Anthropic Economic Index: Tracking AI’s role in the US and g…

夏威夷的旅游行程规划、马萨诸塞的科学研究、印度的网页开发——表面看这三件事彼此关系不大，但恰恰是这些用途在各自地区里最为“过度代表”的 Claude 使用场景。

这并不意味着它们是全球最常见的任务：在几乎所有国家和地区，软件工程仍然遥遥领先。这里要说明的是，相较于其他地方，马萨诸塞的用户更常向 Claude 寻求科学研究方面的帮助；又比如，巴西的 Claude 用户在语言相关的用途上格外热衷，翻译与语言学习的使用频率约为全球平均水平的六倍。

这些结论来自我们第三期 《 Anthropic Economic Index 》 报告。本次我们把对 AI 早期采用模式的记录进一步扩大化，旨在展示 AI 如何开始重塑工作与经济。主要扩展包括：

• 在美国国内：首次提供各州间 AI 使用差异的详细评估。我们发现，各州经济结构会影响人均 Claude 使用率——有意思的是，人均使用最高的州并非代码占比最高的州。
• 在国家层面：新分析显示各国的 Claude 使用强烈与收入相关，且低使用率国家的用户更倾向于用 Claude 来自动化工作，而非协作增强。
• 随时间变化：与 2024 年 12 月—2025 年 1 月以及 2025 年 2–3 月的数据对比显示，“指令式”（directive）自动化对话的比例从 27% 急剧升至 39%，表明 AI 承担责任的比例（以及用户信任）在快速上升。
• 在商业用户层面：我们首次加入来自 Anthropic 一方 API 客户的匿名数据（同时仍包含 Claude.ai 的用户数据），允许我们分析企业端的互动。结果显示 API 用户远比个人消费者更倾向于用 Claude 自动化任务，这预示着潜在的重大劳动力市场影响。

我们在报告中做了总结，也制作了交互式网站，便于公开探索数据。你可以在网站上按州、按职业查看 Claude.ai 的使用趋势与结果；若想基于我们的分析继续研究，我们也将数据集开源发布。

地域分布

我们把地理数据纳入 《 Anthropic Economic Index 》，下面是关于各国与美国各州 Claude 使用模式的主要发现。

跨国比较

按总量计算， 美国 在全球 Claude 使用中占比最高（约 21.6%），其后为 印度 、 巴西 、 日本 和 韩国，几国占比相近。考虑到国家人口规模差异后，我们用每国在 Claude.ai 使用占比除以该国占全球劳动人口的比重，得到所谓的 Anthropic AI Usage Index（简称 AUI）。AUI 大于 1 表示该国的 Claude 使用超出其劳动人口比重的预期。

在 AUI 排行中，得分靠前的二十国里以以色列、 新加坡、 澳大利亚、 新西兰 和 韩国 排在前五。这部分可用收入来解释：我们发现人均 GDP 每上升 1%，平均会伴随 AUI 上升约 0.7%。这也合乎常理——高使用率国家通常互联网连通性更好、经济更偏向知识型工作而非制造业。不过这也提出了一个关于经济分化的担忧：像电气化或内燃机这样的通用技术既带来了巨大的经济增长，也扩大了全球生活水平差距。如果 AI 的影响主要集中在富裕国家，其经济后果可能类似。

美国内部模式

在美国各州层面，人均 GDP 与人均 Claude 使用量间的正相关更强：人均 GDP 每增加 1%，与人均调整后的 Claude 使用量增加约 1.8% 相关。但在州内收入对使用差异的解释力不及跨国层面强，说明还有其他因素在起作用——我们认为最重要的是各州经济构成。

例如，AUI 最高的是 华盛顿特区（District of Columbia，AUI ≈ 3.82），该地区偏多的用途包括文档编辑和信息检索，与当地以知识工作为主的经济相符。加州在编码相关任务上占比尤其高；纽约则偏向金融相关任务。即便在人均使用较低的州，如 夏威夷 ，Claude 在旅游相关任务上的使用率仍约为全国平均的两倍。我们的交互式网站列出了更多类似统计。

