人类独立思考的能力。”
但当AI抬高入门级招聘门槛时,学校是否已在为学生做好充分准备来应对这种趋势?莫纳汉指出,目前备受重视的“人类技能”中,一个关键方面是“专业领域知识”。换句话说,就是人类能够具备识别AI工具何时出错的能力与知识。她表示,学生在进入职场时应当具备这种专长,但他们却在退步。美国国家教育进展评估(NAEP)数据显示,全国高三学生的阅读成绩跌至自1992年以来的最低水平,数学成绩同样下滑。约三分之一的学生不具备基本阅读能力。与此同时,美国高校的外国留学生入学人数也在急剧下降,导致长期依赖这些毕业生填补科学、技术和医学领域人才缺口的毕业生群体规模缩减。
莱斯科维克和莫纳汉的观点与沃顿商学院(Wharton School)的一项研究结果不谋而合,该研究关注AI应用的长期趋势。该研究估计,目前40%的劳动力收入可能受到生成式AI自动化的冲击,但在这些岗位中,实际会被自动化取代的任务比例仅为23%。该研究引用了麻省理工学院的一篇论文,探讨了有多少企业可能选择不全面采用AI,其研究结果与人口普查局数据所揭示的下降趋势类似。
莫纳汉表示,她认为人们未必真的不愿意使用AI,但她也不确定人们是否完全信任AI生成的内容,“这使得人工评估变得至关重要”。她补充说,这也是为何Upwork在8月发布的首份《月度招聘报告》中显示,除了对AI技术技能的需求外,企业对事实核查、翻译和领域专业知识的需求同样旺盛。她表示:“数据表明,企业并没有放弃AI,而是将其与人类独有的优势结合起来。”(*)
译者:刘进龙
审校:汪皓
德意志银行(Deutsche Bank)称之为“让AI变得糟糕的夏天”。事实上,今年夏天英伟达(Nvidia)首席执行官黄仁勋与Anthropic首席执行官达里奥·阿莫迪就后者的预测展开激烈争论。阿莫迪曾预测,人工智能将取代一半的白领岗位。整个夏天,两位高管都在频频交锋,互相抛出关于AI在“第四次工业革命”中会让多少岗位消失的预言。
但从劳动节的视角来看,情况已大不相同。8月下旬,市场意外受到冲击,原因竟是一份来自麻省理工学院(MIT)的调查:95%的大型企业生成式AI试点项目都以失败告终。这一结果引发了科技股抛售,并引起市场关于AI是否正在形成股市泡沫的讨论。而另一项最新数据也浮出水面:美国人口普查局(Census Bureau)发现,大型企业的AI采用率正在开始下降。经历两年快速试验和备受瞩目的试点项目后,许多公司似乎正在重新评估将AI纳入长期运营的实际价值。
凯莉·莫纳汉是Upwork研究院的常务董事,她掌握着来自自由职业市场的大量数据。今年9月,Upwork发布了一份关于招聘趋势和最紧缺技能的新报告。莫纳汉对《财富》杂志表示:“我看到的情况是,人类正重新回到关键环节中。我们确实看到,人类技能正变得格外稀缺和重要。人们正在意识到,即便是最先进的AI模型,依旧有10%到12%的误差。我们还无法彻底克服这一统计学问题……如今人们使用AI生成内容时发现,他们需要进行事实核查。而唯有人类才能提供事实核实。”
这一下降趋势的数据来自美国人口普查局每两周开展一次的《商业趋势与前景调查》(BTOS)。该调查覆盖超过120万家企业,能够提供不同规模企业的最新技术采用情况。最新数据显示,以六期调查的移动平均值计算,大型企业(即员工人数超过250人的公司)的AI采用率已从今年年初的峰值14%下降至2025年夏末的12%。
这一逆转是在此前几个季度的快速攀升之后出现的。大型企业的AI采用率从2023年9月的3.7%升至2024年12月的5.7%,并在2025年第二季度达到9.2%。中型企业的采用率依然较低,最高约为4.8%;而最小型企业,尤其是员工在1-4人的企业,则保持在5.5%左右,呈现温和而稳定的增长。
阿波罗全球管理公司(Apollo Global Management)首席经济学家托尔斯滕·斯洛克指出,调查中有一项问题是:企业在过去两周内是否使用了机器学习、自然语言处理、虚拟代理或语音识别等AI工具来协助生产商品或服务。斯洛克在9月7日的《Daily Spark》简报中写道:“核心结论是,美国人口普查局的双周数据已经开始显示,大型企业的AI采用进程正在放缓。”
对整体经济的影响
这一转变具有特别重要的意义,因为大型企业在塑造供应链与劳动力市场的技术趋势中发挥着巨大作用。就在一年前,“AI淘金热”还推动企业争相引入生成式AI和自动化解决方案,因为它们承诺能够带来显著的成本节约和生产率飞跃。而如今,相关表述变得更为谨慎,甚至带有怀疑的色彩。
人们同样对AI对就业市场的影响充满担忧,尤其是在入门级岗位方面。斯坦福大学经济学家进行的一项开创性研究显示,AI的显著影响才刚刚开始显现。由AI领域思想领袖埃里克·布林约尔弗森主导的研究发现,自2022年以来,22至25岁、处于职业早期的劳动者,在受AI冲击最大的岗位上就业人数相对下降了13%。美国银行研究院(Bank of America Institute)指出,从同一时间点开始,应届毕业生的失业率就已高于整体失业率。
