Anthropic 算了一笔账：如果 AI 真的跑到最快，美国经济会变成什么样
Anthropic Ran the Numbers: What Happens to the US Economy if AI Develops at Maximum Speed
📌 官方交互工具：Scenarios for our Economic Future Anthropic Institute：https://www.anthropic.com/institute/econ-scenarios 工作论文：Economic Scenarios for Transformative AI（Anthropic Institute Working Paper No. 2026-02） 论文 PDF：https://www-cdn.anthropic.com/files/4zrzovbb/website/cf58f84d46a4a76bf5a5b039ac695fba6b80041c.pdf 作者：Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter, Peter McCrory 发布日期：2026 年 9 月 9 日
BLUF：Anthropic 发布了一份罕见老实的经济报告。它没有断言 AI 会毁掉工作，也没有断言 AI 会让所有人受益——它只是建了一个模型，把”AI 进化速度”这一个变量拨到三个刻度，看美国 2030 年的 GDP、失业率、工资分配会走向哪里。最刺眼的结论在极端情景里：认知工作者失业率 17.9%，整体失业率 11.9%，但 GDP 膨胀到 44.4 万亿美元。经济总量更大，但人能拿到多少，取决于分配机制是否跟得上。
这份报告解决的是什么问题
过去两年，“AI 会不会抢工作”这个问题被讨论了无数次，但大多数讨论有一个共同问题：把定性判断当定量结论用。
Anthropic 这次换了一种做法：不预测 AI 会发展到哪里，而是问一个更干净的问题——给定 AI 的进化速度，经济会怎么走？
结果是三套可定量的情景，每套情景都有精确的 GDP 数字、失业率、工资变化和劳动/资本分配比例。用户还可以自己拨动五个维度的滑块（AI 能力、采用速度、自主程度、生产率增益、职业过渡时间线），生成个性化预测，并和 10,980 人的全国调查对比。
三个情景，三张账单
情景一：温和（Modest）——AI 是互联网级别的
核心假设：AI 对知识工作的帮助程度大约和互联网差不多，是辅助工具，不是替代者。
| 指标 | 数值 |
|---|---|
| 2030 GDP | $34.1T（较无 AI 基准 +1.6%） |
| 年均增速 | 2.4%（基准约 2%） |
| 整体失业率 | 3.9%（在历史正常范围内） |
| 劳动收入份额 | 59.4%（基准 60.0%，几乎不变） |
结论：经济有增益，但属于正常技术进步的范畴，不剧烈。
情景二：实质（Substantial）——AI 能做一半知识工作
核心假设：到 2030 年，AI 能自主完成大约一半的知识工作，实际渗透速度稍慢于这个上限。
| 指标 | 数值 |
|---|---|
| 2030 GDP | $36.3T（较基准 +8.3%） |
| 年均增速 | 5.4%（超过 1990 年代互联网泡沫高峰期的 4.7%） |
| 认知工作者失业率 | 4.5%（高于基准） |
| 认知工作者工资 | -0.3%（相对基准，几乎持平但略降） |
| 非认知工作者工资 | +5.9%（受益于总体经济扩张） |
| 劳动收入份额 | 56.1%（资本拿走更多） |
这是调查中位数所对应的情景——普通美国人的直觉预期，大致落在这里。增长够快，但知识工作者的工资不再随经济增长水涨船高。
情景三：极端（Extreme）——AI 几乎接管所有认知工作
核心假设：AI 在几乎所有知识任务上全面超越人类，并且大量自主运行。
| 指标 | 数值 |
|---|---|
| 2030 GDP | $44.4T（较基准 +32.4%） |
| 年均增速 | 15.4%（经济约每 4.5 年翻倍） |
| 认知工作者失业率 | 17.9% |
| 整体失业率 | 11.9%（超过 2008 年金融危机峰值） |
| 认知工作者工资 | -11.5%（相对基准大幅下降） |
| 非认知工作者工资 | +33.6%（劳动力稀缺溢价） |
| 劳动收入份额 | 45.2%（资本首次拿走超过一半：54.8%） |
经济体量增长 32%，但近五分之一的知识工作者失业。这不是分配不均——这是分配机制的系统性崩溃，需要政策介入才能把增长红利传递到受冲击的群体。
模型刻意排除了什么
这份报告有一个不常见的诚实声明：它明确列出了自己没有建模的东西：
- 政策响应：政府税收、社会保障、再分配如何调整
- 商业周期：正常的经济扩张/衰退节奏
- 需求侧冲击：大规模失业会压制消费，消费萎缩会反过来压制 GDP——这个循环没有纳入
- 金融市场冲击：资产泡沫、信贷危机
