深翼 DeepWing深翼 DeepWing
EN中文

深度调研 · 就业与职业 · 2026-10-24Deep Research · Jobs & Careers · Oct 24, 2026

三个数字，三回事：被引错的 AI 就业数字，到底在数什么Three Numbers, Three Different Things: What the Misquoted AI Jobs Figures Actually Measure

作者By 深翼 DeepWingDeepWing·发布Published 2026-10-24·最后核查Last checked 2026-10-24·阅读Read 约 10 分钟~14 min

AI 与就业AI & Workforce阅读时间：约 10 分钟Reading time: About 14 min

研究问题Research question
最近一年被反复揉成一句话的三个 AI 就业数字（就业降 13%、找工作难 14%、新人占比从 27% 降到 9%）各自到底在数什么？它们是怎么被引错的？Three AI jobs figures have been kneaded into one frightening sentence over the past year (employment down 13 percent, job hunting 14 percent harder, juniors cut from 27 to 9 percent of hires). What does each of them actually measure, and how exactly do they get misquoted?
研究方法Method
一手原文逐条对照：斯坦福工作论文、Anthropic 经济指数研究、Anthropic 研究所报告三份原文逐项核对口径、版本与限定语；8 条混引样本逐条打开原文核对标题与原句；另以三份无关的第三方数据（哈佛、SignalFire、耶鲁大学预算实验室）作边界对照（其中哈佛一篇只核到摘要，经斯坦福数字经济实验室研讨会页核对）；全篇按登记册纪律推进：三份一手研究已入就业台登记册（RES-001、RES-004、HIR-001）并按站上四档证据等级定级，事实与推断分开标注。Line by line checking against the primary texts: the Stanford working paper, the Anthropic Economic Index study, and the Anthropic Institute report, each checked for caliber, vintage and qualifiers. Eight conflation samples were opened and checked against their headlines and quoted wording. Three unrelated third party datasets (Harvard, SignalFire, and the Budget Lab at Yale) supply boundary cross checks (the Harvard paper was checked in abstract only, via the Stanford Digital Economy Lab seminar page). The whole piece follows register discipline: the three primary studies are held in the jobs desk registers (RES-001, RES-004, HIR-001) and graded on the site's four evidence grades, with fact and inference labeled apart.
证据等级Evidence grade
一手已核Primary verified
核查截止日期Verified through
2026-10-242026-10-24
核心发现Key findings
① 三个数数的是三种不同的量：22 至 25 岁高暴露职业就业人数的相对降幅（斯坦福，ADP 数据）、求职成功率（Anthropic，CPS 数据）、一家公司内部的初级招聘占比（Anthropic 内部人事数据）；把它们并排放在一起，是这场讨论最大的误会来源。② 13% 是控制企业层面冲击后的相对数，同一篇里的原始人数降幅只有约 6%，且同一估计随版本变为 16%（2025 年 11 月修订版）与 19% 差距口径（2026 年 8 月修订版）；Anthropic 转述的「6% 至 16%」是同一篇论文里的两个不同的数，不是误差范围（这一拆解是深翼解读，Anthropic 原文没有自己定义区间的端点）。③ 三份第三方对照方向一致（新人入口在收窄），但无一支持「就业人数普降 13%」或「初级招聘已经崩了」；一手原文能支持的只有：同期技术岗招聘绝对人数仍在增长，降的是新人占比。1. The three numbers measure three different things: a relative decline in employment for 22 to 25 year olds in the most AI exposed occupations (Stanford, ADP payroll data), a job finding success rate (Anthropic, CPS panel data), and the junior share of hiring inside one company (Anthropic's own personnel records). Placing them side by side as one finding is where most of the confusion starts. 2. The 13 percent is a relative figure after controlling for firm level shocks; the raw headcount fall in the same paper is about 6 percent, and the same estimate runs to 16 percent in the November 2025 revision and a 19 percent gap measure in the August 2026 revision. The "6 to 16 percent" range in Anthropic's retelling is two different numbers from the same paper, not an error band (this decomposition is DeepWing's own reading; Anthropic does not define the endpoints itself). 3. The three cross checks agree on direction, a narrowing entry level, but none supports a claim that employment fell 13 percent across the board or that junior hiring collapsed; what the primary text supports is only that technical hiring kept growing in absolute terms while the junior share fell.
研究局限Limitations
本文核查的是数字的口径与引用方式，不直接计量 AI 对青年就业的因果影响；三份第三方对照量的是不同的总体，测不到 22 至 25 岁高暴露职业这一小格子里的变化；Anthropic 报告通篇没有给出初级岗绝对人数，因此「初级岗人数涨了」与「初级岗人数崩了」两种说法，本文都无法支持。This piece checks what the numbers count and how they are quoted; it does not itself measure AI's causal effect on youth employment. The three cross checks measure different populations and cannot see inside the small cell of 22 to 25 year olds in the most exposed occupations. The Anthropic report nowhere gives the absolute number of junior hires, so this piece can support neither "junior hiring grew" nor "junior hiring collapsed".
后续观察Watch next
斯坦福论文后续修订版的数值与口径变化；Anthropic 经济指数与研究所报告的后续版本；耶鲁大学预算实验室的月度更新与 SignalFire 的下一年度报告；新的混引样本是否继续出现。Later revisions of the Stanford paper and their figures; follow up versions of the Anthropic Economic Index study and the Institute report; the Budget Lab's monthly updates and SignalFire's next annual report; whether new conflation samples keep appearing.

