「AI 吞噬世界」全景图系列 · 公共行政篇 · GOVTHE «AI IS EATING THE WORLD» ATLAS · PUBLIC ADMINISTRATION
公章之前全面失守,
公章之后纹丝不动Before the Seal, Total Collapse —
After the Seal, Not an Inch
一枚公章的距离 · The Distance of One Seal · 填核写看算 vs 批签章责The Distance of One Seal · five verbs fallen, four verbs standing
AI 已经把公共部门的「填、核、写、看、算」大规模自动化:公文起草成了事实标准,秒批把零裁量事项做成无人审批,热线全量质检取代 1% 抽检,电网调度离了 AI 已转不动。但「批、签、章、责」四个动作几乎原地不动——因为一次行政行为必须锚定一个能被追责的人。AI 可以把材料备到 99%,最后一枚公章的落下,必须是一只人的手:AI 优化的是效率,而公共部门的核心产出是合法性。AI has automated the public sector’s filling, checking, writing, watching and computing at scale: document drafting is the de facto standard, instant approval runs zero-discretion items unmanned, hotlines audit every call instead of 1%, and the grid no longer turns without AI. Yet the four verbs — decide, sign, seal, answer — have barely moved, because an administrative act must anchor to a person who can be held to account. AI can prepare the file to 99%; the seal’s fall must be a human hand: AI optimises efficiency, and the public sector’s core product is legitimacy.
本图相信THIS MAP BELIEVES
公文大模型、秒批、热线智能化、电网 AI 调度的渗透是真的(四源+官方文件);《政务大模型指引》《数据共享条例》的制度推动是真的(A);新能源装机 60.4% 逼出的调度革命是真的(A);86:1 的编制热是真的(A/B)。The penetration of document LLMs, instant approval, hotline AI and grid dispatch is real (four sources plus official documents); the institutional push of the Government-LLM Guideline and the Data-Sharing Regulation is real (A); the dispatch revolution forced by 60.4% renewable capacity is real (A); the 86:1 exam fever is real (A/B).
本图不相信THIS MAP DOES NOT BELIEVE
不信「AI 公务员取代公务员」——减的是活,不是编;不信厂商准确率(95%/98%/「七成自主处置」,一律 D);不信效率能替代合法性——快而越权=无效;不信技术能修复考核错位——指挥棒不改,AI 只是形式主义的涡轮增压;不信「数字利维坦」与「效率革命」的任何单边叙事。Not «AI civil servants replacing civil servants» — the work shrinks, the establishment doesn’t; not vendor accuracy (95%, 98%, «70% autonomous handling» — all D); not efficiency substituting legitimacy — fast but ultra vires is void; not technology fixing misaligned incentives — with the baton unchanged, AI is formalism’s turbocharger; and not either one-sided tale, digital Leviathan or efficiency revolution.