- 灾难性风险：模型完全不考虑 AI 失控等极端情况
- 实体机器人：只建模了认知工作，不含物理劳动替代
这意味着极端情景的 GDP 预测（+32.4%）很可能是高估——因为需求端的崩塌没有被计入。真实结果可能更复杂，不是单纯的增长，而是增长和收缩同步发生在不同部门。
公众的预期：中位数落在实质情景
Morning Consult 在 2026 年 8 月对 10,980 名美国成年人做了调查，核心问题和模型的五个维度对齐。
中位调查结果：整体失业率约 4.6%，接近实质情景。普通美国人的直觉——AI 会带来可观增长，但不会是毁灭性的——大体和研究中间情景吻合。
真正的挑战不是增长，是分配
Anthropic 在报告里给出了一个核心论断：
“The main challenge is not achieving economic growth, but making sure the benefits are broadly shared and the costs aren’t unequally dispersed.”
Anthropic 联合创始人 Jack Clark 在报告发布后补充：AI 技术进步会”极快且持续”，但经济渗透”比多数人预期的更慢”。他提到，经济增长带来的税收红利本可以用于援助被替代的知识工作者——但当前的政治环境让这条路”难以想象”。
这是这份报告真正的提示：数字只是情景，政策才是分叉点。同样的 GDP 数字，配不同的再分配机制，可以是”所有人都更富”，也可以是”经济整体更大但大多数人生活更差”。
如何使用交互工具
Anthropic 的 Econ Scenario Explorer 允许用户自己输入五个维度的预测：
- AI 能力：到 2030 年 AI 在知识工作中的上限能力
- 采用速度：企业和个人的实际渗透率
- 自主程度：AI 独立完成工作 vs. 辅助人类完成工作的比例
- 生产率增益：AI 对每个工作单元产出的放大倍数
- 职业过渡时间线：被替代的工人转型到新工作需要多久
输入后，工具会生成你的个人情景，并和全国调查中位数对比。这是目前把”AI 经济冲击”量化到个人预测层面最清晰的工具之一。
这份报告的意义
过去两年，AI 对经济影响的讨论基本上是两个极端：要么”没什么大不了”，要么”所有工作都会消失”。Anthropic 这次做的事不是提供答案，而是提供了一个让讨论变得可量化的框架。
当认知失业率 17.9% 不再是 X 上随口说出的数字，而是一套模型在特定假设下输出的结果，对话就可以从”会不会”转向”在什么条件下”——这才是有效政策讨论的起点。
© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用，须注明作者姓名及原文链接，不得去除署名后以原创发布。
📌 Official interactive tool: Scenarios for our Economic Future Anthropic Institute: https://www.anthropic.com/institute/econ-scenarios Working paper: Economic Scenarios for Transformative AI (Anthropic Institute Working Paper No. 2026-02) Paper PDF: https://www-cdn.anthropic.com/files/4zrzovbb/website/cf58f84d46a4a76bf5a5b039ac695fba6b80041c.pdf Authors: Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter, Peter McCrory Released: September 9, 2026
BLUF: Anthropic has published an unusually honest economic report. It doesn’t claim AI will destroy jobs or benefit everyone — it builds a model, sets “AI development speed” to three levels, and tracks where US GDP, unemployment, and wage distribution end up by 2030. The starkest finding is in the extreme scenario: 17.9% cognitive unemployment, 11.9% overall unemployment, but $44.4 trillion GDP. A bigger economy — but whether most people share in it depends entirely on whether redistribution mechanisms keep pace.