导语Introduction

最近一年，三个数字被反复揉成一句话：AI 让年轻人找不到工作了，证据是就业降了 13%，找工作难了 14%，连卖 AI 的公司都把新人占比从 27% 砍到 9%。Over the past year, three numbers have been kneaded into a single frightening sentence: AI is costing young people their jobs, and the proof is that employment fell 13 percent, job hunting got 14 percent harder, and even the company selling AI cut juniors from 27 percent of its hires to 9 percent.

我们把三份原文逐条对完，结论是：这三个数，数的根本不是一回事。它们来自三份不同的研究，用三套不同的数据，量的是三种不同的东西：一个是就业人数的相对降幅，一个是求职成功率，一个是一家公司内部的招聘占比。把它们并排放在一起吓人，是这场讨论里最大的误会源头。We checked all three against the original texts, line by line. The three numbers measure three different things, from three different studies, on three different datasets: a relative decline in employment levels, a job-finding success rate, and a hiring share inside one company. Placing them side by side as if they were one finding is where most of the confusion starts.

先从我们自己说起。深翼此前在周刊预告与就业台中也用过"少招新人"这一说法，已更正，见变更记录。下面这把标尺，先量我们自己，再量别人。A disclosure about our own copy first. DeepWing previously used the "hiring fewer juniors" formulation in a Weekly newsletter preview and on our jobs desk. That phrasing has been corrected; see the change log. The standard applied below applies to us first.

先看它们是怎么被引错的First, how they get mixed up

来源Source日期Date怎么混的What got mixed up
Lucerna 研究简报Lucerna research brief2026-03-092026-03-09写「达拉斯联储发现就业下降 13%」，又称「国际货币基金组织确认 13% 至 16%」。两家机构都不是这三个数的来源，13% 出自斯坦福用 ADP 数据的研究Credited a "13% employment decline" to the Dallas Fed and a "13 to 16%" range to the IMF. Neither institution produced any of these numbers; the 13% is Stanford's, on ADP data
OpenTools 新闻站OpenTools news site2026-03-112026-03-11把斯坦福两个版本年份的 13% 和 16% 拼成「13% 至 16% 的招聘率相对下降」，放进一篇讲 Anthropic 研究的报道里。斯坦福的数成了 Anthropic 的数，就业量成了招聘率Spliced Stanford's 13% and 16% (two vintages of one estimate) into a "13 to 16 percent relative decline in hiring rates," inside a story about Anthropic's research. Stanford's number became Anthropic's, and employment became hiring
《财富》中文网Fortune China2026-07-012026-07-01把斯坦福的结论写成「就业人数下降了 13%」。原文是控制企业层面冲击后的相对降幅，同一篇里的原始降幅只有约 6%Reported Stanford's result as "the number employed fell 13 percent." The original is a relative decline after controlling for firm-level shocks; the raw fall in the same paper is about 6%
一位 Substack 专栏作者A Substack columnist2026-07-292026-07-29把斯坦福的就业相对降幅说成「招聘挨了 16% 的打击」，再和 Anthropic 的求职成功率 14% 并列，两个数被当成同一种量Recast Stanford's relative employment decline as a "16% hit to hiring," then set it beside Anthropic's 14% job-finding figure as if both measured hiring
Adi Insights，MediumAdi Insights, Medium2026-10-182026-10-18标题写「初级招聘一年内从 27% 崩到 9%」。原文里 27% 到 9% 是占比，不是招聘人数Headline: "Junior hiring collapsed from 27% to 9% in a single year." In the source, 27% to 9% is a share, not a count of hires
GenZ NewZGenZ NewZ2026-10-192026-10-19标题写「初级技术招聘已经崩了」，正文却写对了，是占比从 27% 降到 9%。标题说的，正文和原文都不支持Headline says junior tech hiring "has collapsed," while its own body correctly reports a share falling from 27% to 9%. The headline asserts what neither the body nor the source supports
亿邦动力ebrun2026-10-192026-10-19标题写「Anthropic 近乎停招初级技术员工」。占比降到 9% 不等于近乎停招，原文还写了技术岗招聘绝对人数仍在增长Headline: "Anthropic has almost stopped hiring junior technical staff." A share falling to 9% is not a near halt, and the source states that technical hiring kept growing in absolute terms
MarsBit 转载 InfoQ 中国MarsBit, republishing InfoQ China2026-10-202026-10-20标题把两个占比搅在一起：27% 到 9% 是招聘占比，标题里的「骤降 33%」算的是另一个占比（员工构成 21% 到 14%）的相对变化Blended two different shares into one headline: 27% to 9% is the hiring share; the headline's "plunges 33%" is the relative change of a different share, staff composition (21% to 14%)

