Open Questions

Will A.I. Still Take Our Jobs?

人工智能还会抢走我们的工作吗?
The jobs apocalypse hasn’t arrived. Something else is happening instead.
工作末日并未到来。正在发生的,是另一回事。
Illustration of a tag and suit
Illustration by Josie Norton

Suppose that your company wants to launch a new product. A substantial budget is available to fund a prototype. You think you have a good idea; so do Frank and Jill, your workplace frenemies. Your boss asks the three of you to confer with your respective teams and draft proposals, which will be circulated to those in charge.

假设你的公司想推出一款新产品,有一笔可观的预算可用于制作原型。你自认为有个好点子;你的职场对手弗兰克和吉尔也是。老板让你们三人各自与团队商议后起草方案,提交给决策层审阅。

In the past, you and your subordinates might have embarked on a lengthy research effort. You would’ve calculated the size of your potential market, surveyed competitors, evaluated barriers to entry, estimated manufacturing costs, and worked up branding ideas. This would have taken weeks—and it would only have been after gathering these materials that you’d have sat down to write your draft. But your team happens to include Claudette, a recent college graduate who is great at using artificial intelligence. She organizes a few brainstorming sessions, records them, then runs the transcripts through a large language model, which creates a research plan. She tweaks it, then delegates various parts of it to A.I. agents, assigning yet more agents as fact checkers. She takes all their work and synthesizes it. A first draft of the proposal, complete with visuals, is done in a week; after everyone gives Claudette feedback, the revised draft is done a few days later. As far as you can tell, the proposal is fantastic. You click “Send.”

过去,你和下属可能得展开漫长的调研:估算潜在市场规模、调查竞争对手、评估进入壁垒、测算制造成本、构思品牌方案。这要花好几周——而且只有收集完这些材料之后,你才会坐下来动笔。但你的团队里恰好有克劳德特(Claudette),一个刚毕业的大学生,极擅长使用人工智能。她组织了几场头脑风暴并录了音,然后把文字记录输入大语言模型,生成了一份研究计划。她稍作调整,把各部分任务分派给AI智能体,又安排了更多智能体做事实核查。她汇总所有成果,整合成文。方案初稿连同视觉设计一周内完成;大家给克劳德特反馈后,修改稿几天后出炉。就你看来,这份方案堪称出色。你点了“发送”。

Soon afterward, you get a look at Frank’s and Jill’s proposals. Apparently, their teams have recent college grads, too. Their proposals, like yours, are shockingly well researched, with lots of data alongside polished writing and images. It seems that the efficiency and quality of everyone’s output has increased. This creates two sets of problems: one for you and one for your bosses. It used to be that the people in charge could evaluate proposals not just for their content but as indirect measures of intelligence, creativity, and ambition; they believed that only the best, most committed workers could produce high-quality results. Now that all proposals are good, though, the underlying realities are harder to discern. You, meanwhile, struggle to prove your own value compared with your competitors’. Also, you used to believe in the work you presented: honing it over long periods, you became sure of its worth. By helping you rush to the finish line, A.I. has deprived you of the time you need to know that you’re right. What if your idea isn’t so good, after all?

不久之后,你看到了弗兰克和吉尔的方案。显然,他们的团队里也有刚毕业的大学生。他们的方案和你的如出一辙:研究扎实得惊人,数据丰富,文字精炼,配图考究。似乎每个人的产出效率和质量都提高了。这带来了两组问题:一组是你的,一组是老板们的。过去,决策者评估方案时,不仅看内容,还把它当作衡量智力、创造力和进取心的间接指标;他们相信只有最优秀、最投入的员工才能产出高质量的成果。如今所有方案都很出色,背后的真实水平反而难以分辨。与此同时,你苦于无法证明自己相比竞争对手的价值。而且,你过去对自己提交的工作是有信念的:经过长时间打磨,你确信它的价值。AI帮你冲刺到终点,却剥夺了你确认自己判断正确所需的时间。万一你的想法其实没那么好呢?

“If every team can generate better AI-based analysis to support their arguments, then the demand for conflict resolution and authority-based decisions will increase dramatically,” the economists Luis Garicano, Jin Li, and Yanhui Wu predict, in “Messy Jobs: The Work That AI Cannot Reach.” They note that many workplace decisions aren’t made only on the merits; they also involve deciding “who gets their way.” Who’s ready to take a big swing, or too inexperienced for heavy responsibilities? What kinds of ideas always sound good but never work? What does the C.E.O. really think, but never say? Such information isn’t explicit, but tacit—it’s known, but not written down—and so it isn’t available to an A.I. system. Moreover, the proliferation of A.I.-generated work can make it harder for decision-makers to collect the tacit information they need. If every cover letter is nicely written, and every memo thorough and well structured, how can a boss know whom to trust? If everyone uses A.I. to generate ideas, how do you know who’s actually creative?