What Problem This Report Actually Solves
For two years, “will AI take our jobs” has been endlessly debated — with a common flaw: qualitative claims dressed up as quantitative conclusions.
Anthropic took a different approach: instead of predicting where AI will go, they asked a cleaner question — given AI’s pace, where does the economy branch?
The result: three quantified scenarios, each with precise GDP numbers, unemployment rates, wage shifts, and labor/capital distribution ratios. Users can also adjust five sliders themselves (AI capability, adoption rate, autonomy level, productivity gain, job transition timeline) to generate a personalized 2030 projection, compared against a national survey of 10,980 adults.
Three Scenarios, Three Ledgers
Scenario 1: Modest — AI as the next internet
Core assumption: AI helps knowledge work roughly as much as the internet did — a useful tool, not a replacement.
| Metric | Value |
|---|---|
| 2030 GDP | $34.1T (+1.6% vs. no-AI baseline) |
| Annual growth | 2.4% (baseline ~2%) |
| Overall unemployment | 3.9% (within historical range) |
| Labor income share | 59.4% (barely changed from 60.0%) |
Verdict: economic gains, but within the range of normal technological progress.
Scenario 2: Substantial — AI handles half of knowledge work
Core assumption: By 2030, AI can autonomously perform roughly half of knowledge work, with actual adoption slightly below that ceiling.
| Metric | Value |
|---|---|
| 2030 GDP | $36.3T (+8.3% vs. baseline) |
| Annual growth | 5.4% (faster than the 1990s dot-com peak of 4.7%) |
| Cognitive worker unemployment | 4.5% |
| Cognitive worker wages | -0.3% (flat, relative to baseline) |
| Non-cognitive worker wages | +5.9% (benefit from aggregate expansion) |
| Labor income share | 56.1% |
This is the scenario the survey median maps to — average Americans’ intuitive expectations land here. Fast enough growth, but knowledge workers no longer ride the rising tide of economic expansion.
Scenario 3: Extreme — AI near-fully replaces cognitive work
Core assumption: AI comprehensively outperforms humans at nearly all knowledge tasks and operates largely autonomously.
| Metric | Value |
|---|---|
| 2030 GDP | $44.4T (+32.4% vs. baseline) |
| Annual growth | 15.4% (economy doubles every ~4.5 years) |
| Cognitive worker unemployment | 17.9% |
| Overall unemployment | 11.9% (exceeds 2008 financial crisis peak) |
| Cognitive worker wages | -11.5% (large real decline) |
| Non-cognitive worker wages | +33.6% (scarcity premium) |
| Labor income share | 45.2% (capital first crosses 50%: 54.8%) |
The economy is 32% larger — but nearly one in five knowledge workers is unemployed. This isn’t inequality — it’s a systemic collapse of the distribution mechanism, requiring active policy intervention to transmit growth gains to displaced workers.
What the Model Deliberately Excludes
The report carries an unusually honest disclaimer, explicitly listing what it did not model:
- Policy responses: tax adjustments, social safety nets, redistribution
- Business cycles: normal expansion/contraction rhythms
- Demand-side effects: mass unemployment suppresses consumption, which suppresses GDP — this feedback loop is excluded
- Financial market disruptions: asset bubbles, credit crises
- Catastrophic risks: AI misalignment scenarios entirely excluded
- Physical robots: only cognitive labor substitution is modeled
This means the extreme scenario’s GDP projection (+32.4%) is likely an overestimate — because the demand-side collapse from mass unemployment isn’t baked in. Reality may be more complex: simultaneous growth and contraction in different sectors.