蓝 = 五动作(填核写看算,AI 已入)Blue = the five verbs (AI in)
章泥红 = 四动作与硬骨头(批签章责)Seal red = the four verbs & the bones
A–D = 证据分级(A 官方 → D 厂商自述)A–D = evidence grade (A official → D vendor)
渗透% = 评估值,非统计值Penetration % = estimate, not statistics
98 : 12026 国考竞争比(371.8 万过审、连续 7 年攀升;招录 3.81 万为 2019 年来首次缩招 · A/B)——「AI 越猛,编制越香」撞上供给收紧The 2026 civil-exam ratio (3.718M qualified, a seventh straight climb; 38.1k hires, the first cut since 2019 · A/B) — «the fiercer the AI, the sweeter the establishment» meets a tightening gate
15 分钟 → 5 分钟浦东 12345 单条工单审核时长(2025 · B)——热线智能化最干净的数字Pudong’s 12345 per-ticket review time (2025 · B) — hotline AI’s cleanest number
60.4%可再生能源装机占比(2026Q1,国家能源局 · A;装机≠发电量)——调度革命的背景板Renewables’ share of installed capacity (2026Q1, NEA · A; capacity ≠ generation) — the dispatch revolution’s backdrop
945 TWhIEA 预测 2030 全球数据中心耗电(2025 报告 · B)——「电网自指」的全球尺度IEA’s 2030 global data-centre power forecast (2025 report · B) — the self-referential grid at planetary scale
⚠️ 口径裁判(先读):① 86:1 为「过审/招录」平均口径,岗位差异极大;② 60.4% 是装机占比不是发电量占比;③ 945TWh 为 IEA 2025 版预测,2026 版口径为「2025 年 +17%、2030 翻倍」,两版勿混引;④ 15→5 分钟为单条工单审核环节,非全流程时长;⑤ 龙岗 95%/98%、「七成事件自主处置」、「节电 10–30%」、「提前 3–8 周/91%」均厂商或营销口径(D);⑥ 福田数字员工口径不一(70 名/90%/成本 20–30%,C),主叙述权在 admin 图,本页只互链;⑦ 各节点渗透%为编者评估值非统计数;⑧「90→22 分钟」案例城市不明(C);⑨ 四份来源已于 2026-07 重跑修订(旧版两份截断、引文缺失;修订版全部完整收束、带可溯源引文)——本页据修订版补强,旧版孤例维持降 C。⚠️ The basis rulings (read first): ① 86:1 is the average qualified-to-hired basis with vast per-post variance; ② 60.4% is installed capacity, not generation; ③ 945TWh is the IEA’s 2025-edition forecast — the 2026 edition reads «+17% in 2025, doubling by 2030»; never mix editions; ④ 15→5 minutes covers the review step, not the whole flow; ⑤ Longgang’s 95%/98%, «70% autonomous handling», «10–30% power savings» and «3–8 weeks / 91%» are vendor or marketing bases (D); ⑥ the Futian digital-staff figures vary by basis (C) and their primary narrative lives in the admin map — this page only cross-links; ⑦ node penetration %s are editorial estimates; ⑧ the «90→22 minutes» case names no city (C); ⑨ the four sources were re-run and revised in Jul 2026 (the old set had two truncations and missing citations; the revision closes complete with traceable links) — this page is reinforced from the revision, with old orphan details still capped C.
中心装置 · 公章原子THE CENTRAL DEVICE · THE SEAL ATOM
AI 把「准备」全包了,把「决定」还给了人AI took over all the preparing — and handed the deciding back to the human
本图的最小分析单元是一次有法律效力的行政行为——「一枚公章的落下」。公章之前的准备动作(填、核、写、看、算)已经大规模失守;公章本身与其后的责任(批、签、章、责)纹丝不动。这不是技术还不够好,而是制度性、政治性的结构约束:行政行为必须锚定可追责的自然人,重大决策终身追责——AI 不能签字、不能盖章、不能坐被告席、不能被问责。2025 年的《政务大模型部署应用指引》把这条线写成了国家规矩:辅助型定位,不得直接作出有行政效力的实质决定。This map’s atomic unit is one legally effective administrative act — «the fall of one seal». Everything before the seal — filling, checking, writing, watching, computing — has fallen at scale; the seal itself and what follows — deciding, signing, sealing, answering — has not moved. Not because the technology falls short, but by institutional and political design: every act must anchor to an accountable person, with lifetime liability for major decisions — AI cannot sign, cannot seal, cannot take the defendant’s seat, cannot be held to account. The 2025 Government-LLM Deployment Guideline wrote that line into national rule: assistive positioning only, no substantive decisions with administrative effect.
效率层 · 前台服务Efficiency · the front desk
一网通办、秒批、免申即享、热线智能化——冲击最可见、口碑最好的一层:办事确实变快了。One-portal service, instant approval, proactive benefits, hotline AI — the most visible, best-reviewed layer: things really did get faster.
科层层 · 后台现实Bureaucracy · the back office
文书、留痕、督检考、编制——AI 在这里是旧激励的涡轮增压:前台越顺滑,后台越可能空转。Paperwork, audit trails, inspections, the establishment — here AI is the old incentive’s turbocharger: the smoother the front, the harder the back may spin in place.
