--- author: House of El - AI date: '2026-05-18' guest: '' layout: post.njk source: https://www.youtube.com/watch?v=5EvIY_2TWN8 speaker: House of El - AI tags: - ai-adoption - job-displacement - jevons-paradox - return-on-investment - total-cost-of-ownership title: AI裁员的真相：为何牺牲员工未能带来回报 summary: 一份Gartner研究显示，80%尝试AI的公司进行了裁员，但这与投资回报率并无关联。真正成功的企业并非用AI取代员工，而是赋能员工以提高生产力。该现象背后涉及“AI粉饰”的公-关策略、未被充分计算的AI运营隐性成本，以及“杰文斯悖论”所揭示的新岗位创造效应。最终数据显示，人机结合才是最具生产力的路径。 insight: '' draft: true series: '' category: ai-application area: tech-engineering project: [] people: [] companies_orgs: - Gartner - OpenAI - Anthropic - World Economic Forum - LinkedIn products_models: [] media_books: [] status: evergreen --- ### AI裁员的悖论：为何牺牲员工未能带来回报 上周发布的一则新闻标题揭示了一个惊人的发现。一项针对350家年收入至少10亿美元的全球商业领袖的**Gartner研究**显示，80%试点了AI或自动化技术的公司都进行了裁员。这是一个令人震惊的数字，但研究的发现并未止步于此。该研究同时指出，这些裁员与更高的投资回报率之间**毫无关联**。因AI而裁员的公司，并非那些从AI中获利的公司。真正实现高投资回报的公司，其做法截然不同：他们留住员工，并利用AI来增强员工的生产力。Gartner的副总裁分析师Helen Poitevin直言不讳地指出：“仅仅通过削减人力来追逐价值，很可能将大多数组织引向一条回报有限的道路。” 换言之，企业正在以AI的名义解雇员工，但这并未奏效。他们正在向一个既未要求也未奖赏祭品的“神”献上“活人祭”。 这引出了一个显而易见的问题：他们为何要这么做？首先，存在一种普遍持有但未必有证据支持的真诚信念，即AI能够大规模取代人类工人。这种逻辑听起来很直观：如果机器能做人的工作，那么人就不再被需要。然而，Gartner的数据表明这种逻辑是不完整的，因为事实证明，AI的价值不在于取代人类判断，而在于**放大**人类判断。回报率最高的公司，正是那些将AI视为员工使用的工具，而非完全替代员工的公司。这更像是“我们为飞行员配备了更先进的仪表，于是飞机准点降落，无人伤亡”，而不是“我们用自动驾驶取代了飞行员”。两者都是自动化故事，但只有一个结局是好的。
Original English Source So, here is a headline that landed last week. A Gartner study of 350 global business executives, all at companies with at least a billion dollars in annual revenue, found that 80% of those who have piloted AI or autonomous technology have carried out workforce reductions. 80% That is a staggering number, and if that's where the finding stopped, this would be a very short and very depressing video. But, that is not where the finding stopped. The study also found that there was no correlation between those layoffs and higher returns on investment. The companies cutting staff because of AI were not the ones making money from AI. The companies actually seeing high return of investment were doing something different entirely. They were keeping their people and using AI to make them more productive. Gartner's VP analyst Helen Poitevin put it very plainly, "chasing value only through headcount reduction is likely to lead most organizations down a path of limited returns." So, companies are firing people in the name of AI, and it's not working. They're performing a human sacrifice to a god that didn't ask for one and isn't rewarding it. That raises an obvious question. Why are they doing it? My name is L. I have a PhD in computer science, and this is the kind of question I find it very difficult to leave alone. So, let's get into it. A quick note before we go further, given some past comments on this channel, if your first instinct is to dismiss survey data, remember that this is the same statistical methodology that approved every medication you've ever taken. Gartner's sample of 350 billion-dollar companies uses the same representative sampling as clinical drug trials. Yes, I know it's not perfect, but dismissing it means dismissing the foundation most of modern science is built on. So, let's engage with the data seriously. Now, why are companies laying people off if it's not generating returns? There seems to be a few things happening simultaneously. First, there is a genuine belief, not necessarily supported by evidence, but widely held, that AI can replace human workers at scale. Another gentler, but equally accurate word for that is hope. The logic sounds intuitive. If machine can do what a person does, you don't need the person. But the Gartner data suggest this logic is incomplete because it turns out the value of AI isn't in replacing human judgment, it's in amplifying it. The companies seeing the highest returns are the ones treating AI as a tool their employees use, not a substitute for having employees at all. Think of it less as we replace the pilot with autopilot, and more we gave the pilot better instruments, and now the plane lands on time and nobody dies ever. Both are automation stories, only one of them ends well.
