给分数定价的生意 · AI 与应试产业 · 2026-07The Business of Pricing a Grade · AI & the Exam Industry · Jul 2026

这个行业卖的从来不是知识,是通过的概率——它给每一分标价。AI 把「学会」的成本推向零,却在考场门口停住了脚:那里是「最后的无 AI 区」 This industry never sold knowledge — it sells the probability of passing, and it prices every single point. AI drove the cost of learning toward zero, then stopped at the exam-hall door: that is the «last AI-free zone»

把镜头拉远,你会看到本图的三重荒诞:数以千万计的人,正用最前沿的 AI,去准备一场禁止使用任何 AI 的闭卷考试,以竞争一个日常工作正在被 AI 替代的岗位。Pull the camera back and you see this page's triple absurdity: tens of millions are using the most advanced AI to prep for a closed-book exam that bans all AI, competing for a job whose daily work AI is already replacing.

AI 彻底重构了「学会」的经济学,却几乎没碰到「筛选」的政治学——它让所有人在同一条独木桥上跑得更快、更整齐,却无法让更多人过桥。桥的宽度,从来不是技术参数。AI has utterly rewritten the economics of learning, yet barely grazes the politics of selection — it lets everyone run faster and more uniformly across the same single-log bridge, but cannot let more people cross. The bridge's width was never a technical parameter.

主脊是应试流水线八节点(择考择岗→备考规划→内容课程→练习批改→模考冲刺→考试·无 AI 区→复试面试→上岸再战),每节点标注传统 vs AI + 强/中/弱渗透。另含最小单元「一分」定价、诚实层(宣传话术 vs 证据强度)、筛选的政治学四支柱、产品指南(2C 考生 / 2B 机构)。这是一张批判性行业解剖,不是报考或培训建议。聚焦成人应试(考公 / 考研 / 考证 / 语言考试);相邻议题见姊妹图:坐下来的意志→growth、思维外包→mind、名师 IP 变现→creator。 The spine is the eight-node exam-prep assembly line (choose exam → plan → content → practice/grading → mock → the exam · AI-free zone → interview → pass-or-retake), each tagged traditional vs AI + strong/medium/weak penetration. Plus the atomic unit — pricing «one point» — an honesty layer (marketing vs evidence grade), the four pillars of the politics of selection, and a product guide (2C candidates / 2B institutions). A critical industry dissection, not application or coaching advice. Focused on adult exams (civil-service / grad-school / certification / language); adjacent topics on siblings: the will to sit down → growth, cognitive outsourcing → mind, star-teacher IP → creator.

传统教学 · 知识Traditional · knowledge

AI 赋能 · 减负AI · efficiency

军备竞赛 · 提分焦虑Arms race · anxiety

无 AI 区 · 筛选信号AI-free zone · signal

人工 · 中性Human · neutral

98:1

2026 国考竞争比:371.8 万人过审、计划招录仅 3.81 万,且近七年首次缩招(减 1602 人)——供给侧被 AI 压平,需求侧岗位刚性稀缺,两条曲线背道而驰2026 civil-service exam ratio: 3.718M passed screening for just 38,100 openings — and the first hiring cut in seven years. AI flattens supply while jobs stay rigidly scarce; the two curves run opposite ways

474→343万

考研报名从 2023 峰值 474 万 连降三年至 2026 的 343 万,首次低于国考——独木桥没有变宽,只是人流改道去挤另一座Grad-school sign-ups fell from a 4.74M peak (2023) to 3.43M (2026), dropping below the civil-service exam for the first time — the bridge didn't widen; the crowd rerouted to another one

148门 / 2 天

Duolingo 两天内一次上线约 148 门 AI 生成课程——而前 100 门课花了 12 年。AI 把「内容生产」的边际成本推向零,「学会」不再稀缺Duolingo shipped ~148 AI-generated courses in two days — the first 100 took 12 years. AI drives the marginal cost of content to zero; «learning» is no longer scarce

+354%

华图 2025 净利同比 +354%、营收 31.98 亿元反超成行业第一;三巨头全部「All in AI」,赌的却是同一批缩招的岗位Huatu's 2025 net profit rose 354%, revenue ¥3.198B — now #1; all three giants went «all in on AI», betting on the same shrinking pool of jobs