合法性层 · 天花板Legitimacy · the ceiling
问责、裁量、授权、程序——制度常量。快但越权/违程序=无效;慢但合法合规=有效:公共部门的 AI 天花板因此比任何行业都低。Accountability, discretion, mandate, procedure — the institutional constants. Fast but ultra vires is void; slow but lawful stands: which is why the public sector’s AI ceiling sits lower than any industry’s.
九个动作的分界线 · 公章为界(渗透为评估值)NINE VERBS, ONE DIVIDE · THE SEAL AS BORDER (PENETRATION = ESTIMATES)
填 · 核Fill · Check
表单预填、OCR 要件核验、电子证照免交、一表通取数——受理准备已大半交给机器(~55–80%);受理决定本身仍需人签。Pre-filled forms, OCR verification, licence-free submission, one-form data pulls — intake preparation largely machine-run (~55–80%); the acceptance decision itself still takes a signature.
写 · 看 · 算Write · Watch · Compute
公文大模型成事实标准(~70%)、热线全量质检替代 1–2% 抽检、视觉巡查扩容「发现」、电网调度离 AI 转不动——信息生产与监测,全面失守。Document LLMs the de facto standard (~70%); hotlines auditing 100% instead of 1–2%; visual patrols multiplying «discovery»; the grid undispatance without AI — information production and monitoring, fully fallen.
批 · 签Decide · Sign
秒批只吃零裁量羁束事项(~80%);含裁量即退回人审(~20%)——裁量权是授予人的权力,不是授予模型的;拍板的判断与责任,AI 拿不到合法授权。Instant approval eats only zero-discretion bound items (~80%); anything discretionary returns to human review (~20%) — discretion is a power vested in persons, not models; the call and its liability carry no lawful mandate for AI.
章 · 责Seal · Answer
电子证照 ~90% vs 落章授权 ≈0%——证是数据,章是责任:电子印章只是把章数字化,加盖的授权仍来自有职权的人;落章的同时落下的,是责任。E-licences ~90% vs seal authority ≈0% — the certificate is data; the seal is liability: the e-seal merely digitises the stamp, whose authority still flows from an office-holder; what falls with the seal is the answerability.
读图法一句话:越是技术问题,AI 越好用;越是权力问题,AI 越受限——平行轨(电水气)与行政轨的对照,本身就是这张图的论点。The reading in one line: the more technical the problem, the better AI works; the more it concerns power, the harder AI is bound — the contrast between the utility track and the administrative track is itself this map’s thesis.
Reading the MapReading the Map
从这张图带走的五条规律Five patterns to take away
立场声明:本页是批判性、祛魅的行业结构分析——四份深度研究交叉整理(两份中途截断、一份自认离线撰写,数字均降级;框架共识部分源于委托预设论纲,已如实声明);厂商准确率一律 D 级,渗透%为评估值。材料 ≈95% 为中国叙事,本页坦率定位为「中国数字政府的 AI 跃迁」,欧美 GovTech 对照留白待补,不凭常识编造爱沙尼亚细节。本页提供制度结构信息,不构成任何办事、报考或采购建议;不使用国徽、五角星或真实机关公章图样。核心判断一句话:AI 把「准备」全包了,把「决定」还给了人——效率是它的战利品,合法性不是。
Stance: a critical, demystifying structural analysis — cross-compiled from four deep-research reports (two truncated, one self-declared offline, figures downgraded; where the four converge partly echoes the commissioning brief, declared honestly); vendor accuracy is D-grade throughout, penetration %s are estimates. The material is ~95% a China narrative — this page frankly frames itself as «the AI leap of China’s digital government», leaving the Western GovTech comparison blank rather than inventing Estonian details. Institutional-structure information only — no advice on procedures, examinations or procurement; no state emblems, stars or real-agency seal imagery are used. The core judgment in one line: AI took over the preparing and handed back the deciding — efficiency is its trophy; legitimacy is not.