### “AI粉饰”与不断变化的叙事 除了错误的信念，另一种现象是OpenAI首席执行官**Sam Altman**所描述的“**AI粉饰**”（AI washing）。他认为，一些公司本就会进行裁员，但将此归因于AI，使其听起来具有战略性和前瞻性，远比承认“我们在疫情期间过度招聘，现在首席财务官终于注意到了”要好听得多。当然，Altman作为全球最著名的AI公司负责人，其言论有其自身动机，他不希望AI与大规模失业联系在一起。这提醒我们，任何公众人物的言论背后都有其未言明的原因。 这种叙事的易变性在Anthropic公司CEO **Dario Amodei**身上也得以体现。他去年曾声称AI可能淘汰约半数的白领入门级工作，而今年他却援引“**杰文斯悖论**”（Jevons paradox）修正了看法，认为AI可能更多是增强而非取代工作。在短短12个月内，预测就从“你一半的工作将消失”变为“情况复杂，或许不会”，这一转变本身就值得注意，尤其是当两位最重要的AI高管都在公开调整他们对工作替代的预测时。
Original English Source Second, there is what Sam Altman, the CEO of OpenAI, has described as AI washing. His argument is that some companies are attributing layoffs to AI when they would have made those cuts anyway. AI becomes a convenient narrative, it sounds strategic, it sounds forward-thinking, it sounds like you're positioning for the future, it sounds a lot better than we over-hired during the pandemic and our CFO finally noticed. Now, Altman has his own motivations for saying this. He runs the most prominent AI company in the world and has an obvious interest in AI not being associated with mass unemployment. That's like the CEO of a fireworks company assuring you that the house fire was probably electrical. I'm not saying he's wrong, I'm just saying that when anyone, CEO, analyst, politician, literally anyone, makes a public statement, they have reasons for making it and those reasons are not always the ones they state. That is not cynicism, it is just how information works. People have motivations, the truth is hard to get to partly because people might not always be forthcoming, but also because even genuine beliefs change constantly as new data emerges. Which brings me to Dario Amodei, the CEO of Anthropic, the cloud people. Last year, he made a widely reported claim that AI could eliminate roughly half of white-collar entry-level jobs. This year, he walked that back, citing something called the Jevons paradox, which I'm going to explain in a moment, because it's generally fascinating, and suggested AI might augment work rather than eliminate it. Though he cautioned that AI is evolving faster than previous technologies and could produce different outcomes. So, we've basically gone from half your jobs will disappear to actually maybe not, it's complicated, in about 12 months. Which, to be fair, is also how long it takes most companies to finish an AI pilot program. So, at least the timeline is consistent. Again, motivations exist. Amodei runs a company that builds AI and has a commercial interest in that technology being adopted widely. But, the fact that two of the most prominent AI executives are now publicly moderating their predictions about job displacement is, at minimum, worth noting.
### 杰文斯悖论：效率提升反向引爆需求 杰文斯悖论是一个听起来违反直觉但历久弥新的经济学理念。1865年，英国经济学家**William Stanley Jevons**观察到，当蒸汽机效率大幅提升，生产相同能量所需煤炭减少时，总煤炭消耗量非但没有下降，反而急剧上升。因为煤炭变得更便宜、更高效，人们为其找到了更多用途，催生了新的产业和应用。效率的提升并未减少需求，而是引爆了需求。 Apollo首席经济学家**Torsten Slok**认为，同样的悖论也适用于AI。如果AI使某些类型的工作变得更便宜、更快速、更易于获取，那么对这些工作的需求非但不会消失，反而会扩张。**世界经济论坛**报告称，AI已在全球创造了超过130万个新工作岗位。LinkedIn的《2026年就业新趋势》报告将“AI工程师”列为美国增长最快的工作，职位发布量同比增长143%，前五名增长最快的职位中有四个与AI相关。这些新角色不仅限于工程，还包括**提示工程师**（prompt engineers）、**AI训练师**（AI trainers）和**AI治理专家**（AI governance specialists）等。一个三年前甚至不存在的职业——提示工程师，如今已成为市场上增长最快的职位之一，其增长率高达135.8%。我们发明了一项如此强大的技术，以至于需要一个全新的专业来专门研究如何正确地向它提问。
Original English Source Now, let's talk about the Jevons paradox, because it's one of those ideas that sounds counterintuitive, but has been surprisingly durable. Back in 1865, an English economist named William Stanley Jevons observed something strange about coal. The steam engine had just become dramatically more efficient, meaning you needed less coal to produce the same amount of energy. You would expect, logically, that total coal consumption would decrease, but it didn't. It actually increased massively. Because when coal became cheaper and more efficient to use, people found more uses for it. New industries became viable, new applications emerged. The efficiency gains didn't reduce demand, they exploded it. The Victorians basically invented, "We'll save money by spending more," 160 years before every SaaS company on Earth independently rediscovered it. Apollo's chief economist, Torsten Slok, has argued that the same paradox now applies to AI. If AI makes certain