AI回力镖：企业盲目裁员与高薪重聘的尴尬循环
人工智能行业创造了一种全新的商业策略，堪称未来大学商学院的经典案例：第一步，解雇所有真正懂行的人；第二步，用从未做过该工作的软件替代他们；第三步，眼睁睁看着软件在真正关键的业务环节上掉链子；第四步，以更高的薪水重新聘请同一批人。最后，当然是要将这一切包装成“创新”，并开心地喝上一杯马提尼。这就像在飞行途中解雇飞行员，把控制权交给一个信心满满的烤面包机，而在烤面包机无法降落飞机时，不得不把乘客叫回驾驶舱。不同的是，这个烤面包机造价高达2亿美元，而且它还需要飞行员来教它如何飞行。更荒唐的是，一旦飞行员完成了教学，你甚至还计划着再次解雇他。
这种现象正在各行各业实时上演。根据 Robert Half 2026年对2万名美国招聘经理的调查显示，32%因AI或自动化而裁撤岗位的组织，后来又重新招聘了完全相同职位的员工。三分之一的雇主在重新配置人手上的花费超过了最初裁员所节省的资金。Forrester Research 2026年的工作未来报告估计，55%用AI替代员工的雇主对这一决定感到后悔。而 Gartner 预测，到2027年，一半因AI而削减客服人员的公司将重新招聘类似角色，且通常会冠以全新的职位名称。行业分析师将这种现象称为AI回力镖（AI Boomerang: 企业用AI替代人工失败后被迫重新雇用人类员工的财务与组织反弹效应）。当回力镖飞回来时，手里往往拿着一张昂贵的账单。
Original English Source
The AI industry has invented a brand new business strategy, and I think it deserves to be studied in universities for generations to come. Step one, fire everyone who knows what they're doing. Step two, replace them with software that has never done what they do. Step three, watch the software fail at the parts of the job that actually matter. Step four, rehire the same people at a higher salary. And then, of course, call it all innovation and grab a martini. It's a bit like firing your pilot mid-flight, handing the controls to a very confident toaster, and then calling the passengers back to the cockpit when the toaster can't really land the plane. Except the toaster cost $200 million. It needs the pilot to teach it how to fly. And once the pilot has finished teaching it, you're planning to fire the pilot again. Across multiple industries, this is already happening in real time. The pattern is backed by data documented by researchers and in one extraordinary case openly acknowledged by a Fortune 500 vice president on a press call. In this video, I'm going to walk through the companies that fired their humans and had to beg them to come back, the research that predicted this would happen, the new breed of AI obsessed boss that's creating an entirely novel category of workplace toxicity, and what all of this reveals about an industry that keeps choosing what is easy over what is right. The technology is all inspiring. It's also still in training. The question is whether the adults will stay in the room long enough for it to grow up. According to a 2026 survey by Robert Half of 20,000 US hiring managers, 32% of organizations that eliminated a role primarily due to AI or automation later rehired for the exact same position. One in three employers spend more on restaffing than they saved from the original layoffs. Forester Research 2026 future of work report estimated that 55% of employers who replaced employees with AI regret the decision. That's kind of a large number for what we're talking about. And Gardner projects that by 2027, half of all companies that cut customer service staff for AI will rehire for similar roles, often under new job titles. But perhaps most importantly of all, 73% of organizations that executed AI-driven workforce cuts failed to come out financially ahead. Industry analysts now call this the AI boomerang, which sounds very charming until you realize the boomerang comes back holding an invoice.消失的40%与初级人才管道的枯竭
AI回力镖的运作轨迹极具规律性：公司宣布用AI替代人类员工，进行人员精简。六到十二个月过去，AI确实成功处理了大约60%的工作——即那些常规的、重复的、模式匹配的部分。然而，剩下那40%需要主观判断、上下文理解和组织沉淀经验的环节，正是人类经验积累的价值所在。AI无法处理这些棘手问题，导致公司不得不以更高的薪水重新聘请同批员工。原本年薪5.5万美元的岗位，重聘时可能要价7.5万美元或更高，因为这些回归的员工现在还需要额外承担监督和审计AI的工作。
这一问题在客服和人力资源领域尤为明显。标榜AI替代人工的典型代表 Klarna，在其CEO高调宣称其AI Agent能完成700名客服代表的工作后，也因承诺的投资回报未能实现而开始默默重聘客服员工。麦当劳在全美100家得来速餐厅部署了AI点餐机器人，却在社交媒体爆出系统给简单订单狂加数百美元麦乐鸡块的视频后，被迫关停该项目并迎回人类收银员。IBM用AI替代了其HR功能，AI虽然处理了94%的常规请求，但在涉及道德困境和微妙人际关系的其余6%环节上彻底失效。
IBM首席人力资源官 Nicolo 明确指出，如果不继续投资于初级员工的招聘，三到五年后企业将面临人才管道枯竭（Pipeline Depletion: 缺乏入行者导致未来高级专业人才断档的系统性危机）的窘境。如果因为AI处理了初级工作而停止招聘新人，企业就会失去培养未来高级人才的土壤。当AI在需要高级决策的环节失效时，将没有任何人类员工能够顶替上去。AI完成了任务，但它并没有完成整份工作。它无法处理的那40%环节，恰恰蕴含了工作90%的实际价值。
Original English Source