样本看完，再逐个拆数。With the samples on the table, here is each number on its own terms.

第一个数：下降 13%，是就业量的相对降幅，出自斯坦福Number one: down 13 percent, a relative decline in employment levels, from Stanford

原文是什么。 斯坦福数字经济实验室的工作论文《煤矿里的金丝雀？》（Canaries in the Coal Mine?），作者 Erik Brynjolfsson、Bharat Chandar、Ruyu Chen，2025 年 8 月 26 日版，用的是美国最大薪资服务商 ADP 的实际雇佣记录，样本为 2021 年 1 月至 2025 年 7 月持续在册的企业，涉及数百万名员工。What the source is. "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence," a Stanford Digital Economy Lab working paper by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, August 26, 2025 version. It uses payroll records from ADP, the largest payroll provider in the United States, for firms present throughout January 2021 to July 2025, covering millions of workers.

口径是什么。 13% 是 22 至 25 岁、在 AI 暴露程度最高的一组职业里就业人数的相对降幅，时间从 2022 年底到 2025 年 7 月。这里的「相对」有确切含义：它是在控制了企业层面冲击（同一家公司整体的冷暖）之后，跟暴露程度最低的一组职业比出来的数。原文摘要的原话是：这一组年轻人经历了 13% 的相对就业下降，「即使在控制企业层面冲击之后」。同一篇论文里，不做这层控制、直接看人数变化，降幅只有约 6%；同组里 35 至 49 岁员工的就业同期还涨了 9% 以上。所以 13% 不是「就业人数掉了 13%」，原始人数降幅约是它的一半。What the number measures. The 13% is a relative decline in employment for workers aged 22 to 25 in the most AI-exposed group of occupations, from late 2022 to July 2025. "Relative" has a precise meaning here: the estimate controls for firm-level shocks (how a given company was doing overall) and is measured against the least-exposed occupations. The abstract's own wording: this group "experienced a 13 percent relative decline in employment even after controlling for firm-level shocks." In the same paper, the unadjusted change in headcount is a fall of about 6%, while employment for workers aged 35 to 49 in the same occupations grew by more than 9%. So 13% is not "13 percent fewer people employed"; the raw headcount fall is roughly half of it.

还有一个常被省掉的细节：这个数有版本。2025 年 8 月版报 13%（数据到 2025 年 7 月）；2025 年 11 月修订版把同一口径的估计更新为 16%（数据到 2025 年 9 月）；2026 年 8 月的修订版又换了差距口径，报 19%。引用时只说一个光秃秃的百分比，不带版本日期，本身就是混引的开始。One more detail that usually gets dropped: the number has vintages. The August 2025 version reports 13% (data through July 2025). The November 13, 2025 revision updates the same estimate to 16% (data through September 2025). The August 2026 revision switches to a gap measure and reports 19%. Quoting a bare percentage with no version date is itself the start of a conflation.