经济学家路易斯·加里卡诺(Luis Garicano)、李进(Jin Li)和吴彦辉(Yanhui Wu)在《混乱的工作:AI无法触及的领域》(Messy Jobs: The Work That AI Cannot Reach)中预测:“如果每个团队都能生成更好的AI分析来支撑自己的论点,那么对冲突解决和权威决策的需求将大幅增加。”他们指出,许多职场决策并非只凭优劣而定;它们还涉及决定“谁说了算”。谁准备好放手一搏?谁又太缺乏经验、担不起重任?什么样的点子总是听起来很好却从不奏效?CEO真正想的是什么,却从不说出口?这类信息不是显性的,而是隐性的——大家都知道,但没人写下来——因此AI系统无法获取。此外,AI生成内容的泛滥会让决策者更难收集所需的隐性信息。如果每封求职信都写得漂亮,每份备忘录都详尽周全、结构清晰,老板怎么知道该信任谁?如果人人都用AI生成创意,你怎么知道谁真正有创造力?

ChatGPT first appeared in 2022; Claude, in 2023. Almost immediately, an imminent jobs apocalypse was predicted. There’s no question that people find A.I. useful: studies and surveys show that an increasing number of office workers are now employing it on a daily basis. Certain fields—coding, recruiting, scientific research, the law—really do seem to be getting transformed. And yet A.I.’s effect, in general, is turning out to be hard to measure. Many workers appear to be using it semi-secretly, on their own devices, perhaps saving themselves time or improving their work in ways that aren’t reflected on the bottom line. Recent college grads are finding it harder to get hired, and customer-service jobs may be disappearing, but job openings for software engineers, which decreased substantially in 2025, increased in 2026. Does this mean that A.I. is creating software jobs? Or is the industry merely rebounding after post-pandemic downsizing? Nobody knows.

ChatGPT于2022年问世;Claude于2023年登场。几乎立刻,人们就预言了一场迫在眉睫的工作末日。毫无疑问,人们觉得AI有用:研究和调查显示,越来越多的办公室职员如今每天都在使用它。某些领域——编程、招聘、科研、法律——确实正在被改变。然而总体而言,AI的影响却难以衡量。许多员工似乎在半秘密地使用它,用自己的设备,可能为自己节省了时间或提升了工作质量,但这些并未体现在公司账面上。应届大学毕业生发现找工作更难了,客服岗位可能在消失,但软件工程师的职位空缺在2025年大幅减少后,2026年又有所回升。这意味着AI在创造软件岗位吗?还是说这个行业只是在疫情后的裁员潮中反弹?没人知道。

“Early evidence is hardly the last word on the future of work in an AI world,” a group of Stanford researchers cautioned, in July. Part of the difficulty is that, with A.I. in the mix, we’re realizing that we don’t necessarily know how work works. Why are the jobs we have set up the way they are, and how much could they change? What is distinctly human in what we do, and what is amenable to automation? What makes working with someone valuable, beyond the work they produce? As more people use A.I., the blunt idea of an A.I.-driven jobs apocalypse is getting replaced with a growing number of challenging questions, with which managers and workers are just beginning to grapple.

斯坦福大学的一组研究人员在7月提醒道:“早期证据远非AI时代工作未来的最终结论。”困难之一在于,当AI介入后,我们意识到自己未必了解工作是如何运作的。为什么我们现有的岗位是现在这个样子?它们能有多大程度的改变?我们的工作中什么是独属于人的,什么是可以自动化的?与某人共事的价值,除了他们产出的成果之外,还有什么?随着越来越多的人使用AI,“AI驱动的工作末日”这个粗线条的说法正在被越来越多棘手的问题所取代,管理者和员工才刚刚开始面对这些问题。

Economists have a term—the production function—for describing how things are made. Imagine you’re having a dinner party for ten. If you decide to make steak frites, then you’ll have to cook the steaks and the frites in the minutes just before your guests sit down to eat. If you only have four burners on your stove, then you’ll need to sear the steaks in batches; if an extra guest arrives, you must cook an extra steak. Alternatively, you could make a giant pot of stew. In that case, you could do almost all the work a day or two beforehand, then put the pot on the stove when your guests arrive. If an extra guest presents himself, there’s probably enough to go around. Steak frites and stew have completely different production functions. If you graphed them, with effort on one axis and results on the other, you’d get totally different curves.