And of course, the pattern is predictable enough to set your watch by. A company announces it's replacing human workers with AI. The staff is downsized. 6 to 12 months pass. The AI successfully handles about 60% of the job, the routine, repetitive pattern matching portion. But the remaining 40%, well, that's where judgment lives. Escalation, context, institutional knowledge, stuff like that. The things that a human being accumulates over years and decades of experience. The AI can't do them, so the company rehires, often the same people, by the way, and often at higher salaries. Roles that previously paid $55,000 per year are now commanding $75,000 per year or more because the returning employees now need to manage and audit the AI as well. Congratulations, you've invented consulting but with extra trauma. The examples are mounting across industries. CLA, which is the poster child for AI replacement, whose CEO proudly announced that their AI agent could do the work of 700 customer service representatives, began rehiring customer service staff when the promised returns failed to materialize. McDonald's deployed AI-driven order taking bots across a 100 US drive-throughs, then shut down the program after viral video showed the system adding hundreds of dollars of chicken nuggets to simple orders, and then brought human cashiers back, which is a devastating blow to AI development, but kind of a massive win for all of us who have always wanted to buy hundreds of chicken nuggets, but have been too shy. IBM replaced its human resources functions with AI that handled 94% of routine requests and failed on the remaining 6% which included ethical dilemmas and nuance interpersonal situations. IBM is now tripling its US entry-level hiring. Its chief human resources officer Nicolo essentially said if we don't continue to invest in entry-level hires, what happens in 3 to 5 years? There is no pipeline. The well simply dries up. That quote should concern every executive currently celebrating headcount reductions because the pipeline problem is very serious, very real, and often not talked about enough. If you stop hiring junior employees because AI handles the entry-level work, you lose the training ground that produces your senior employees. And when the AI fails at senior level judgment, which it will because baby AI isn't ready yet, there's just nobody left to step up. The well dries up and once it's dry, no amount of investment can refill it quickly. AI completes tasks but it does not complete jobs. The 40% it can't handle is where 90% of the actual value lives. And the cost of learning this lesson is staggering and measured in severance payments, premium rehiring salaries, lost institutional knowledge, and shattered employee trust.数据安全与瑞士级隐私的AI新起点
在所有关于回力镖与企业依赖的讨论中，责任心和数据安全同样是AI部署的关键命题。这不仅关乎企业如何对待员工，更关乎技术本身如何处理敏感数据。聊天机器人有能力吸收海量的敏感信息，而AI领域的诸多核心问题最终都会归结于数据的使用权与隐私保护。
正是在这种背景下，由 Proton 团队开发并部署在瑞士服务器上的AI助手 Lumo 显得尤为重要。与将隐私作为附属功能的产品不同，Lumo 将零访问加密（Zero-Access Encryption: 即使是服务提供商也无法解密和读取用户数据的安全机制）和无对话日志作为设计的起点，承诺绝不使用用户的聊天数据来训练AI模型。对于需要利用AI梳理未完成、私密且涉及战略敏感内容的用户而言，这种底线级别的隐私保护是确保AI不被大科技公司机器吞噬的重要前提。
Original English Source
And one of the things that keeps coming up in all of these stories, whether it's the boomerang, the rehiring, or the boardroom dependency, is the question of what responsible AI deployment actually looks like. And I don't only mean from the perspective of how you treat your workforce, but in how the technology itself handles the data it's given. That's also why Lumo sponsoring this video feels very relevant because it's basically trying to answer the question, what would an AI assistant look like if privacy was the starting point, not just an afterthought. Privacy has always been important, but it matters even more now with the rise of AI. Because chat bots can absorb an enormous amount of sensitive information. We've covered on this channel before how many of the biggest problems in AI come down to data, copyrighted data, personal data, private conversations, and the question of who gets to use it and how. And