使用趋势

自 2024 年 12 月起，我们通过一种隐私保护的分类方法（ CLIO），将匿名化的对话按美国政府的职业与任务分类数据库 O*NET 中定义的任务组进行标注，从而分析任务类别与人机协作方式的变化。

任务类型上，自 2024 年 12 月以来，计算机与数学类用途占比保持在约 37–40% 的高位。但“知识密集型”领域在过去九个月持续增长：教育教学类对话从 9% 提升到 13%（增长逾 40%），物理与社会科学相关任务从 6% 增至 8%（增长约三成）。与此同时，传统商务类任务的相对份额下降：管理类从 5% 降至 3%，商业与金融运营类从 6% 减半至 3%。（需要说明的是，各类别对话的绝对数量仍显著增加。）总体来看，软件开发仍是我们追踪的每个国家中最常见的用途。

随着国家人均 GDP 上升，Claude 的用途呈现出从单一的计算机/数学类向更为多元的活动（教育、艺术设计、行政支持、物理与社会科学等）转移的趋势，尽管这一模式存在相当噪声。

交互模式上，我们通常把任务分为自动化（AI 在较少人工输入下直接产出工作）和增强（人机协作完成任务）。自动化内部再分为“指令式”（directive，最少人工交互）与“反馈回路”（feedback loop，需向模型反馈真实世界结果）；增强则分为学习（请求信息或解释）、任务迭代（与 Claude 协同工作）和验证（请求反馈）。

自 2024 年 12 月以来，“指令式”对话比例从 27% 上升至 39%，使得总体上自动化（49.1%）首次略超增强（47%）。这一变化可能反映了模型能力的提升（在最早收集数据时，最新的 Claude 版本为 Sonnet 3.6），用户因而更愿意信任模型在首次输出时就完成复杂工作的能力。

有趣的是，人均 Claude 使用率较高的国家反而更倾向于增强式交互，而低使用率国家更偏好自动化。控制任务组合后，人口调整后的 Claude 使用每增加 1%，自动化倾向大约减少 3%。我们尚未完全清楚原因，可能是早期采纳者更愿意与模型协作，也可能与文化或经济因素相关。

企业端的使用

我们用同样的隐私保护方法，对一部分 Anthropic 一方 API 客户的交互做了抽样，这是首次对企业端使用行为的系统分析。API 客户（多为企业与开发者）与通过 Claude.ai 访问的用户使用方式明显不同：他们按 token 付费，并通过自有程序发起请求。

在我们的样本中，API 流量在编码与行政任务上的集中度更高：约 44% 映射为计算机或数学类任务，而在 Claude.ai 上这一比例为 36%。同时，API 中教育类占比仅 4%，而在 Claude.ai 上为 12%；艺术与娱乐类在 API 上为 5%，在 Claude.ai 上为 8%。

更重要的是，API 客户更常用 Claude 来自动化任务：样本中约 77% 的 API 对话显示出自动化模式，其中绝大多数为指令式；仅约 12% 属于增强。在 Claude.ai 上，自动化与增强的比例则近乎平分。历史上任务自动化常伴随重大经济转型与生产率提升，这一差异可能预示着重要的经济影响。

鉴于 API 的计费方式，我们还检验了任务成本（由消耗的 token 数决定）是否影响企业购买哪些任务。结果显示，任务类别的平均成本与其使用份额正相关：成本更高的任务类别往往被使用得更频繁，表明对企业而言，模型的核心能力与其能创造的经济价值，往往比完成该任务的直接费用更重要。

结论

《 Anthropic Economic Index 》旨在以实证方式早期评估 AI 对就业与经济的影响。目前的主要发现是：AI 的采用极不均衡——高收入国家更常使用 Claude、更多倾向于协作式使用而非自动化、并在代码以外追求更广泛的用途；在美国，不同行业主导的地方经济显著影响 AI 的使用模式；企业比个人更愿意赋予 Claude 自主权。

特别值得注意的是，在过去九个月里，Claude.ai 上的指令式自动化大幅增长——这反映出人们对 AI 的信任在上升，同时也说明我们仍在共同决定把多少责任交给这些工具。到目前为止，趋势显示我们正在变得更愿意让 AI 代劳。我们也会在未来继续更新分析，观察随着模型改进，用户选择是否以及如何稳定下来。

