入口收窄,不是失业潮 · 「AI 吃白领」汇总母图 · 2026-07The entrance narrows — not a jobs crash · the parent map of «AI eats white-collar» · Jul 2026

失业潮没有来。来的是一件更安静的事:职业阶梯的第一级,被拆走了——整体失业率纹丝不动,应届生却站不上梯子,而站在上面的人,价码还在涨 The jobs crash never arrived. Something quieter did: the first rung of the career ladder was removed — headline unemployment hasn't moved, yet new graduates can't step on, while those already up the ladder cost more than ever

最小单元=「一次匹配」:一份劳动被定价、被成交。看清冲击的关键在于:企业没有在裁员报表上动手,而是悄悄不再开新坑——初级岗承担的标准化信息处理,恰好是 LLM 的主场。所以正确的问题从来不是「失业率什么时候起飞」,而是入口有多窄、任务改写了多少、信号还可不可信、组织怎么重新分工。失业率 4.2% 与入口坍塌同时为真——这正是传统统计对它失明的原因。The atomic unit is «one match»: a unit of labour priced and closed. The key to seeing the impact: firms never touched the layoff line — they quietly stopped digging new seats, because the standardised information work of junior roles is exactly the LLM's home turf. So the right question was never «when does unemployment spike», but how narrow is the entrance, how many tasks got rewritten, can signals still be trusted, how is the division of labour redrawn. A 4.2% headline rate and a collapsing entrance are true at once — which is precisely why the statistics are blind to it.

诚实底色=信号通胀 + 副业数学。求职者用 AI 海投,HR 用 AI 去筛 AI 写的简历,单岗申请翻倍(每分钟 8,200–11,000 份)——文本一旦免费,可信就开始收费:作品、内推、弱关系这些造不了假的信号,全在涨价。至于「副业致富」:副业月收入中位数 200 美元;这个生态里最稳的现金流,属于卖副业课的人。The honest baseline: signal inflation + side-hustle math. Seekers mass-apply with AI while HR uses AI to screen AI-written résumés; applications per opening double (8,200–11,000 a minute) — once text is free, credibility starts charging: portfolios, referrals and weak ties — the signals that cannot be faked — are all repricing upward. As for «get rich on a side hustle»: the median hustle nets $200 a month, and the steadiest cash flow in this ecosystem belongs to whoever sells the course.

主脊是劳动生命周期八节点 × 雇佣轨/自雇轨双轨:技能形成→求职→信号包装→筛选面试→在职→过渡→再就业转自雇→退出。AI 在做一件双手互搏的事——一只手收窄雇佣轨的入口,另一只手拆掉自雇轨的门槛。这张图是「AI 吃白领」系列的母图:四条白领赛道的 junior 同步坍塌,各自的深挖在四张子图——软件→code、设计→design、财会→ledger、法律→law;一人公司→startup、卖课生态→creator。 The spine: the eight-node labour lifecycle × the employment/self-employment double track — skilling → search → signalling → screening → on-the-job → transition → re-employment or independence → exit. AI is playing both hands at once — one hand narrowing the employment entrance, the other dismantling the barrier to self-employment. This is the parent map of «AI eats white-collar»: four white-collar tracks collapsing at the junior level in lockstep, each dug deep in its own child map — software → code, design → design, finance → ledger, law → law; the one-person company → startup, the course economy → creator.

传统节点Traditional

入口收窄 · 梯子第一级 · 危机Narrowing entrance · crisis

雇佣轨 · AI 重构 · 信号通胀Employment track · AI · signal inflation

自雇轨 · 一人公司Self-employment · one-person co.

判断层 · 可验证信号Judgment · verifiable signal

9.7% vs 3.6%

同一个劳动市场的两种天气:20–24 岁本科毕业生失业率 9.7%(2025-09,一年前 6.8%),25–34 岁学历者仅 3.6%——门外排队,门内加薪Two climates in one labour market: unemployment for 20–24 bachelor's grads at 9.7% (Sep 2025, up from 6.8%) vs just 3.6% for degree-holders 25–34 — a queue outside the door, raises inside it

−50%

毕业不足一年者的「新岗位起步数」五年腰斩(SignalFire,2019–2024),销售/工程/HR/设计/财务/法务无一幸免;大科技应届占招聘 15%→7%New-role starts for those under a year out of school halved in five years (SignalFire, 2019–2024) — sales, engineering, HR, design, finance, legal, none spared; big tech's grad share of hires 15%→7%

8,200–11,000/分

LinkedIn 每分钟收到的申请数(8,200 官方口径/1.1 万 2025-07 报告)。AI 海投遇上 AI 筛选,申请成本归零的代价是信号全体失真Applications LinkedIn receives per minute (8,200 official / 11,000 per a Jul 2025 report). AI mass-applying meets AI screening — the price of free applications is the debasement of every signal