最常见的错读。 一是安错出处：这个数常被写成 Anthropic 的发现，它跟 Anthropic 没有关系。二是丢掉「相对」和「控制之后」：写成就业人数普降 13%。三是把三个版本年份的数字当成三个互相印证的发现。How it is most often misread. First, wrong owner: it is routinely credited to Anthropic, with which it has nothing to do. Second, the words "relative" and "after controlling" get dropped, turning it into a universal 13% fall in employment. Third, the three vintages get cited as three separate findings that confirm each other.

第二个数：下降 14%，是求职成功率，出自 AnthropicNumber two: down 14 percent, a job-finding success rate, from Anthropic

原文是什么。 Anthropic 经济指数的研究《AI 对劳动力市场的影响：新测度与早期证据》（Labor market impacts of AI），作者 Maxim Massenkoff 与 Peter McCrory，2026 年 3 月 5 日发布。它用的是美国人口普查局 CPS（现行人口调查）的面板数据，逐月追踪同一批人有没有换到新工作。What the source is. "Labor market impacts of AI: A new measure and early evidence," an Anthropic Economic Index study by Maxim Massenkoff and Peter McCrory, published March 5, 2026. It uses the panel dimension of the US Current Population Survey (CPS), following the same individuals month to month to see who starts a new job.

口径是什么。 14% 降的是求职成功率：22 至 25 岁的人里，每个月成功进入一份新工作（且是高暴露职业）的比例，较 2022 年的平均水平下降了约 14%。它不是就业人数，不是失业率，也不是招聘岗位数。换算成直观的量：高暴露职业这一比例每月约降半个百分点，相对 2022 年基线降约 14%。作者自己对这个结果的显著性只给了一句限定，原话是 just barely statistically significant，意思是刚刚够到统计显著的边。同一篇研究里还有一个很少被引用的结论：高暴露职业的失业率自 2022 年底以来没有系统性上升，与失业有关的差距变化小到和零没有区别。What the number measures. The 14% is a drop in the job-finding success rate: among 22 to 25 year olds, the share who successfully start a new job in a high-exposure occupation in a given month, down about 14% relative to its 2022 average. It is not employment, not unemployment, and not a count of job postings. In plain units: in the most exposed occupations, this rate fell by roughly half a percentage point per month, a drop of about 14% relative to its 2022 baseline. The authors attach their own caution to this result, verbatim: it is "just barely statistically significant." The same study carries a second finding that is seldom quoted: unemployment in highly exposed occupations has shown no systematic increase since late 2022, and the change in the unemployment gap is indistinguishable from zero.

最常见的错读。 一是把求职成功率写成就业下降或失业上升，跟原文同篇的失业结论正好相反。二是省掉「刚刚显著」这句作者自陈，把一个作者本人都说证据尚弱的结果，写成已成定论的事实。How it is most often misread. First, a job-finding rate gets reported as falling employment or rising unemployment, the opposite of what the same paper found on unemployment. Second, the authors' "just barely significant" gets dropped, and a result the authors themselves describe as early and tentative is reported as settled fact.

第三个数：27% 到 9%，是一家公司内部的招聘占比Number three: 27% to 9%, a hiring share inside one company

原文是什么。 Anthropic 研究所的报告《Anthropic 的就业正在如何变化》（How employment is changing at Anthropic），2026 年 10 月 16 日发布，研究对象是 Anthropic 自己一家公司，数据是它内部的招聘与人事记录。报告里写到，这家公司在过去一年里规模增至约原来的 3 倍（原文原句：「It has tripled in size in the last year」）。注意这是一个独立的时间窗：过去一年；它和下文占比的「自 2025 年初以来」不是同一个窗，别混读。What the source is. "How employment is changing at Anthropic," an Anthropic Institute report published October 16, 2026. Its subject is Anthropic itself: one company, studied on its own internal hiring and personnel records. The report states that the company tripled in size in the last year, in its own words: "It has tripled in size in the last year." Note that this is a separate time window, the last year; it is not the same window as the "since early 2025" attached to the shares below, and the two should not be read as one.