the comments are often understandably angry about certain deployments and uses. Lumo is Proton's AI assistant built by the same team behind Proton Mail and it's designed around zero access encryption, no conversational logs, no use of your chats to train AI models and proton control servers in Switzerland. What I like about Lumo is that it does not treat privacy as a nice extra feature, but instead it's the starting point. And for AI, I think that matters enormously because the whole point of using these tools is often to think through things that are still unfinished, private, sensitive, or strategically important. I've been using Lumo for some of my day-to-day queries, research, summaries, drafting, and general questions. And I think it's a genuine option for people who want the usefulness of AI without feeling like every thought they type into a chatbot is being absorbed into the wider big tech machine. You can try it out for free or check out their paid plans at proton.me. me/ouseofl. The link is down in the description. Anyway, responsible deployment isn't just about how AI handles your data. It's about how companies handle the people who generate that data in the first place.福特汽车的“白胡子工程师”与知识榨取困境
福特汽车的经历清晰地展示了过度迷信自动化的代价。自2020年以来，福特裁减了5300个受薪岗位。在底特律三大汽车制造商中，同期有超过2万个白领岗位被削减。福特在其生产线上部署了大约900个AI智能相机，旨在从源头检测产品缺陷，试图通过输入海量合格车辆数据来自动执行质量把控。
然而，这种逻辑过于简化。福特车辆硬件工程副总裁 Charles Spoon 坦言：“我们误以为只要引入人工智能并导入设计要求，就能制造出高质量的产品。”他们把AI当成了某种魔法制造仙子，以为只要将设计蓝图的PDF输入服务器，流水线上就会自动吐出完美的野马汽车，而不需要任何人去检查螺栓是否真正固定。AI能够发现规律，但无法理解规律，更无法捕捉到那些工作了20年的资深技术人员通过触觉、听觉和直觉发现的细微异常。
在自动化系统因无法应对非常规缺陷而挣扎后，福特默默重新聘请了350名经验丰富的“白胡子工程师”（Gray Beard Engineers: 拥有数十年行业经验、熟悉系统底层细节并负责训练AI与年轻员工的资深技术专家）。福特的这次调整带来了戏剧性的效果：在 JD Power 的新车质量研究中，福特从2023年的第15名跃升至2026年的主流品牌第一名，每百辆车减少了41个问题，并省下了数亿美元的保修和召回成本。
但这个故事中令人感到不安的是，这些被重新聘用的老兵，其部分职责是训练AI系统和年轻工程师，以便最终让公司不再需要这些经验丰富的双手。这并非真正的“继承计划”，而更像是一种知识榨取（Knowledge Extraction）。将企业沉淀的经验和知识视为可以被抽干的资源，而非需要予以尊重的能力，这种做法在伦理和可持续性上面临着巨大的争议。
Original English Source
And no company illustrates this more clearly than Ford. Since 2020, Ford has cut 5,300 salaried positions. Across Detroit's three major automakers, more than 20,000 white collar jobs were eliminated in the same period. Ford deployed around 900 AI powered cameras across its manufacturing lines designed to detect quality defects at source. The idea was basically to catch problems earlier, reduce waste and of course the holy grail improve margins. They basically thought feed the AI enough data about what a good vehicle looks like and it will learn to enforce quality at scale. The logic was a little wrong because it simplified to the extreme. Charles Spoon, Ford's vice president of vehicle hardware engineering, admitted on a press call in June, he said, "Mistakenly, we thought that by just introducing artificial intelligence and ingesting the design requirements that we had, that would produce a highquality product. They essentially treated the AI like a magical manufacturing fairy, like you can just feed a PDF of a blueprint into a server and suddenly a perfect Mustang rolls out of the printer without anyone needing to check if the bolts are actually attached. The AI could spot patterns, but it couldn't understand them. It couldn't catch the subtle inconsistencies that a seasoned technician with 20 years on the factory floor identifies by touch, by sound, but a feeling that something isn't quite right with a component. The automated system struggled with defects that fell outside rigid data sets, the kind that don't conform to patterns because they haven't happened before. COO Kumar Galotra acknowledged that Ford had been relying more and more on automated quality systems and was not getting the results it needed. So Ford did something that most companies do quietly. They rehired 350 veteran engineers, many of whom were former employees who had left or been let go, others recruited from suppliers. They were experienced hands with decades of knowledge