不过,莫纳汉在随后向《财富》提供的另一份澄清声明中强调,人类技能的崛起并不仅仅是对AI采用率下滑的反应。她表示:“每一次重大技术转型都会重塑技能结构,把人们推向价值链更高端的位置。”她举例说,自动取款机并没有取代银行柜员,而是让他们转向更高价值的服务,这些服务更需要人际交往和解决问题的能力。她指出,AI也遵循着类似路径:“随着自动化承担部分任务,市场对特定人类技能的需求正在上升,包括判断力、沟通能力和情境理解能力。”
教育领域正在应对AI与人类技能的冲击
斯坦福大学(Stanford)计算机科学教授、AI初创公司联合创始人尤尔·莱斯科维克对《财富》杂志表示,几年前GPT-3的出现曾在校园里引发一场“存在危机”,但“人类的专业知识如今比以往任何时候都更重要”。他的学生似乎也理解这一点:他们主动要求将居家开卷考试改回到手写、人工批改的形式,以检验人类知识水平,而不依赖AI工具。莱斯科维克表示,AI是一种极其强大却“非常不完美”的工具,无论是学生还是专业人士,都必须学会如何使用。“我们必须既能考察人类使用工具的能力,也能考察人类独立思考的能力。”
但当AI抬高入门级招聘门槛时,学校是否已在为学生做好充分准备来应对这种趋势?莫纳汉指出,目前备受重视的“人类技能”中,一个关键方面是“专业领域知识”。换句话说,就是人类能够具备识别AI工具何时出错的能力与知识。她表示,学生在进入职场时应当具备这种专长,但他们却在退步。美国国家教育进展评估(NAEP)数据显示,全国高三学生的阅读成绩跌至自1992年以来的最低水平,数学成绩同样下滑。约三分之一的学生不具备基本阅读能力。与此同时,美国高校的外国留学生入学人数也在急剧下降,导致长期依赖这些毕业生填补科学、技术和医学领域人才缺口的毕业生群体规模缩减。
莱斯科维克和莫纳汉的观点与沃顿商学院(Wharton School)的一项研究结果不谋而合,该研究关注AI应用的长期趋势。该研究估计,目前40%的劳动力收入可能受到生成式AI自动化的冲击,但在这些岗位中,实际会被自动化取代的任务比例仅为23%。该研究引用了麻省理工学院的一篇论文,探讨了有多少企业可能选择不全面采用AI,其研究结果与人口普查局数据所揭示的下降趋势类似。
莫纳汉表示,她认为人们未必真的不愿意使用AI,但她也不确定人们是否完全信任AI生成的内容,“这使得人工评估变得至关重要”。她补充说,这也是为何Upwork在8月发布的首份《月度招聘报告》中显示,除了对AI技术技能的需求外,企业对事实核查、翻译和领域专业知识的需求同样旺盛。她表示:“数据表明,企业并没有放弃AI,而是将其与人类独有的优势结合起来。”(*)
译者:刘进龙
审校:汪皓
Deutsche Bank called it “the summer AI turned ugly,” and there was certainly a heated war of words between Nvidia CEO Jensen Huang and Anthropic CEO Dario Amodei over the latter’s prediction that artificial intelligence would wipe out half of all white-collar jobs. The two executives spent much of the summer trading predictions about how many jobs would be lost as AI transformed the workforce in a “fourth industrial revolution.”
But from the standpoint of Labor Day, things are looking very different. Markets got a shock in late August from an unlikely place: a survey by MIT finding that 95% of generative AI pilots at large companies were failing. That prompted a tech sell-off and talk of whether AI was forming into a stock-market bubble. And another piece of the puzzle just fell into view: the Census Bureau finds that AI adoption rates are starting to decline among major firms. After two years of rapid experimentation and headline-grabbing pilot projects, many corporations appear to be reassessing the real-world value of integrating AI into their operations for the long haul.