types of work cheaper, faster, and more accessible, the demand for that work will not disappear, but rather it will expand. New applications emerge, new roles get created, new industries become viable that were not viable before. Is he right? I don't know. Nobody really does, but the early data is very interesting. Let's have a look at that. The World Economic Forum reports that AI has already created more than 1.3 million new jobs globally, including roles that either didn't exist or existed in negligible numbers before 2023. LinkedIn's 2026 Jobs on the Rise report ranked AI engineer as the number one fastest growing job title in the United States, with job postings rising 143% year-over-year. Four of LinkedIn's top five fastest growing positions are AI-related. AI and machine learning job posting surged 163% from 2024 to 2025. And these are not just engineering roles, by the way. Prompt engineers design inputs for AI systems to produce reliable outputs. AI trainers evaluate and correct AI-generated content to improve model performance. AI governance specialists and model evaluators make sure the systems don't do something catastrophic. There is an entire ecosystem of roles emerging around the question of how to make AI systems work properly, roles that require human judgment, domain expertise, and contextual understanding that the technology itself doesn't have right now. Three years ago, prompt engineer wasn't even a job title. Now, it's one of the fastest growing positions in the market with a 135.8% growth rate. We have invented a technology so powerful that it requires a brand new profession dedicated entirely to asking it questions in the right way. That is either a breakthrough or the most expensive auto complete in history and honestly it might be both.
### 自动化的隐形成本：从固定薪资到“老虎机预算” 即使是构建AI模型的公司，也依赖庞大的人力队伍来训练它们。AI并非自我训练，其背后是一个由教导它如何行为的“隐形”全球劳动力组成的体系。这使得AI更像一个昂贵且会犯错的实习生，而非一个卓越的新员工。在这一背景下，另一个值得关注的维度是**成本**。一个深刻的观点是：“工人的薪水你可以控制，AI代理的成本你无法控制。” 数据支持了这一观点。与成本固定且已知的受薪员工不同，AI代理的成本与**代币消耗、API调用、云端计算和交互量**等难以预测的变量挂钩。集成成本经常超出初步估算的30%到50%，而生产一个AI代理的初始开发成本仅占公司三年内总支出的25%到35%。其余的都是代币、基础设施、安全、监控、再训练等一系列未在原始提案中列出的项目。例如，一家SaaS公司发现，由于各团队未经集中监督而启动了23个未记录的AI服务，每月产生了28万美元的未入账云支出。此外，**供应商锁定**（vendor lock-in）问题也日益突出，企业围绕特定AI供应商构建运营体系后，会发现自己被该供应商锁定，从而降低了灵活性并增加了长期成本。你解雇了一个薪水可以年度谈判的员工，却换来一个定价模型由供应商酌情每季度更改的服务，其成本规模在你收到账单前无法预测。你用一个已知的成本换来了一张“神秘账单”，恭喜你，你已经将你的预算自动化成了一台“老虎机”。
Original English Source Even the companies building AI models rely on enormous human workforces to train them. Data labeling companies like Search AI hit over 1.2 billion dollars in annual revenue by connecting thousands of domain experts, doctors, lawyers, scientists, engineers with AI labs that need human feedback to improve their models. The AI isn't training itself. It requires an invisible global workforce of people teaching it how to behave which if you really think about it makes AI less like a brilliant new employee and more like a very expensive intern who keeps confidently doing things wrong until a senior person corrects them and teaches them better. Except the intern costs 4 billion dollars and occasionally hallucinates. So the picture is more complex than AI takes everyone's jobs. AI is taking some jobs. AI is also creating new jobs and nobody can really tell you yet which effect will be larger in the long run. That uncertainty is definitely generally uncomfortable and I understand why people find it frightening rather than reassuring. It might work out as not exactly a pension plan. But there is another dimension to this that I think deserves more attention and it has to do with costs. There is a line that I came across that stuck with me a lot. Workers had salaries you could control. AI agents have costs you cannot. And the data backs this up actually. AI operations fluctuate unpredictably. Unlike a salaried employee whose cost is fixed and known AI agent cost are tied to token consumption, API calls, cloud compute, and interaction volume. All of which vary and are difficult to forecast. Integration costs regularly exceed initial estimates by 30 to 50%. The initial development of a production AI agent represents only 25 to 35% of what a company will spend over 3 years. The rest is tokens, infrastructure, security, monitoring, retraining, and a dozen other line items that weren't in the original proposal. One SaaS company, for example, discovered $280,000 in monthly unaccounted cloud spend from 23 undocumented AI services that various teams have spun up without centralized oversight. 