$200/月

副业月收入中位数(Bankrate 2025,还在跌:2024 年是 250)。「自雇自由」的分布长尾冰冷;53% 创作者的变现方式是卖课——教你做副业,就是他们的副业Median monthly side-hustle income (Bankrate 2025, still falling: $250 in 2024). The long tail of «independence» runs cold; 53% of creators monetise by selling courses — teaching you the hustle is their hustle

口径警告:本页是批判性行业分析,不是求职或投资建议;官方统计、权威调研与厂商自述并置。失业率两口径(2026-06=4.2% BLS 最新/2025-09=4.4% 四年最高,时点不同不矛盾);应届失业率 9.7%(20–24 岁本科)与 5.7%(纽约联储应届生)是不同口径;每分钟申请量 8,200 与 1.1 万两读数并呈。junior 降幅五个数字不可相加:−8%(哈佛·AI 公司六季度)/−10%(软件初级 18 个月)/−20%(斯坦福·22–25 岁软件自 2022 峰值)/−50%(SignalFire·新岗起步)/−24%(Randstad·金融 0–2 年)——研究、赛道、时窗各不相同。AI 是否主因四方分歧(加息论/Anthropic/丹麦/斯坦福 ADP),本页结论只有一句:AI 加速了既有下行,不是唯一原因。厂商自述与预测(含 Amodei「1–5 年砍半初级白领」)一律标 D。流传最广的那几句话反而进不了证据链:「前端工程师正在灭绝」「某职业 90% 的任务可被 AI 替代」这类说法找不到可追溯的一手来源或同行评议支持——以前者为例,它出自某风投基于 6.5 亿科技工作者数据的内部分析,方法细节与具体数字均未公开,也无学术或官方复核,故本页只在局限中标注、不纳入结论。这正是本图题眼的由来:被数据支持的是入口端的结构性收窄,不是那些更耸动、更好传播的整体消失叙事。卡片右上角 A/B/C/D=证据强度。 Basis warning: a critical industry analysis, not career or investment advice; official statistics, serious surveys and vendor claims sit side by side. Unemployment carries two readings (Jun 2026 = 4.2% latest BLS / Sep 2025 = 4.4% four-year high — different dates, no contradiction); grad unemployment 9.7% (20–24 bachelor's) vs 5.7% (NY Fed new grads) are different bases; applications per minute shows both 8,200 and 11,000. The five junior-decline figures do not add: −8% (Harvard, AI firms, six quarters) / −10% (junior software, 18 months) / −20% (Stanford, 22–25 software from the 2022 peak) / −50% (SignalFire, new-role starts) / −24% (Randstad, finance 0–2 yrs) — different studies, tracks, windows. Whether AI is the cause splits four ways (rate hikes / Anthropic / Denmark / Stanford ADP); this page concludes one sentence only: AI accelerated an existing decline and is not its sole cause. Vendor claims and forecasts (incl. Amodei's «halve entry-level in 1–5 years») are graded D. The most-repeated lines are exactly the ones that fail to enter the evidence chain: claims such as «front-end engineers are going extinct» or «90% of the tasks in occupation X can be replaced» have no traceable primary source and no peer-reviewed support — the first traces to a venture firm internal analysis of data on 650 million tech workers, with neither method nor underlying figures published and no academic or official replication, so this page records it among limitations rather than conclusions. Which is where the title of this map came from: what the data supports is a structural narrowing at the entrance, not the more dramatic, more shareable story of wholesale disappearance. Card badges A/B/C/D = evidence strength.

◆ 中心结构判断 · 梯子第一级被抽掉The core structural read · the first rung is pulled out

下面的人上不来,上面的人在涨价(seniorization)Those below can't climb on; those above cost more (seniorization)

LinkedIn 首席经济机会官 Aneesh Raman 的判断是:AI 正在摧毁「职业阶梯的底层横档」。机制并不戏剧化——没有裁员公告,只有不再放出的新编制:初级岗的本质是标准化、可重复、有正确答案的信息工作,而这正是 AI 定价最低的那类任务。PwC 给这个现象起了名字:「资深化 seniorization」——招聘启事要求 22 岁的人交出 35 岁的能力证明。被拆掉的不只是某一代人的第一份工作:初级岗同时是资深判断力的育苗床,今天不种,十年后无树。Aneesh Raman, LinkedIn's chief economic-opportunity officer, puts it plainly: AI is destroying «the bottom rungs of the career ladder». The mechanism is undramatic — no layoff memo, just new headcount that never opens: junior work is standardised, repeatable information processing with checkable answers, exactly the class of task AI prices lowest. PwC named the result «seniorization» — job ads asking a 22-year-old for a 35-year-old's proof of ability. And what got removed is more than one cohort's first job: the junior seat is the seedbed of senior judgment. Skip the planting today, and there is no tree in ten years.