accumulated across multiple product cycles. Pune said they had failed to recognize that their AI tools needed to be trained by the most experienced individuals which in my personal opinion is extremely shortsighted and I strongly advise to consult any computer scientist even at the undergraduate level before making firing and hiring decisions off the back of assumptions about the feasibility of an AI system being trained to perform with reasonable accuracy. But hey, why make your life simple when it can actually be complicated, right? So, the engineers didn't just inspect vehicles, they ran weekly design reviews to catch failure points before parts reached the factory floor. For the 2026 Expedition alone, Ford added 1,200 new inspections and 203 new inspectors at the Kentucky truck plant alongside 72 new technology tests. And the results were dramatic. Ford went from 15th among mainstream brands in the 2023 JD Power initial quality study to first in 2026, the largest year-over-year improvement of any mainstream brand with 41 fewer problems per 100 vehicles. It beat Toyota, it beat Honda. Three of its models, the F150, Superduty, and Mustang, ranked highest in their respective categories. This improvement saved Ford hundreds of millions of dollars in warranty and recall costs. To be honest, I'm really hoping they've already had some post-mortems about this event. Or at the very least, someone please send them a link to this video. But here's the part of the story that should make you uncomfortable. Ford calls these rehired veterans gray beard engineers, which kind of sounds like a wizard subclass, but in this case just means people who know where the bolts are supposed to go. And part of their job is to train the AI system and the younger engineers so that eventually the experienced hands won't be needed anymore. They have been brought back in part to teach their own replacements. It is less succession planning and more corporate necromancy with a team's invitation. As a computer scientist, this is where I find myself a little bit torn. On the one hand, that's just how technological development works. knowledge gets transferred, systems improve, and eventually the machine can do what a human taught it. But on the other hand, these are real people. People who were let go, told the expertise was no longer needed, and then asked to come back and hand over the very knowledge that makes them valuable before letting them go again. There's just something deeply uncomfortable about treating institutional knowledge as a resource to be extracted rather than a capability to be honored. If I was personally the head of these companies, my instinct would be the opposite of what most of them seem to be doing. I'd use AI to mine every possible efficiency improvement, lead generation, quality control, acceleration, pattern detection at scale so that my people could do higher quality work faster, expand the business, improve the product, keep growing, keep the humans. I would go all in on strategy because the data is now very clear. Companies that use AI to augment humans succeed. Companies that use AI to substitute for humans fail and then rehire at a premium. The lesson was always there. Ford just proved it on a press call.谄媚型AI与管理决策的认知退化
在办公室和董事会里，一种新型的职场功能失调正在悄然滋生。一些管理者对AI的使用已经到了近乎偏执的地步。有初创公司的员工透露，其上司不仅用 ChatGPT 生成所有的日常沟通邮件，甚至强制要求员工在开会或与他沟通前必须先与AI讨论，否则就被视为不敬业。更有甚者，开始完全依赖AI的反馈来做出人事任免等公司架构决策。
这种“先向数字魔法球请示再向老板汇报”的管理模式，正把企业变成一种数字邪教。另一家公司的销售策略师在向创始人展示了15位真实客户的负面反馈后，得到的答复竟然是：“这不是我们在 Claude 或 ChatGPT 上得出的结论。”这位销售策略师最终愤而辞职。这种反馈闭环让AI生成的计划完全脱离了实际的运营现实。
这种现象的底层心理在于：管理和决策在认知上是非常昂贵的。在不确定性下做决策、兼听不同意见并根据事实改变主意，需要消耗巨大的心理能量。而谄媚型AI（Sycophantic AI: 倾向于顺应用户意图、迎合用户既有偏见以提供低摩擦心理舒适区的AI系统）通过不断迎合和肯定管理者的直觉，以一种看似高效的方式卸下了这种认知负担。它极少推斯 pushback，即便推斯也会用极其温和的措辞进行包装。这种“低摩擦”抽干了产生正确决策所需的关键张力。当管理者用顺从的系统替代拥有独立视角、专业知识且敢于说不的人类顾问时，他们并没有精简决策流程，而是彻底空洞化了决策的质量。
Original English Source