经济学家有一个术语——生产函数——用来描述事物是如何被制造出来的。想象你要为十个人办一场晚宴。如果你决定做牛排薯条,那么你必须在客人入座前的几分钟内煎好牛排、炸好薯条。如果你的炉灶只有四个灶头,你就得分批煎牛排;如果多来一位客人,你就得多煎一块。或者,你可以做一大锅炖菜。那样的话,你几乎可以提前一两天完成所有工作,客人到了之后把锅放上炉子就行。如果临时多来一位客人,大概率也够吃。牛排薯条和炖菜有着完全不同的生产函数。如果你把它们画成图,横轴是投入,纵轴是产出,你会得到截然不同的曲线。

“Messy Jobs” deals, among other subjects, with the precise ways in which A.I. changes production functions at work. A.I., the authors argue, creates a “new shape of progress” for what we do—and the shape isn’t simply up and to the right. They describe a study in which artists were given A.I. tools that helped them quickly deliver a finished product—an illustration of a scene from a novel. The artists reached, in half an hour, “a quality level that would have taken two hours by hand”—and yet, at that point, progress slowed. Because the artists had used A.I. to “get a polished image before they had thought enough about the composition,” they struggled to improve it; “further gains were barely noticeable, even as artists kept tweaking prompts and patching details.” Ultimately, the artists split into two groups: those who simply suspended their work after about an hour, and those who kept working fruitlessly.

《混乱的工作》讨论的主题之一,就是AI如何精确地改变工作中的生产函数。作者认为,AI为我们的工作创造了一种“新的进步形态”——而这个形态并非简单的向右上方攀升。他们描述了一项研究:给艺术家提供AI工具,帮助他们快速完成成品——一幅小说场景插画。艺术家们在半小时内就达到了“手工需要两小时才能达到的质量水平”——然而到了那个节点,进展就放缓了。因为艺术家们借助AI“在充分思考构图之前就得到了精修图像”,他们反而难以进一步改进;“即使不断调整提示词、修补细节,进一步的提升也几乎察觉不到。”最终,艺术家们分成两组:一组在大约一小时后停止了工作,另一组则继续徒劳地修改。

Garicano, Li, and Wu call this the “90/10 production function.” A.I. makes it easy to build something that’s ninety-per-cent great. But if you want to get to a hundred per cent—that is, if you want to achieve true excellence—it will often leave you stranded. The challenge, therefore, is to distinguish between “commodity tasks,” for which good enough is actually good enough, and “star tasks,” for which excellence is required. (In some fields, of course, excellence is required nearly all the time.) Workers have to make this distinction in advance, because these two kinds of tasks might demand entirely different approaches: you can’t decide, halfway through cooking your stew, to turn it into steak frites. And bosses have to figure out how to identify and reward excellence in star tasks. “How do you distinguish workers who pushed to 98 percent from those who stopped at 91 percent?” the authors ask. If you fail to look closely, “the surface looks identical.”

加里卡诺、李和吴称之为“90/10生产函数”。AI让你很容易做出九成好的东西。但如果你想达到十成——也就是说,如果你想追求真正的卓越——它往往会让你陷入困境。因此,挑战在于区分“商品型任务”和“明星型任务”:前者够好就真的够了,后者则需要卓越。(当然,在某些领域,几乎所有时候都需要卓越。)工作者必须提前做出这种区分,因为这两类任务可能需要完全不同的方法:你不可能炖菜炖到一半,决定把它改成牛排薯条。而老板们则必须想办法识别并奖励明星型任务中的卓越表现。“你如何区分那些推进到98%的员工和停在91%的员工?”作者问道。如果你不仔细看,“表面看起来一模一样。”

Companies don’t only make products. They also foster workers—ideally, experienced ones, who have grown in knowledge, wisdom, and trustworthiness over time. The 90/10 production function has profound implications for the experienced-worker production function. How will less-experienced workers gain skills when the creation of good-enough work is automated? Junior associates at law firms learn, in part, by producing the simpler documents that A.I. can now create; trainee physicians learn to diagnose difficult cases by figuring out easier ones. In doing such tasks, younger employees acquire work histories and networks of relationships that allow their superiors to see how trustworthy they can be. In order to occupy a position of authority—to decide “who gets their way”—you have to understand the big picture; you have to know that Geoff in accounting is always slow, that Linda in the marketing department is always worth consulting, and that the worst time to pitch an idea to the C.F.O. is on Thursday mornings, before she meets with her boss. You have to owe people favors, so that they owe you favors. How will you gain all this if all you do is prompt?