23 rogue AI subscriptions, that is not exactly a technology strategy. It's more corporate haunting. And then there's vendor lock-in. Companies that build their operations around a specific AI provider's ecosystem, their APIs, their models, their infrastructure, find themselves locked into that vendor, reducing flexibility and increasing long-term costs. You fire the person whose salary you negotiated annually, you replace them with a service whose pricing model changes quarterly at the vendor's discretion, and whose cost scaling ways you cannot predict until the invoice arrives. You swapped unknown costs for a mystery bill. Congratulations, you've automated your budget into a slot machine. This doesn't mean AI is a bad investment, by the way. The companies in the Gartner study that used AI for people amplification are seeing strong returns, but it does mean that the calculus of replace workers with AI to save money is far less straightforward than the pitch deck makes it sound. You're not exactly swapping a salary for a subscription, you're swapping unknown costs for an unknown one and hoping the math works out.
### 结论：神祇需要的是操作员，而非祭品 让我们将所有线索汇集起来。2026年3月和4月，AI是导致裁员的首要原因。然而，Gartner的研究告诉我们，这些裁员并未产生预期的回报。杰文斯悖论暗示AI可能创造比摧毁更多的工作，而就业市场数据也显示全球已创造超过一百万个与AI相关的新职位。同时，成本数据表明，用AI取代人类并非一个清晰的成本节约方案。这一切都不意味着万事大吉，人们正在真实地失去工作，而技术变革的阵痛从未被均匀地分配过。 然而，数据确实揭示了最具生产力的前进道路并非用AI取代人类，而是将两者**结合**起来。取得最佳成果的公司不是那些解雇员工的公司，而是那些用能使其更快、更敏锐、更强大的工具来武装员工的公司，这种能力是人类或AI单独无法达到的。那位“神祇”不想要你的员工做祭品，它想要一个称职的操作员。
Original English Source Here is where I want to bring it all together. AI was the leading reason cited for layoffs in March and April of 2026. Outplacement firm Challenger, Gray, and Christmas found that 49,135 jobs have been attributed to AI so far this year, nearly matching the entire total of 2025, and it's only May. Those are real people, real livelihoods. That number is not exactly abstract, but the Gartner study tells us those layoffs aren't generating the returns companies expected. The Jevons paradox suggests AI might create more work than it destroys. The job market data shows over a million new AI-related roles already created globally, and the cost data shows that replacing humans with AI isn't a clean savings exercise that the term automation implies. The word automation is doing an enormous amount of heavy lifting in boardroom presentations right now. It's carrying the entire pitch on its back while the actual numbers quietly file for workers compensations. None of this, of course, means that everything is fine. People are losing their jobs right now, today, in this economy, and telling them the Jevons Paradox suggests your suffering may be temporary is not exactly helpful when rent is due on the 1st. The pain of technological change is never distributed evenly. It never has been. We actually covered the parallels with the former industrial revolution in a separate video that I'm going to link in the description. Some industries will be hit harder than others. Some roles will disappear entirely. Some people would benefit enormously while others bear the cost. That is not a prediction about AI. It's just a description of how every major technological transition in history has played out. What the data does suggest is that the most productive path forward isn't replacing people with AI. It's rather combining them. The companies seeing the best results are not the ones firing the workforce. They're the ones arming their workforce with tools that make them faster, sharper, and more capable than either humans or AI could be alone. The god doesn't want your people. It wants a competent operator. But if the data says combining works, why are so many people still furious about AI? It's not because they haven't seen the data. It's because the data isn't really the point. I made a video exploring where that anger actually comes from and why it might be aimed at the wrong target. That's the one that I would watch next. Thank you all so much for watching this one. I'll see you all in the next one.