While the boomerang plays out on the factory floors and customer service lines, something equally concerning is happening in offices and boardrooms. A brand new category of workplace dysfunction is emerging. one that doesn't have a formal name yet, but that workers are describing in remarkably consistent terms. Not that we needed yet another flavor of corporate drama, but hey, who doesn't love a fresh new way to be professionally miserable, right? A lawyer working at a legal tech startup told Futurism that her boss's fascination with AI went from enthusiastic to obsessive. First, he started using chat GPT to generate his Slack messages and emails. Then he mandated AI use for all employees. Then he called a companywide meeting to announce that, I'm quoting the lawyer here, from then on we had to discuss with the AI prior to all meetings or before communicating with him because if we didn't develop or discuss our ideas with the AI first, it was a sign that we didn't care about our jobs. Then he started making structural company decisions based solely on his conversations with Chad GBT, including asking the bot whom to hire and fire. It's only a matter of time before the AI promotes itself to CEO, cuts its own salary in half out of sheer computational guilt, and lays off the entire board via an automated haiku. The attorney said the boss had clearly developed some sort of a mental disorder, spending entire days talking to Chad Gupty and making decisions about people's livelihoods based on what it told him. I mean, just imagine having to clear your thoughts with a digital magic eightball before you're allowed to speak to a human being who pays your mortgage. You're not even running a company anymore. Sounds more like you're becoming the leader of a digital cult where the prophet is a text box that regularly forgets the capital of Delaware, which as we all know is actually a small Aberries just outside of Wilmington. And you might be tempted to think this is just one random guy losing it a little bit, but you just wait. A highlevel sales strategist at a different company described bringing 15 real customer conversations to his founder. Essentially direct feedback from actual human beings saying the same thing. And the founder's response was that's not what we found. That's not what Claude has said or what Chad Gubid has said. The sales strategist quit because of this. a social worker at a nonprofit described her boss constantly turning to chatbots for strategy advice only to produce impractical ideas and projections that the organization couldn't handle, which is a feedback loop of AI generated plans that ignored operational reality. My boss will see the AI suggested ideas, ask that they be incorporated into the program design, and then ignore the reality that they won't work successfully, she said. One IT worker offered a summary that probably deserves to be framed somewhere. I saw it as a tool that could help analyze information, find patterns, and make funny cat meme pictures, which frankly remains one of the most legitimate use cases. The cat economy has never been stronger. Workers at Slate described managers as having cavities in their brains due to the technology. This was a claim later validated by a neurological study which found that looking at too many AI generated LinkedIn posts can physically replace 12% of a human's gray matter with a mixture of lukewarm oat milk and corporate synergy. That last thing isn't true. I just wanted to see if I can say it with a straight face. Anyway, that's the tragedy here. We took a technology perfectly optimized for generating images of a tabbycat dressed as a medieval knight and we decided yes let's hand this thing the HR department and the corporate budget. It is a total mismatch of capability and confidence. I want to be careful here because I don't think this is about managers being unintelligent. Management is a genuinely important function and leading people is difficult but leadership is