公司不仅制造产品,还培养员工——理想情况下,是那些随着时间推移在知识、智慧和可信度上不断成长的经验丰富的员工。90/10生产函数对经验型员工的生产函数有着深远的影响。当“够好”的工作成果可以被自动化生产时,经验较少的员工如何获得技能?律所的初级律师部分是通过撰写那些AI现在就能生成的简单文件来学习的;实习医生是通过先诊断简单病例来学会处理疑难病例的。在做这些任务的过程中,年轻员工积累了工作履历和人际关系网络,让上级得以判断他们有多值得信赖。要占据一个有权威的位置——决定“谁说了算”——你必须了解全局;你必须知道会计部的杰夫总是慢半拍,市场部的琳达总是值得请教,而向CFO推销想法最糟糕的时机是周四上午,在她去见老板之前。你得欠别人人情,这样别人才会欠你人情。如果你做的只是输入提示词,你怎么获得这一切?

If work is this messy, then does A.I. even have a place in it? The authors of “Messy Jobs” argue that the technology actually has substantial value for many companies and workers. (It’s because of this value, presumably, that people are adopting the tools on their own, outside of their bosses’ supervision.) In their view, there are some jobs that will certainly be out-and-out eliminated by A.I.—jobs centered on discrete, predictable tasks that tend to be accomplished by individuals in isolation. And yet, they argue, lots of jobs are actually bundles of tasks, only some of which can be automated. Whether such jobs will be reshaped by A.I. depends on whether their tasks form a “weak bundle,” which can be broken up without penalty, or a “strong bundle,” in which “separation destroys value.”

如果工作如此“混乱”,那么AI还有立足之地吗?《混乱的工作》的作者认为,这项技术对许多公司和员工来说实际上具有相当大的价值。(大概正是因为这种价值,人们才在老板的监管之外自行采用这些工具。)在他们看来,有些工作肯定会被AI彻底淘汰——那些以离散、可预测的任务为核心、往往由个人独立完成的工作。然而,他们指出,许多工作实际上是任务的组合体,其中只有一部分可以自动化。这类工作是否会被AI重塑,取决于它们的任务构成的是“弱捆绑”——可以无代价地拆分开来——还是“强捆绑”,即“分离会破坏价值”。

Lots of jobs are strong bundles. When we go to the doctor, we want the person who prescribes our medicine to be the same one who diagnoses us; we also want that person to be legally liable if things go wrong. When we enter into a contract with a vender, we want the person we negotiate with today to be the same one we’ll negotiate with two years from now, and we want them to understand what they’re selling and to track its implementation. Both the physician and the salesperson have strong bundles. Are you called upon to do different kinds of tasks at different times? Are you better at some parts of your job because you do other parts? If your answer to such questions is yes, then your bundle is probably strong.

许多工作都是强捆绑。我们去看医生时,希望开药的人和诊断的人是同一个;我们还希望如果出了问题,这个人要承担法律责任。我们与供应商签订合同时,希望今天谈判的人就是两年后还要谈判的那个人,而且希望他了解自己在卖什么,并跟踪合同的执行情况。医生和销售员都拥有强捆绑。你是否需要在不同时间处理不同类型的任务?你是否因为做了工作中的某些部分而更擅长其他部分?如果你对这些问题的回答是肯定的,那么你的捆绑很可能是强的。

Garicano, Li, and Wu think that people with strong bundles have a lot to gain from A.I. In one scenario, which they call “robots above,” workers get access to expertise they’d otherwise lack; they need to escalate to their bosses less frequently, and can accomplish more on their own initiative. In another, which they call “robots below,” A.I. “takes over the routine problems of the job while the human handles the exceptions.” In both cases, the value of a worker—especially an experienced one—goes up. In some industries, workers who are “upskilled” by A.I. will still be in trouble: if all workers can do more, but demand doesn’t increase, then fewer workers will be necessary. But in many industries demand can grow. There is an insatiable need for cheaper, better health care; if A.I. helps clinicians do more, there will be more for them to do.