口径是什么。 27% 到 9% 是占比：技术岗新招员工里，初级员工所占的比例，由约 27% 降到约 9%。同一份报告里还有另一个占比，新人在技术部门全体员工中的构成，自 2025 年初以来由 21% 降到 14%，两个占比不能混用。这里要分清一个二手转述常见的添字：原文里「自 2025 年初以来」（since early 2025）只写在员工构成那一句上，原句是「the share of junior-level employees in technical departments at Anthropic slipped from 21% to 14% since early 2025」；招聘占比是紧接着的下一句，只说它「降得更快」（fell even faster），从 27% 降到 9%，原文没有给它另设起点。关键还在分母：原文明确写了，同期该公司技术岗招聘绝对人数仍在增长。占比 = 初级人数 ÷ 技术招聘总人数；总人数在涨，占比照样可以大跌。原文通篇没有给出初级岗的绝对人数。所以从这份报告出发，能说的只有：技术岗招聘绝对人数仍在增长，降的是新人占比。「初级岗人数涨了」原文不支持，「初级岗人数崩了」原文同样不支持，两个说法都超出了这份材料能证明的范围。What the number measures. The 27% to 9% is a share: junior employees as a proportion of technical hires, down from about 27% to about 9%. A second share in the same report, juniors as a proportion of all technical staff, fell from 21% to 14% since early 2025; the two shares are not interchangeable. One piece of precision matters here, because secondary coverage routinely gets it wrong: in the original, the words "since early 2025" appear only in the staff-composition sentence ("the share of junior-level employees in technical departments at Anthropic slipped from 21% to 14% since early 2025"). The hiring share follows in the next sentence, introduced only as having "fallen even faster," from 27% to 9%, with no start date of its own. Attaching "since early 2025" to the 27% to 9% is an addition made in retelling. The denominator is the key: the report states, in its own words, that over the same period "technical hiring has continued growing in absolute terms." A share is junior hires divided by total technical hires; when the total is growing, the share can fall sharply on its own. The report nowhere gives the absolute number of junior hires. So the only statements this source supports are: technical hiring grew in absolute terms, and the junior share fell. "Junior hiring grew in absolute terms" is not supported by this source, and neither is "junior hiring collapsed." Both go beyond what the text can prove.

最常见的错读。 一是把占比当人数，写成「近乎停招」「招聘崩了」。二是把一家公司的内部构成，当成整个劳动力市场的行情。Anthropic 是一家在过去一年里规模增至约 3 倍的前沿 AI 公司，它的新人占比变化是它自己的组织选择，不是全行业的统计。How it is most often misread. First, a share gets read as a headcount, producing "almost stopped hiring" and "hiring collapsed." Second, one company's internal composition gets read as a labor market statistic. Anthropic is a frontier AI lab that tripled in size in the last year; how it rebalanced its junior share is a fact about its own organization, not a measurement of the whole market.

三个数放在一起之前，先对一件事：6% 到 16% 是怎么来的Before putting the three together: where "6 to 16 percent" comes from

Anthropic 在转述斯坦福研究时，写过一句「22 至 25 岁人群在暴露职业中的就业下降了 6% 至 16%」。这个区间常被当成不确定范围，或者被取上限当成「斯坦福的结论」引用。Citing Stanford, Anthropic's report writes that Brynjolfsson et al. report a fall in employment of 6 to 16 percent in exposed occupations among workers aged 22 to 25. This range is often treated as a margin of error, or its top end is quoted alone as "the Stanford finding."

深翼解读（这一段是我们拿两边原文对出来的拆解，Anthropic 原文没有自己定义区间的端点）：6% 和 16% 是同一篇斯坦福论文里的两个不同的数。6% 是不做控制的原始人数降幅；16% 是控制企业层面冲击后的相对降幅，而且是 2025 年 11 月修订版的数值。前面说的 13%，是同一个受控估计在 2025 年 8 月版的数值，正好落在这个区间之内，它不是第三个互相竞争的发现，而是 16% 的上一版。把区间当误差范围、把点估计当独立证据，都是从这一步开始错的。DeepWing reading (this decomposition is ours, worked out by matching the two texts; Anthropic does not define the endpoints itself): the two endpoints are two different numbers from the same Stanford paper. The 6% is the raw, unadjusted fall in headcount. The 16% is the relative decline after controlling for firm-level shocks, in the November 2025 revised version. The 13% discussed above is the same controlled estimate in its August 2025 vintage, and it falls inside the range; it is not a third, competing finding, but the earlier version of the 16%. Treating the range as an error band, or a point estimate as independent confirmation, is where the error begins.