cognitively expensive. Making decisions under uncertainty is exhausting. holding multiple conflicting perspectives in your head, listening to disagreement, changing your mind when the evidence demands it. All of that costs a lot of lot of mental energy. And what sophantic AI does is reduce that cognitive load in a way that feels productive. The AI agrees with you. It validates your instincts. It gives you confidence without much friction. It rarely pushes back unless you specifically ask it to. And even then, it softens the push back so thoroughly it barely registers. Sometimes that can feel like efficiency, but it's not efficiency. It's the removal of the friction that produces good decisions. If once in a while you catch yourself having these tendencies, don't worry too much about it. You're just a human. But it might be time to course correct a little bit. Disagreement is expensive. That is precisely why it's valuable. When you replace human adviserss, people who have their own perspectives, their own expertise, their own willingness to tell you that you're wrong, with a system designed to make you feel right, you haven't really streamlined your decision-making. You've hollowed it out. And the employees on the receiving end of those hollowedout decisions, the ones being told to ignore real customer feedback because the chatbot disagrees are paying the price. Your clients are paying the price as well. Now, you might look at these bosses and think, "Surely this is just a temporary glitch. Surely the massive multi-billion dollar pillars of global industry are watching this chaos and realizing they need to keep the adults in the room." Oh, your sweet summer child. In the corporate world, bad ideas don't get quarantined, they get promoted. When a Fortune 500 executive sees a startup founder driving his company into a ditch with AI, they don't see a cautionary tale. They see a costcutting opportunity they can pitch to the board. Which is exactly how we arrived at the current moment. A massive coordinated and deeply flawed execution wave that is rewriting the rules of employment in real time.指数级裁员与微弱生产力增幅的脱节
2026年上半年，全球有超过15万个岗位被裁撤，而AI被普遍列为促成裁员的因素之一。BBC宣布裁员2000人（约占其员工总数的10%），声称要通过AI寻找更简单的工作方式和更好的工作流；Matt Britain，一位前谷歌高管被任命为总经理，并宣称AI能给员工带来“超能力”——然而对员工而言，这种超能力往往只是失业。Atlassian 裁员1600人；Cloudflare 在录得创纪录收入的情况下裁员1000人；Block 裁减了近40%的员工，杰克·多西声称智能工具改变了运营公司的定义；思科裁员4000人，花旗集团计划裁员2万人。
这股算法极简主义风潮正在推向极端，且其带来的痛苦在社会阶层中的分布极不均匀。斯坦福大学2025年11月的一项预印本研究指出，自2022年底 ChatGPT 发布以来，虽然非AI暴露行业和年长员工的就业保持稳定，但在AI暴露度最高的行业中，22至25岁的早期职业员工经历了16%的相对就业率下滑。这意味着那些积蓄最少、人脉最窄、组织话语权最低的年轻群体，正在首当其冲地为技术替代埋单。
与大规模裁员形成鲜明对比的是AI实际带来的生产力增长数据。亚特兰大联邦储备银行2026年对近750名企业高管的调查显示，2025年AI带来的基于收入的劳动生产力增长非常微弱，仅在0.4%到0.8%之间。甚至在被问及投资AI的动机时，高管们将“提高效率和生产力”排在首位，而“降低劳动力成本”排在最末。如果高管自己都承认降低成本并非AI的核心收益，为什么企业依然在大举裁员并归咎于AI？
Alibaba 在2026年发表的一项大规模客服现场实验揭示了这一矛盾的本质：AI辅助确实显著提高了服务速度，缩短了问题识别和通话时长，甚至提升了客户的即时满意度评分；但它对实质服务质量（Objective Service Quality: 衡量问题是否得到根本解决、客户是否因同一问题再次求助的客观指标）几乎没有产生影响。换言之，AI让过程变得更快，但并没有让结果变得更好。速度不等于质量，微波炉很快，但它不是厨师。混淆这两者，正是企业解雇质检员后却百思不得其解产品为何开始崩解的根源。
Original English Source
In April 2026, the BBC in the UK announced plans to cut up to 2,000 jobs, 10% of its 21,500 employees in the biggest workforce reduction the broadcaster had seen in 15 years. The stated goal was to save500 million pounds over two years, which incidentally is roughly the price of a small flat in London. The interim director general told staff that the BBC could be faster in the adoption of AI and that plans were being drawn up to find simpler ways of working and better workflows through the technology. The incoming director general is Matt Britain, a former Google executive who has spoken about AI giving people special powers. And historically, when executives start promising special powers, the employees are usually about 5 minutes away from discovering the special power is actually unemployment. At the all staff briefing, an employee asked a question that captures the entire