▚ 职业阶梯 · 第一级被 AI 抽掉(示意)The career ladder · the first rung removed by AI (illustrative)

资深 / 架构师Senior / architect判断、架构与担责没有平替,暂时安全甚至溢价;资深工程师产能 2x–5xJudgment, architecture and accountability have no substitute — safe, even a premium; senior productivity 2x–5x失业率 3.6%3.6% unempl.

中级Mid-level经验 + 会指挥 AI 的人仍被要;岗位重心向「AI 指挥者」上移Experience plus the ability to conduct AI still sells; the role shifts up to «AI conductor»需求仍在still hiring

初级 juniorJunior梯子第一级 · 被 AI 抽掉:标准化任务被自动化,编制静悄悄冻结The first rung · removed by AI: standardised tasks automated, headcount frozen without a sound招聘 −8%~−10%hiring −8~−10%

应届生New grad站不上去:入行通道被截断,全部溢出压力落在这一层;新岗起步数 2019–2024 −50%Can't step on: the entry channel is severed and all the spillover lands here; new-role starts −50% (2019–2024)失业率 9.7%9.7% unempl.

−50%

SignalFire·新岗起步数(2019–24)SignalFire·new-role starts

−8%

哈佛·AI 公司 junior 招聘(6200 万工人)Harvard·AI-firm junior hiring (62M workers)

−20%

斯坦福·22–25 岁软件(2022 峰值起)Stanford·22–25 software (from 2022 peak)

−24%

Randstad·金融 0–2 年岗招聘帖Randstad·finance 0–2 yr postings

⚠️五个 junior 降幅是五把不同的尺子——研究、赛道、时窗都不同(哈佛 −8% 量的是「采用 AI 的公司六个季度」,斯坦福 −20% 量的是「22–25 岁软件自 2022 峰值」,SignalFire −50% 量的是「新岗起步数」),不可相加。但五把尺子量出同一个形状:入口在收窄,而失业率对此不设刻度。⚠️The five junior-decline figures are five different rulers — different studies, tracks and windows (Harvard's −8% measures «AI-adopting firms over six quarters», Stanford's −20% «22–25 software from the 2022 peak», SignalFire's −50% «new-role starts») — they do not add. Yet all five rulers trace the same shape: a narrowing entrance, for which the unemployment rate has no gradation.

◆ 判断层杠杆排序 · 出路结论卡The leverage stack · what actually works

赛道与雇主选择Track & employer choice ＞面试技巧interview skill

先选有物理护城河或与 AI 互补的赛道,再谈技巧。BOSS:电工工资 +56%、招聘需求 11 倍——「AI 无法拧螺丝」Pick a track with a physical moat or AI complementarity before polishing technique. BOSS: electrician pay +56%, demand 11× — «AI can't turn a screw»

可验证作品Verifiable work ＞简历措辞résumé wording

措辞可以由 AI 无限生成,作品不能。作品集/开源/真实交付=造不了假的信号;雇主的排序早已是:有实习经验 > 4.0 GPA 无经验Wording is infinitely generatable; work is not. Portfolio / open source / real delivery = signals that cannot be faked; employers already rank internship experience > a 4.0 without any

弱关系内推Weak-tie referral ＞海投mass-applying

内推是拿别人的信誉抵押你的可信度:录用率 28.5% vs 海投 2.7%(哥大);占申请 7% 却贡献 30–50% 录用,留存还高 46%A referral pledges someone else's credibility against yours: 28.5% hire rate vs 2.7% cold (Columbia); 7% of applications, 30–50% of hires, 46% better retention

现金储备的跑道长度Cash-runway length ＞履历完美度a perfect CV

过渡期在变长(首个 offer 中位 68.5 天,+22%)。跑道长度决定你能不能等得起好机会、转得起自雇——它是议价权,不是存款Transitions are lengthening (median 68.5 days to a first offer, +22%). Runway decides whether you can afford to wait for the right seat or pivot independent — it is bargaining power, not savings

底层逻辑一句话:当申请文本可以免费生成,一切「写出来」的信号作废,一切「做出来、有人担保」的信号涨价。求职竞赛已经从「比谁投得多」改写成「比谁的信号更难造假」。The logic in one line: once application text is free to generate, every «written» signal is void and every «made and vouched-for» signal appreciates. The competition has been rewritten from «who applies most» to «whose signal is hardest to fake».

从这张图带走的五条规律Five patterns to take away

求职与职业工具地图(按你所处环节)Career tools, by node