口径警告:本页是批判性行业分析,非报考 / 培训 / 投资建议。三份研究文档聚焦成人应试(考公 / 考研 / 考证 / 语言考试),K12 与硬件作外溢层。市场规模「8700 亿→1.2 万亿」口径不一(弗若斯特沙利文等,量级参考)。凡「提分 / 降本 80% / 批改提速」类几乎全是厂商自述(D 级),已逐条标注;松鼠 AI 提分率、Bloom「2 sigma」原文出处,三份文档均查无,故不上图不杜撰。信号理论 Caplan「80% 来自信号」有争议(反方估 8%–35%,须并陈)。国考 / 考研数据为官方通报、Duolingo 为 SEC 8-K(A 级);张雪峰 2026-03 猝死为据报道;字节河马爱学 / 豆包爱学文档未覆盖故不列。每张公司卡片右上角 A/B/C/D＝证据强度;产品指南非推荐背书。归因纪律:同期发生的巨变不一定由 AI 造成——新东方停掉 K12 学科辅导是「双减」所迫(2024 年 20-F 仍以此为重心),公开材料里看不到它大规模部署 AI 助教或自适应练习;印度 edtech 融资从 2021 年 41 亿美元跌到 2023 年 3.21 亿(−87%)、2024 上半年回到 1.64 亿,这条曲线跟的是疫情退潮与资本环境,不是模型能力。凡本图把变化算到 AI 头上的地方,都先排掉这两类解释。 Basis warning: this is a critical industry analysis, not application / coaching / investment advice. The three source reports focus on adult exams (civil-service / grad-school / certification / language); K12 and hardware are spillover. Market-size figures («¥870B→¥1.27T») disagree across sources (Frost & Sullivan et al., order-of-magnitude). Any «score boost / 80% cost cut / faster grading» claim is almost always a vendor claim (grade D), flagged individually; Squirrel AI score-lift rates and the primary source for Bloom's «2 sigma» appear in none of the three reports, so they are neither charted nor invented. Signal theory's «Caplan 80% signaling» is contested (critics estimate 8%–35%; both shown). Civil-service/grad-school data are official; Duolingo is SEC 8-K (grade A); Zhang Xuefeng's Mar-2026 death is per reports; ByteDance's edu apps aren't covered by the reports and are omitted. Each company card's top-right A/B/C/D = evidence strength; the product guide is not an endorsement. Attribution discipline: upheavals that happen in the AI era were not necessarily caused by AI — New Oriental ended K12 academic tutoring because the «double reduction» policy forced it (its 2024 Form 20-F still frames the pivot that way), and public filings show no large-scale deployment of AI tutors or adaptive practice; Indian edtech funding fell from $4.1B in 2021 to $321M in 2023 (−87%), recovering to $164M in H1-2024 — a curve that tracks the end of the pandemic and the cost of capital, not model capability. Wherever this map attributes a change to AI, those two explanations are ruled out first.

◆ 最小单元 · 地基The atomic unit · foundation

「一分」的定价与「测评→反馈→再练」闭环Pricing «one point» & the assess→feedback→drill loop

赌博的最小单元是「一次拉杆」,成瘾是「变率奖励回路」——应试是给「多拿一分」标价的一次交易。行测多一分、申论多两分、考研数学再抢五分,每一分都对应「上岸概率」的一个微小增量。所有产品(网课 / 题库 / 批改 / 押题 / 协议班 / AI 刷题班),归根到底都在为「多拿一分的边际成本」定价。Gambling's atom is «one pull of the lever», addiction's is «the variable-reward loop» — the exam industry is a transaction that prices «one more point». One point on the aptitude test, two on the essay, five more on grad-school math — each maps to a tiny increment of «the probability of passing». Every product (courses / question banks / grading / predicted questions / guarantee classes / AI drill classes) ultimately prices «the marginal cost of one more point».

▸ 定价单元The pricing unit

1 分point

对外讲「教育改变命运」,用户真正买的是「更可见、可认证、有通过概率的信号」。商业语言不是「课程完成率」,而是「一分定价 + 概率定价」——这是全行业最反直觉、也最诚实的底层。The pitch is «education changes destiny»; what users actually buy is «a more visible, certifiable signal with a probability of passing». The business language isn't «course completion» but «pricing per point + pricing by probability» — the industry's most counter-intuitive, most honest substrate.

▸ 生产单元 · AI 在此爆发The production unit · where AI erupts

测评Assess → 反馈Feedback → 再练Drill ↺

闭环密度 > 听课时长。做题—反馈—纠错的闭环转得越密,提分越快。AI 恰好在这里爆发:批改、答疑、出题、诊断的边际成本被推向零,让这个闭环第一次可以无限、即时、个性化地转。但请把这句话钉在墙上——用「听课 / 看 AI 讲解」替代「自己做题」,是头号假学习陷阱。Loop density > lecture hours. The tighter the do–feedback–correct loop spins, the faster the score rises. And here is exactly where AI erupts: the marginal cost of grading, Q&A, item-generation and diagnosis goes to zero, letting the loop spin infinitely, instantly, personally for the first time. But nail this sentence to the wall — replacing «doing problems» with «watching lectures / AI explanations» is the number-one fake-learning trap.

立场声明:本页是批判性、祛魅的行业结构分析——拆开机制,是为了让你看清「你买的到底是知识,还是通过的概率」。不美化、不教唆,不构成报考或培训消费建议;产品指南是市场地图而非推荐背书。凡厂商自述的效率 / 提分数字,一律显式降权标注。 Stance: a critical, demystifying structural analysis — mechanisms are taken apart so you can see whether you're buying knowledge or the probability of passing. Nothing glamorized or instructed; not application or coaching advice; the product guide is a market map, not an endorsement. Every vendor-claimed efficiency/score figure is explicitly downweighted.

从这张图看到的五条规律Five patterns this map makes visible