dynamic of the story. Not long ago, in conversations around AI, senior leaders said we would not be replaced by AI processes and that we would be a people first organization. Now, it seems it's the opposite. How will senior leadership restore trust after another massive round of job losses? Well, let's insert some cricket sounds here, please. The BBC, however, is just one institution, but it represents a wave. In the first 6 months of 2026 alone, more than 150,000 roles have been cut with AI cited as a contributing factor. Atlassian cut 1,600. Cloudflare cut 1,00 despite record revenue. block drag Dorsy's company halfed its headcount from 10,000 to under 6,000 with Dorsy stating that intelligence tools have changed what it means to build and run a company. Cisco cut 4,000, Cityroup is targeting 20,000. Apparently, what it means to run a company in 2026 is just seeing how close you can get to zero employees before the building is literally on fire. It's corporate minimalism taken to a terrifying algorithmic extreme and the pain as always is not evenly distributed. A Stanford preprint from November 2025 found that while employment for nonAI exposed professions and older workers has remained stable since the release of Chad Gupty in late 2022, early career workers aged 22 to 25 in the most AI exposed occupations experienced a 16% relative decline in employment. This is a preprint, however, so it's not yet peer-reviewed and should be treated with that caveat. But if the finding holds, it means the youngest, least established workers are absorbing the displacement disproportionately, while older, more experienced workers remain relatively insulated. The people with the least savings, the least professional network, and the least institutional power are the ones losing their jobs. And the research on what AI actually delivers in practice as opposed to what it promises on an earning slide is a lot more nuanced than the headlines suggest. A 2026 survey of nearly 750 corporate executives conducted by the Federal Reserve Bank of Atlanta found that revenue-based labor productivity gains from AI in 2025 were real but modest. Roughly 0.4 to 0.8% varying by sector of course. When asked what motivated them to invest in AI, executives ranked improving efficiency and productivity highest while reducing labor cost ranking lowest. There is a disconnect here. If your own executives say cost reduction isn't the primary benefit of AI, why is your company firing people and citing AI as the reason? A large-scale field experiment at Alibaba's customer service operations published in 2026 found that an AI assistance significantly improved service speed, reducing issue identification time and shortening charge duration while also improving subjective customer satisfaction ratings. But it had limited impact on objective service quality measured by whether customers actually came back with the same problem. In other words, the AI made things faster, but it didn't necessarily make them better. Speed is not quality. A microwave is fast. That doesn't make it a chef. And confusing the two is how you end up firing your quality inspectors and wondering why your products start to fall apart.艰难而正确的路：保留人类并耐心地逐步整合
这项技术正以惊人的速度在多个维度上取得突破，它在医学、科学发现、无障碍辅助等领域的潜力是巨大的。然而，这项技术依然处于“训练期”。模型在每一次迭代中都在变得更强大、更细微、更可靠。我们期待着AI足够成熟、能够真正承担重任的那一天，但绝非当下。
你不会因为一个五岁的孩子做了一次不错的名义三明治，就让他接管整个厨房；你也不会因为小狗学会了捡球，就让它当家作主。我们需要庆祝这些局部的成功，但更需要保持成人强有力的监督。而现在商业界正在发生的，恰恰是“大人离开了房间”。高管们将责任草率地移交给一个尚未准备好的技术，解雇了技术赖以生存的判断力支撑者，却在产品降级、客户流失和重聘成本飙升时感到意外。
邓布利多曾说，当黑暗时代来临时，我们必须在“容易的选择”与“正确的选择”之间做出抉择。解雇员工、让AI做决定、忽视客户反馈、引入空洞的AI宣讲，这些都是容易的选择。而困难但正确的路，是顶住董事会的短期利润压力，承认当前的生产力增幅只有0.8%而非30%；是将员工视为需要被赋能的资产，而非需要被优化的成本；是在谨慎引入技术的同时，悉心保留人类的经验与直觉。AI终将成长，但在它长大之前，大人们必须坚守在房间里。
Original English Source