如果你想自己探索数据，可访问我们专门的 《 Anthropic Economic Index 》 网站，那里有各国、各州与职业的数据交互可视化。完整报告与可下载的数据集也已发布；我们希望这些资料能帮助政策制定者、经济学家等更好地为 AI 带来的机遇与风险做准备。

附注要点

• 关于 犹他州（Utah）：在对犹他州的数据深入检查时，我们发现其中一部分使用活动显示出协调滥用的迹象，这也反映在该州较高的“指令式”自动化得分上。但经稳健性检验后，我们认为这类活动并未主导总体结果。
• 方法学：除使用 ONET 分类外，我们还用 Claude 的自有任务分类作为补充，以弥补 ONET 可能的覆盖空白；隐私保护的分析方法细节见 CLIO 相关说明。
• API 数据样本：本节数据基于 2025 年 8 月约 100 万条对话记录，随机抽自占我们一方 API 使用量约一半的客户池。我们按既定隐私与保留政策管理数据，分析亦符合相关条款与合同约定。

Travel planning in Hawaii, scientific research in Massachusetts, and building web applications in India. On the face of it, these three activities share very little in common. But it turns out that they’re the particular uses of Claude that are some of the most overrepresented in each of these places.

That doesn’t mean these are the most popular tasks: software engineering is still by far in the lead in almost every state and country in the world. Instead, it means that people in Massachusetts have been more likely to ask Claude for help with scientific research than people elsewhere – or, for instance, that Claude users in Brazil appear to be particularly enthusiastic about languages: they use Claude for translation and language-learning about six times more than the global average.

These are statistics we found in our third Anthropic Economic Index report. In this latest installment, we’ve expanded our efforts to document the early patterns of AI adoption that are beginning to reshape work and the economy. We measure how Claude is being used differently…

• …within the US: we provide the first-ever detailed assessment of how AI use differs between US states. We find that the composition of states’ economies informs which states use Claude the most per capita – and, surprisingly, that the very highest-use states aren’t the ones where coding dominates.
• …across different countries: our new analysis finds that countries’ use of Claude is strongly correlated with income, and that people in lower-use countries use Claude to automate work more frequently than those in higher-use ones.
• …over time: we compare our latest data with December 2024-January 2025 and February–March 2025. We find that the proportion of ‘directively’ automated tasks increased sharply from 27% to 39%, suggesting a rapid increase in AI’s responsibility (and in users’ trust).
• …and by business users: we now include anonymized data from Anthropic’s first-party API customers (in addition to users of Claude.ai), allowing us to analyze businesses’ interactions for the first time. We find that API users are significantly more likely to automate tasks with Claude than consumers are, which suggests that major labor market implications could be on the horizon.

We summarize the report below. In addition, we’ve designed an interactive website where you can explore our data yourself. For the first time, you can search for trends and results in Claude.ai use across every US state and all occupations we track, to see how AI is used where you live or by people in similar jobs. Finally, if you’d like to build on our analysis, we’ve made our dataset openly available, alongside the data from our previous Economic Index reports.

Geography

We've expanded the Anthropic Economic Index to include geographic data. Below we cover what we've learned about how Claude is used across countries and US states.

The US uses Claude far more than any other nation. India is in second place, followed by Brazil, Japan, and South Korea, each with similar shares.

However, there is huge variation in population size across these countries. To account for this, we adjust each country’s share of Claude.ai use by its share of the world’s working population. This gives us our Anthropic AI Usage Index, or AUI. Countries with an AUI greater than 1 use Claude more often than we’d expect based on their working-age population alone, and vice-versa.

From the AUI data, we can see that some small, technologically advanced countries (like Israel and Singapore) lead in Claude adoption relative to their working-age populations. This might to a large degree be explained by income: we found a strong correlation between GDP per capita and the Anthropic AI Usage Index (a 1% higher GDP per capita was associated with a 0.7% higher AUI). This makes sense: the countries that use Claude most often generally also have robust internet connectivity, as well as economies oriented around knowledge work rather than manufacturing. But it does raise a question of economic divergence: previous general-purpose technologies, like electrification or the combustion engine, led to both vast economic growth and a great divergence in living standards around the world. If the effects of AI prove to be largest in richer countries, this general-purpose technology might have similar economic implications.