Kelly Monahan is managing director of the Upwork Research Institute, where she is plugged into reams of data from the freelance market. In September, Upwork launched a new report into the hiring trends and skills that are most in demand. “What I’m seeing happening is the humans are coming back into the loop,” Monahan told Fortune. “We’re actually seeing the human skills coming into premium,” she said. “I think what people are realizing is even the best AI models still hallucinate 10% to 12% of the time. “We just cannot necessarily overcome that statistical problem yet … I think what people are seeing, now that they’re using AI-generated content, is that they need fact-checking.” Only a human can provide that.
The AI adoption decline figures come from the Business Trends and Outlook Survey (BTOS), conducted biweekly by the U.S. Census Bureau, which covers more than 1.2 million firms and captures a unique, up-to-date view of technology adoption across different business sizes. The most recent data, reflected in a six-survey moving average, shows that the AI adoption rate among large companies—defined as those with more than 250 employees—has dipped from a peak of 14% earlier this year to 12% as of late summer 2025.
This reversal follows a steep climb over previous quarters, where large firm adoption jumped from 3.7% in September 2023 to 5.7% by December 2024, and reached 9.2% in the second quarter of 2025. Medium-sized firms remain less likely to adopt AI, with maximums around 4.8%, while the smallest businesses, especially those with one to four employees, still report a modest but steady adoption rate of 5.5%.
Apollo Global Management chief economist Torsten Slok notes that one question in the survey is whether a business has used AI tools such as machine learning, natural language processing, virtual agents, or voice recognition to help produce goods or services in the past two weeks. “The bottom line is that the biweekly Census data is starting to show a slowdown in AI adoption for large companies,” Slok wrote in his Daily Spark newsletter on September 7.
Implications for the broader economy
This shift is particularly significant given the outsized role large companies play in shaping technology trends across supply chains and labor markets. Just a year ago, the “AI gold rush” saw firms racing to integrate generative AI and automation solutions, fueled by promises of dramatic cost savings and productivity leaps. Today, the narrative is more cautious—or even skeptical.
There’s also considerable anxiety around what AI is doing to jobs—especially at the entry level, where a first-of-its-kind study by Stanford University economists suggested the beginning of a considerable impact. The research, led by AI thought leader Erik Brynjolfsson, revealed a 13% relative decline in employment for early-career workers aged 22 to 25 in the most AI-exposed jobs since 2022. The Bank of America Institute has pointed out that the unemployment rate for recent graduates started trending ahead of the overall unemployment rate starting at exactly the same time.
To be sure, Monahan said in a separate, clarifying statement provided to Fortune, she thinks the rise of human skills is not simply a reaction to declining AI adoption. “Every major technological shift has reshaped the skill mix by moving people up the value chain,” she said, offering the example of how the ATM didn’t eliminate bank tellers, but elevated their role toward higher-value services that required more interpersonal and problem-solving skills. AI is following a similar pattern, she said. “As automation takes on certain tasks, demand is growing for distinctly human skills—judgment, communication, and contextual understanding.”
Education grapples with AI and human skills
Stanford computer science professor and AI startup co-founder Jure Leskovec told Fortune that the arrival of GPT-3 years ago triggered a sort of “existential crisis” on campus, but that “human expertise matters much more than it ever did.” Leskovec’s students seem to grasp that: They requested that he go from take-home, open-book exams back to a hand-written and hand-graded model, to test human knowledge without AI tools. He called AI an amazingly powerful and “very imperfect” tool that students and professionals alike need to learn how to use, “and we need to be able to both test the humans being able to use the tool and humans being able to think by themselves.”
But are schools preparing students properly for this moment, when AI is raising the bar on entry-level hiring? Monahan said that a key aspect of the “human skills” that are valued at a premium right now is “domain expertise,” or in other words, a human with the skills and knowledge to spot when an AI tool is making a mistake. Just as students need to be entering the job market with this expertise, they are sliding backwards. High school seniors nationwide just recorded their worst reading scores since 1992, with math scores also falling. About a third of them didn’t have basic reading skills, according to the National Assessment of Educational Progress. U.S. universities are also seeing a sharp drop in foreign student enrollment, reducing the size of a group of graduates long relied upon to fill talent gaps in science, technology, and medicine.
Leskovec and Monahan’s experience line up with another study, by the Wharton School, on the long-term likelihood of how AI adoption will play out. While the research estimates that 40% of current labor income is potentially exposed to automation by generative AI, only 23% of actual tasks in those roles will be automated. The study cited an MIT paper that tackled the subject of how many companies would simply choose not to bother with full adoption—and found something like the dip revealed by the Census Bureau.
Monahan said that she doesn’t think people are shying away from using AI necessarily, but she’s not sure that people fully trust AI-generated content, and “this makes human evaluation essential.” She thinks that’s why Upwork’s inaugural Monthly Hiring Report in August showed strong demand for fact-checking, translation, and domain expertise alongside technical AI skills, she added. “The data shows businesses aren’t abandoning AI, but instead are pairing it with uniquely human strengths.”