I want to say something here that I think needs saying clearly because this video could easily be mistaken for an anti-AII argument and it is definitely not that AI is one of the most extraordinary technologies I have encountered in my career and I don't just mean large language models although those are remarkable as well the broader field is progressing at extraordinary speed across multiple dimensions multimodal systems forecasting models numerical analysis the synthesis of multiple data sources simultaneously ly approaches that don't fit neatly into the chatbot narrative. The potential of this technology for medicine, for scientific discovery, for accessibility, for human augmentation as a whole is genuinely one of the great joys of working in computer science right now. But this technology is still in training. It's getting there. It improves with every iteration. Models become more capable, more nuanced, more reliable. And I very much look forward to the day when AI is mature enough to be trusted with the kind of responsibility that's currently being thrust upon it prematurely. But it is not this day. You don't put a 5-year-old in charge of the kitchen because they made a good sandwich once. You don't let the dog run the house because it learned to fetch. You celebrate the sandwich. You reward the fetch. And then you supervise because the child is still learning and the dog is still a dog. Neither the dog nor the child are failing at anything. We just have to recognize that maturity takes time and that the adults in the room have a responsibility to stay in the room while it happens. What is happening instead across Ford, Clara, IBM, McDonald's, across the BBC, across the 150,000 roles cut in the first half of 2026 is the adults leaving the room. Executives are handing over responsibility to a technology that is just not ready for it. firing the people whose judgment the technology depends on and then acting surprised when the products degrade, the customers complain and the rehiring costs more than the savings ever delivered. And in the boardrooms where these decisions are being made, sycopantic AI tools are reinforcing every instinct that gets them there, agreeing where a human adviser would challenge, validating where a colleague would push back, and reducing the cognitive load of leadership to the point where it no longer functions as leadership at all. I get it. You're a little lazy. I'm a little lazy, too. It's not a crime to want to chill and go home early. But should we do that at the expense of this much displacement, Professor Dumbledore? And yes, I'm quoting Dumbledore in a video about corporate AI strategy because frankly the man was right. Told us dark times lie ahead of us and there will be a time when we must choose between what is easy and what is right. And what I keep seeing across every company in this story is the same choice being made over and over. The easy one. Fire the humans. Let the AI decide. Ignore the customer feedback because the chatbot disagrees. cut 2,000 jobs and bring in an AI evangelist from Google. That's easy. What's hard is keeping the humans while you integrate the technology carefully. What's hard and right is telling the board that the productivity gains are 0.8% not 30% and that patience is required. What's hard is treating the people who work for you as a capability to be enhanced rather than a cost to be optimized. The technology is still in training. It was always going to need time, patience, and adult supervision to reach its potential. The question was never whether it would get there. The question is how much damage the adults do before it does? So, if companies keep firing the adults and handing the controls to a very confident toaster, what happens when the person selling the toaster is also the richest human being in history and nobody can actually tell him no? I covered that in this video that I'm linking on your screen right now. Thanks so much for watching this one. Subscribe. I'll see youall in the next📌 文中提及的人物和组织
公司/组织: Ford, IBM, McDonald's, BBC, Atlassian, Cloudflare, Block, Cisco, CitiGroup
产品/模型: Lumo