The link between per capita GDP and per capita use of Claude also holds when comparing between US states. In fact, use rises more quickly within income here than across countries: a 1% higher per capita GDP inside the US is associated with a 1.8% higher population-adjusted use of Claude. That said, income actually has less explanatory power within the US than across countries, as there’s much higher variance within the overall trend. That is: other factors, beyond income, must explain more of the variation in population-adjusted use.

What else could explain this adoption gap? Our best guess is that it’s differences in the composition of states’ economies. The highest AUI in the US is the District of Columbia (3.82), where the most disproportionately frequent uses of Claude are editing documents and searching for information, among other tasks associated with knowledge work in DC. Similarly, coding-related tasks are especially common in California (the state with the third-highest AUI overall), and finance-related tasks are especially common in New York (which comes in fourth).1 Even among states with lower population-adjusted use of Claude, like Hawaii, use is closely correlated to the structure of the economy: people in Hawaii request Claude’s assistance for tourism-related tasks at twice the rate of the rest of America. Our interactive website contains plenty of other statistics like these.

Trends in Claude use

We’ve been tracking how people use Claude since December 2024. We use a privacy-preserving classification method that categorizes anonymized conversation transcripts into task groups defined by O*NET, a US government database that classifies jobs and the tasks associated with them.2 By doing this, we can analyze both how the tasks that people give Claude have changed since last year, and how the ways people choose to collaborate—how much oversight and input into Claude’s work they choose to have—have changed too.

Since December 2024, computer and mathematical uses of Claude have predominated among our categories, representing around 37-40% of conversations.

But a lot has changed. Over the past nine months, we’ve seen consistent growth in “knowledge-intensive” fields. For example, educational instruction tasks have risen by more than 40 percent (from 9% to 13% of all conversations), and the share of tasks associated with the physical and social sciences has increased by a third (from 6% to 8%). In the meantime, the relative frequency of traditional business tasks has declined: management-related tasks have fallen from 5% of all conversations to 3%, and the share of tasks related to business and financial operations has halved, from 6% to 3%. (In absolute terms, of course, the number of conversations in each category has still risen significantly.)

The overall trend is noisy, but generally, as the GDP per capita of a country increases, the use of Claude shifts away from tasks in the Computer and Mathematical occupation group, and towards a diverse range of other activities, like education, art and design; office and administrative support; and the physical and social sciences. Compare the trend line in the first graph below to the remaining three:

All that said, software development remains the most common use in every single country we track. The picture looks similar in the US, although our sample size limits our ability to explore in more detail how the task mix varies with adoption rates.

As we’ve discussed previously, we generally distinguish between tasks that involve automation (in which AI directly produces work with minimal user input) and augmentation (in which the user and AI collaborate to get things done). We further break automation down into directive and feedback loop interactions, where directive conversations involve the minimum of human interaction, and in feedback loop tasks, humans relay real-world outcomes back to the model. We also break augmentation down into learning (asking for information or explanations), task iteration (working with Claude collaboratively), and validation (asking for feedback).

Since December 2024, we’ve found that the share of directive conversations has risen sharply, from 27% to 39%. The shares of other interaction patterns (particularly learning, task iteration, and feedback loops) have fallen slightly as a result. This means that for the first time, automation (49.1%) has become more common than augmentation (47%) overall. One potential explanation for this is that AI is rapidly winning users’ confidence, and becoming increasingly responsible for completing sophisticated work.

This could be the result of improved model capabilities. (In December 2024, when we first collected data for the Economic Index, the latest version of Claude was Sonnet 3.6.) As models get better at anticipating what users want and at producing high-quality work, users are likely more willing to trust the model’s outputs at the first attempt.

Perhaps surprisingly, in countries with higher Claude use per capita, Claude’s uses tend towards augmentation, whereas people in lower-use countries are much more likely to prefer automation. Controlling for the mix of tasks in question, a 1% increase in population-adjusted use of Claude is correlated with a roughly 3% reduction in automation. Similarly, increases in population-adjusted Claude use are associated with a shift away from automation (as in the chart below), not towards.