第三方对照：别人的数据怎么说Third-party cross-checks: what other data say

这三个数之外，我们还查了三份与 Anthropic、斯坦福都无关的数据。没有一份在量同一件事，但拼起来能看出故事的边界在哪。Beyond these three sources, we checked three datasets that belong to neither Anthropic nor Stanford. None of them measures the same thing, but together they show where the story's boundaries are.

哈佛 Hosseini 与 Lichtinger（简历与招聘启事数据，近 6,200 万名员工）： 在采用生成式 AI 的企业里，初级员工的就业自 2023 年初起相对未采用企业明显下降，六个季度后，比未采用企业相对低约 9%，且主要来自招聘放缓，不是裁员增加；资深员工同期还在涨。这与斯坦福的机制方向一致，但它量的是「采用 AI 的企业」，人群按资历分，不按年龄分，数值也不是 13%。Hosseini Maasoum and Lichtinger, Harvard (US résumé and job-posting data, nearly 62 million workers): at firms adopting generative AI, junior employment declined sharply relative to non-adopters from early 2023, reaching about 9% lower after six quarters, driven by slower hiring rather than more separations, while senior employment kept rising. This matches Stanford's mechanism on a completely different dataset, but its population is adopting firms, its junior measure is seniority-based rather than age-based, and its magnitude is not 13%.

SignalFire《科技人才报告 2025》（平台追踪数据）： 大型科技公司里，应届生占新招员工的比例只剩约 7%，新招人数比 2023 年少了约 25%。但 SignalFire 自己在报告里提醒：把这全算在 AI 头上并不准确，更大的推手可能是低利率时代过度招聘之后的回调。它证实了科技行业新人入口在收窄，同时复杂化了「全是 AI 干的」这个解释。SignalFire, State of Tech Talent Report 2025 (platform tracking data): at large technology companies, new graduates account for only about 7% of hires, and new-graduate hiring counts are down about 25% from 2023. SignalFire itself cautions that blaming AI alone is not accurate, and that the larger driver may be the correction after the zero-interest-rate era's overhiring. It corroborates a narrowing entry level in tech while complicating the all-AI explanation.

耶鲁大学预算实验室（CPS 数据，更新至 2026 年 9 月）： 从整个经济看，职业结构的变化还看不出与 AI 引入清晰对齐的痕迹，AI 使用量与就业、失业的变化没有测出关联。它量的是全局，测不出 22 至 25 岁高暴露这个小格子里的变化，但它划出了一条边界：把前三个数读成全行业、全年龄的就业冲击，超出了现有证据。The Budget Lab at Yale (CPS data, updated September 15, 2026): at the level of the whole economy, the occupational mix is not yet changing in ways that clearly align with AI, and AI usage measures show no connection to changes in employment or unemployment. It measures the whole economy and cannot see inside the small cell of 22 to 25 year olds in the most exposed occupations, but it draws a boundary: reading the three numbers above as an economy-wide, all-ages employment shock goes beyond the current evidence.

三份对照的共同点：方向上，新人入口收窄在多个独立数据集里都出现；程度上，没有任何一份支持「就业人数普降 13%」或「初级招聘已经崩了」这样的说法。The three cross-checks agree on direction: a narrowing entry level shows up in several independent datasets. They agree on degree as well: none of them supports a claim that employment fell 13 percent across the board, or that junior hiring has collapsed in absolute terms.

结论：先问这个数，数的是什么Conclusion: ask what the number is counting

这三个数本身没有错，错的是引用它们的方式。下次再刷到「AI 让年轻人就业降了百分之十几」这样的话，先问三句：None of the three numbers is wrong. The ways they are quoted are. The next time a headline says AI cut young people's employment by some percent in the teens, ask three questions.

第一，它数的是人数、求职成功率，还是占比？三样东西，慌的程度完全不同。First, what is being counted: people, a success rate, or a share? The three call for very different levels of alarm.

第二，它跟谁比？是跟另一组职业比出来的相对数，还是人数本身的增减？Second, relative to what: is it a relative figure measured against another group, or an actual change in a headcount?

第三，它是哪份研究、哪一年的哪个版本？同一个研究，光版本年份不同，数字就能从 13% 变成 16%。Third, which study, and which vintage of it? In a single research program, the same estimate runs from 13% to 16% purely because the version changed.