Public Expectations: Median Falls Near Substantial
Morning Consult surveyed 10,980 US adults in August 2026 using questions aligned with the model’s five dimensions.
Median result: overall unemployment ~4.6%, landing near the substantial scenario. Average Americans’ intuition — AI brings meaningful growth but not destruction — aligns roughly with the middle scenario.
The Real Challenge Isn’t Growth, It’s Distribution
Anthropic’s central thesis:
“The main challenge is not achieving economic growth, but making sure the benefits are broadly shared and the costs aren’t unequally dispersed.”
Co-founder Jack Clark added after the release: AI progress will be “at a very, very fast and sustained rate,” but economic diffusion will be “more slowly than most people assume.” He noted that tax revenue windfalls from growth could fund displaced workers — but the political path to that outcome is currently “unimaginable.”
This is the report’s real signal: the numbers are scenarios, policy is the fork. The same GDP figure, paired with different redistribution structures, can mean “everyone is richer” or “the economy is larger but most people are worse off.”
How to Use the Interactive Tool
The Econ Scenario Explorer lets users input predictions on five dimensions:
- AI capability: the ceiling of AI performance in knowledge work by 2030
- Adoption rate: actual penetration across businesses and individuals
- Autonomy level: ratio of AI completing work independently vs. assisting humans
- Productivity gain: output multiplier per unit of work
- Job transition timeline: how long displaced workers need to shift into new roles
The tool generates your personal scenario and compares it against the national survey median. It’s currently the clearest tool for quantifying “AI economic impact” at the level of individual predictions.
What This Report Actually Means
For two years, the conversation about AI’s economic impact has split between “nothing major” and “all jobs will disappear.” Anthropic didn’t provide an answer — they provided a framework for making the debate quantifiable.
When cognitive unemployment of 17.9% stops being a throwaway number on social media and becomes the output of a model under specified assumptions, the conversation can shift from “will it happen?” to “under what conditions?” — which is where effective policy discussion actually begins.
© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution. You must credit the author and link to the original; removing attribution and republishing as original is not permitted.
关于本站 · 免责声明
🍄 Mushroom Research Blog 是非营利、免费公开的个人科技观察博客与公众号 XStack18，不接受商业合作、不代表任何企业或机构立场，也不谋求商业利益。我们以个人视角客观中立地记录和分析 AI、Web3 等领域的最新模型发布与技术动态——不止转述新闻标题或二手信息，而是给出有独立思考的深入分析，希望帮更多人获得有价值的一手科技认知。
⚠️ 文中介绍的开源代码与模型，仅供学习交流与技术借鉴。它们大多仍处于早期阶段，有待进一步研究和验证，请勿直接用于工作或生产环境；如需采用，请先自行充分测试，并核实其许可证与安全性。
Open-source code and models featured here are shared for learning and reference only. Most are early-stage and still need further study and verification — please don't use them directly in your work or in production. Test them thoroughly and check their licenses and security first.
- 本站文章均为作者基于公开信息的个人研究与观点整理，不代表文中提及的任何公司、产品、模型的官方立场，未与其构成商业关联或合作关系。
- 科技行业信息更新极快，我们尽力保证内容准确、及时，但不对完整性、实时性做绝对保证，具体请以相关企业/项目官方公告为准。
- 文中引用的第三方商标、产品名称、图片、数据等版权归原权利人所有，我们会尽量注明来源；如你认为存在版权疑问或侵权，请通过下方邮箱联系我们，收到通知后会尽快核实处理（更正、加注来源或删除）。
- 文章内容仅为技术科普与个人观点，不构成投资、法律或其他专业建议，据此进行任何决策的后果需自行判断和承担。
📮 侵权 / 勘误 / 合作咨询：hello@mushroom.cv
💬 评论与讨论
使用 GitHub 账号登录后发表评论