We’re not yet sure why this is. It could be because early adopters in each country feel more comfortable allowing Claude to automate tasks, or it could be down to other cultural and economic factors.

Businesses

Using the same privacy-preserving methodology we use for conversations on Claude.ai, we have begun sampling interactions from a subset of Anthropic’s first-party API customers, in a first-of-its-kind analysis.3 API customers, who tend to be businesses and developers, use Claude very differently to those who access it through Claude.ai: they pay per token, rather than a fixed monthly subscription, and can make requests through their own programs.

These customers’ use of Claude is especially concentrated in coding and administrative tasks: 44% of the API traffic in our sample maps to computer or mathematical tasks, compared to 36% of tasks on Claude.ai. (As it happens, around 5% of all API traffic focuses specifically on developing and evaluating AI systems.) This is offset by a smaller proportion of conversations related to educational occupations (4% in the API relative to 12% on Claude.ai), and arts and entertainment (5% relative to 8%).

We also find that our API customers use Claude for task automation much more often than Claude.ai users. 77% of our API conversations show automation patterns, of which the vast majority are directive, while just 12% show augmentation. On Claude.ai, the split is almost even. This could have significant economic implications: in the past, the automation of tasks has been associated with large economic transitions, as well as major productivity gains.

Finally, given how API use is paid for, we can also explore whether differences in the cost of tasks (caused by differences in the number of tokens they consume) affect which tasks businesses choose to “buy”. Here, we find a positive correlation between price and use: higher-cost task categories tend to see more frequent use, as in the graph below. This suggests to us that fundamental model capabilities, and the economic value generated by the models, matters more to businesses than the cost of completing the task itself.

Conclusion

The Economic Index is designed to provide an early, empirical assessment of how AI is affecting people’s jobs and the economy. What have we found so far?

Across each of the measures we cover in this report, the adoption of AI appears remarkably uneven. People in higher-income countries are more likely to use Claude, more likely to seek collaboration rather than automation, and more likely to pursue a breadth of uses beyond coding. Within the US, AI use seems to be strongly influenced by the dominant industries in local economies, from technology to tourism. And businesses are more likely to entrust Claude with agency and autonomy than consumers are.

Beyond the fact of unevenness, it’s especially notable to us that directive automation has become much more common in conversations on Claude.ai over the past nine months. The nature of people’s use of Claude is evidently still being defined: we’re still collectively deciding how much confidence we have in AI tools, and how much responsibility we should give them. So far, though, it looks like we’re becoming increasingly comfortable with AI, and willing to let it work on our behalf. We’re looking forward to revisiting this analysis over time, to see where—or, indeed, if—users’ choices settle as AI models improve.

If you’d like to explore our data yourself, you can do so on our dedicated Anthropic Economic Index website, which contains interactive visualizations of our country, state, and occupational data. We’ll update this website with more data in future, so you can continue to track the evolution of AI’s effects on jobs and the economy in the ways that interest you.

Our full report is available here. We hope it helps policymakers, economists and others more effectively prepare for the economic opportunities and risks that AI provides.

Open data

As with our past reports, we're releasing a comprehensive dataset for this release, including geographic data, task-level use patterns, automation/augmentation breakdowns by task, and an overview of API use. Data are available for download at the Anthropic Economic Index website.

Work with us

If you’re interested in working at Anthropic to help build the systems powering this research, we encourage you to apply for our Research Engineer role.

1. As for Utah, in second: when further investigating Utah’s activity, we discovered that a notable fraction of its use appeared to be associated with indicators of coordinated abuse – which is also reflected in Utah’s much higher “directive” automation score than average. However, we ran robustness checks and believe that this activity is not driving the results.

2. We supplement this with a ‘bottom-up’ task classification in which Claude classifies conversations according to its own taxonomy, in order to address any gaps in the O*NET categories. The full details of our privacy-preserving analysis methodology are available here.

3. Data in this section covers 1 million transcripts from August 2025, sampled randomly from a pool of 1P API customers constituting roughly half of our 1P API usage. We continue to manage data according to our privacy and retention policies, and our analysis is consistent with our terms, policies, and contractual agreements.

Generated by RSStT. The copyright belongs to the original author.

Source