把这三句问完，三个数就各回各的位置：一个是特定人群就业的相对变化，一个是作者自陈刚刚显著的求职成功率，一个是一家公司内部的招聘构成。它们值得认真对待，但没有一个能替你喊出「年轻人没工作了」。Asked and answered, the three numbers return to their places: a relative change in employment for one narrow group, a job-finding rate its own authors call barely significant, and a hiring share inside one company. Each deserves to be taken seriously. None of them, on its own, says that young people have lost their jobs.

出处与核查Sources and verification

一手来源（均为原文，核查日期 2026-10-24）：Primary sources (original texts, checked October 24, 2026):

  1. Brynjolfsson, Chandar & Chen，《Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence》，斯坦福数字经济实验室与 SIEPR 工作论文，2025-08-26 版、2025-11-13 修订版及 2026 年 8 月修订版。2026 年 8 月修订版作者说明（2026-08-12，差距口径 19% 出处）；作者说明（2026-02-09）。Brynjolfsson, Chandar and Chen, "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence," Stanford Digital Economy Lab and SIEPR working paper, August 26, 2025 version, November 13, 2025 revision, and August 2026 revision. authors' note on the August 2026 revision (August 12, 2026; source of the 19% gap measure); authors' explanatory note (February 9, 2026).
  2. Massenkoff & McCrory，《Labor market impacts of AI: A new measure and early evidence》，Anthropic，2026-03-05。Massenkoff and McCrory, "Labor market impacts of AI: A new measure and early evidence," Anthropic, March 5, 2026.
  3. Wilmers, Catipon, Massenkoff & McCrory，《How employment is changing at Anthropic》，Anthropic 研究所，2026-10-16。Wilmers, Catipon, Massenkoff and McCrory, "How employment is changing at Anthropic," Anthropic Institute, October 16, 2026.

第三方对照来源：Hosseini Maasoum & Lichtinger，《Generative AI as Seniority-Biased Technological Change》，工作论文（摘要经斯坦福数字经济实验室研讨会页核对）；SignalFire，《State of Tech Talent Report 2025》；耶鲁大学预算实验室，《Tracking the Impact of AI on the Labor Market》，2026-07-16 发布、2026-09-15 更新。Third-party cross-checks: Hosseini Maasoum and Lichtinger, "Generative AI as Seniority-Biased Technological Change," working paper (abstract checked via the Stanford Digital Economy Lab seminar page); SignalFire, "State of Tech Talent Report 2025," ; The Budget Lab at Yale, "Tracking the Impact of AI on the Labor Market," published July 16, 2026, updated September 15, 2026.

混引样本来源：Lucerna（2026-03-09）、OpenTools（2026-03-11）、《财富》中文网（2026-07-01）、一位 Substack 专栏作者（2026-07-29）、Adi Insights / Medium（2026-10-18）、GenZ NewZ（2026-10-19）、亿邦动力（2026-10-19）、MarsBit 转 InfoQ 中国（2026-10-20），均已逐条打开原文核对标题与原句，核查日期 2026-10-24。其中 Substack 一篇。Conflation samples: Lucerna (2026-03-09), OpenTools (2026-03-11), Fortune China (2026-07-01), a Substack columnist (2026-07-29), Adi Insights / Medium (2026-10-18), GenZ NewZ (2026-10-19), ebrun (2026-10-19), MarsBit republishing InfoQ China (2026-10-20). Each was opened and checked against its headline and quoted wording on October 24, 2026. The Substack piece.

本台登记口径：以上三份一手研究已入深翼就业台登记册（RES-001、RES-004、HIR-001），证据等级为一手已核与多源已核。本文事实部分只用一手原文；推断部分单独标为「深翼解读」。Register note: the three primary studies are held in DeepWing's jobs desk registers (RES-001, RES-004, HIR-001) with evidence grades of primary-source verified and multi-source verified. Facts above come only from primary texts; inference is labeled separately as "DeepWing reading."

署名：深翼 DeepWingByline: DeepWing

引用本文Cite this piece

深翼 DeepWing：《三个数字，三回事：被引错的 AI 就业数字，到底在数什么》，2026-10-24。https://deepwingai.com/research/three-numbers/DeepWing, "Three Numbers, Three Different Things: What the Misquoted AI Jobs Figures Actually Measure", Oct 24, 2026. https://deepwingai.com/research/three-numbers/