加里卡诺、李和吴认为,拥有强捆绑的人能从AI中获益良多。在一种他们称之为“机器人在上”的情景中,员工获得了原本缺乏的专业知识;他们需要向上级请示的频率降低了,能够更多地自主完成工作。在另一种他们称之为“机器人在下”的情景中,AI“接管工作中的常规问题,而人类处理例外情况。”在这两种情况下,员工——尤其是经验丰富的员工——的价值都上升了。在某些行业,被AI“提升技能”的员工仍会面临困境:如果所有员工都能做更多事,但需求没有增长,那么需要的员工就会减少。但在许多行业,需求是可以增长的。人们对更便宜、更好的医疗有着永不满足的需求;如果AI帮助临床医生做更多的事,他们就会有更多的事可做。

Entrepreneurs are perfectly positioned to take advantage of A.I.; there’s no org chart to hem them in. But employees inside big companies need managers to notice what’s possible, so that workflows can change. During this transitional period, the authors argue, the scarcest workers are the people who can help organizations adapt to A.I. The people best positioned to help aren’t chief information officers or consultants, they suggest, but rank-and-file workers who are intimately familiar with how work is actually done. They cite, as a positive example, the Spanish bank BBVA—a large company with eighty million customers and more than a hundred and twenty thousand employees. Instead of equipping all of them with A.I., BBVA’s leaders offered it only to the most enthusiastic workers—people across the company who were excited about the technology. They then rewarded those who came up with good ways to use it. The transformation was led from within, instead of from the top down, by individuals who were simultaneously technically literate, familiar with the details of day-to-day work, and cognizant of the workplace politics that resist change. There aren’t many people with those three competencies—and their scarcity, the authors argue, is what is keeping A.I. from transforming work at the speed some predicted. Smart bosses will start hunting for candidates.

创业者处于利用AI的绝佳位置;没有组织架构图束缚他们。但大公司内部的员工需要管理者注意到可能性所在,工作流程才能改变。作者认为,在这个过渡时期,最稀缺的员工是那些能帮助组织适应AI的人。最适合提供帮助的人不是首席信息官或咨询顾问,而是那些对实际工作方式了如指掌的一线员工。他们举了一个正面例子:西班牙银行BBVA——一家拥有八千万客户和超过十二万名员工的大公司。BBVA的领导层没有给所有员工配备AI,而是只把它提供给最热心的员工——公司里对这项技术充满热情的人。然后,他们奖励那些想出好用法的人。这场转型是由内部推动的,而非自上而下,推动者是那些同时具备技术素养、熟悉日常工作细节、并且了解抵制变革的职场政治的人。同时具备这三种能力的人并不多——作者认为,正是这种稀缺性,使得AI没有以一些人预测的速度改变工作。聪明的老板会开始寻找这样的人选。

Is “Messy Jobs” right? In many respects, it’s a reassuring book. According to its analysis, A.I. isn’t a bubble but a complex and valuable new technology that takes time to figure out; instead of causing a generalized jobs apocalypse, it will hurt some workers while helping others to do more, both for themselves and for their companies. Two alternative views loom on either side, both less encouraging. It could be that bosses will simply be bad at implementing A.I.—that, instead of comprehending its potential to increase the value of workers, they will see it mainly as a cost-cutting tool, hollowing out their companies as a result. And it could also be that the progress of artificial intelligence will accelerate further, making some kinds of work so cheap and easy that many strong bundles become weak. No one can predict the future; only time will tell. Our best bet, right now, is to make sure that we’re asking the right questions. ♦

《混乱的工作》说得对吗?在许多方面,这是一本令人安心的书。根据它的分析,AI不是泡沫,而是一项复杂而有价值的新技术,需要时间去摸索;它不会造成普遍的工作末日,而会伤害一些员工,同时帮助另一些员工为自己和公司做更多的事。两种相反的观点分列两侧,都不那么令人鼓舞。一种可能是,老板们就是不擅长实施AI——他们不会理解AI提升员工价值的潜力,而主要把它当作削减成本的工具,结果掏空了自己的公司。另一种可能是,人工智能的进步会进一步加速,使某些类型的工作变得如此廉价和容易,以至于许多强捆绑变成弱捆绑。没人能预测未来;只有时间能给出答案。我们眼下最好的选择,是确保自己在问正确的问题。♦

2 free previews left — register to get 30 translations a month