核心发现：
- 我们近期对 81,000 名 Claude 用户进行的调查显示，在 AI 接触程度较高岗位工作的人，对 AI 导致岗位替代的担忧也更多。这种担忧在职业生涯早期的受访者中同样更为突出。
- 收入最高和最低职业群体报告的生产力提升幅度最大，最常见的原因是工作范围扩大（从事新任务）。
- 因 AI 而获得最大效率提升的受访者，对岗位替代的担忧程度也更高。
为了让公众了解我们观察到的 AI 带来的经济变化，我们的经济指数分享了 Claude 被要求执行哪些工作，以及在哪些岗位中 Claude 承担了最大比例的任务。然而，迄今为止，我们尚缺乏关于这些使用模式如何映射到人们对 AI 的看法和印象的信息。
我们近期对 81,000 名 Claude 用户进行的调查研究，提供了一种将人们的经济担忧与我们在 Claude 流量中量化的情况联系起来的方法。
该调查询问了人们对 AI 发展的愿景和恐惧。人们分享的许多想法都涉及经济话题。我们了解到，许多人担心岗位替代——尽管他们也感觉自己在工作中效率更高、能力更强。在某些情况下，AI 使他们能够创业，或为他们腾出时间处理更重要的事情；而在另一些情况下，AI 则令人感到压抑，或是被雇主强加于身。
调查结果提供了初步证据，表明观察到的接触程度（我们衡量 AI 替代风险的指标）与围绕 AI 的经济担忧相关。在接触程度高的职业（根据观察到的 Claude 执行的任务来定义）中，人们对经济替代更为焦虑。这与人们普遍意识到 AI 的扩散及其潜在影响是一致的。我们将在下文详细阐述我们的发现。
谁在担忧岗位替代？
“嗯，就像如今任何一个拥有白领工作的人一样，我百分之百担心，几乎每时每刻都在担心自己最终会被 AI 抢走工作。”——软件工程师。
在我们的调查中，五分之一的受访者对经济性替代表示担忧。有些人对此表达了抽象层面的忧虑：一位软件开发人员警告说，“当前状态下的人工智能存在被用于取代初级岗位的可能性。”另一些人则感叹他们的工作，或工作中的某些部分，正在被自动化取代。一位市场研究员表示：“在提升我的能力方面，这毫无疑问。[但]未来人工智能可能会取代我的工作。”在某些工作中，人们觉得AI让工作变得更难了。一位软件开发人员观察到，“当AI出现后，项目经理开始分配越来越难的任务和漏洞需要解决。”
在本报告中，我们使用基于Claude的分类器，从受访者的回答中推断其属性和观点。例如，许多参与者在谈话中会提及自己的行业，或提供有关其工作生活的信息性细节，这使我们能够推断出他们的职业。同样，我们通过提示Claude识别并解读受访者明确表示自身岗位面临AI驱动替代风险的直接引述，来量化对失业的担忧。我们在附录中给出了示例提示词。
受访者感知到的AI威胁与我们自己测量的实际暴露度指标相关，该指标反映了某项工作中使用Claude完成的任务所占百分比。当受访者的实际暴露度测量值较高时，他们对AI的担忧也更大。例如，小学教师对自己被替代的担忧程度低于软件工程师，这与Claude的使用偏向编码任务这一事实相符。
我们在下面的图1中展示了这一点。纵轴是特定职业中表示AI已经或可能很快取代其岗位的受访者百分比。横轴是实际暴露度。该图显示，平均而言，处于暴露度较高职业的人更倾向于表达对工作被自动化取代的担忧。暴露度每增加10个百分点，感知到的职业威胁就增加1.3个百分点。暴露度处于前25%的人群提及这种担忧的频率是后25%人群的三倍。
另一个重要的劳动者特征是职业阶段。在先前的研究中，我们报告了美国近期毕业生和早期职业工作者招聘放缓的初步迹象。在本调查中，约半数受访者我们能够从其回答中推断出职业阶段。² 我们发现，早期职业工作者对岗位被取代的担忧程度远高于资深工作者。
谁从 AI 中受益？
我们使用 Claude 对调查回答进行评估，将受访者自我报告的 AI 生产力提升程度按 1 至 7 分进行评分，其中 1 分代表“效率降低”，2 分代表“无变化”，后续每个等级代表更大的提升。获得 7 分的回答包含这样的评价：“以前需要几个月才能建好的网站，我（用 AI）4 到 5 天就做完了”；Claude 对“原本可能需要四个小时的工作，现在一半时间就完成了”这样的表述给出了 5 分，而对“就我个人而言，AI 帮我修复了网站上的代码。但反复尝试了好几次才得到我想要的结果”这样的表述给出了 2 分。³
总体而言，受访者报告了有意义的平均生产力提升。平均生产力评分为 5.1，对应“效率显著提升”。当然，我们的受访者是活跃的 Claude 用户，并且愿意参与调查。这可能使他们比普通用户更倾向于报告生产力提升。约 3% 的受访者报告了负面或中性影响，42% 的受访者未就生产力方面给出明确反馈。
这种差异在一定程度上与收入水平相关。图3左侧面板显示，高薪职业人群（如软件开发者）从AI中获得的生产力提升最为显著。这一结果并非仅由编程领域驱动；即便排除计算机与数学相关职业，结论依然成立。这印证了此前经济指数报告中的发现——同样倾向于高薪工作者：在需要更高教育水平的任务中，Claude往往能更大比例地缩短任务完成时间（相较于不使用AI的情况）。
部分低收入工作者也描述了较高的生产力提升。其中包括一位客服代表，他利用“AI根据现有回复快速生成新回复，节省了大量时间”。在某些情况下，低薪岗位人员还将AI应用于技术类副业项目。例如，一名送货司机正使用Claude开展电商业务，而一位园林设计师则在构建音乐应用程序。
我们在图3右侧面板中对此进行了更详细的分析，展示了按主要职业类别划分的推断生产力增益。排名最高的是管理类职业。这些受访者大多是使用Claude创业的企业家。4 排名第二的是计算机与数学类职业，包括软件开发者。生产力提升幅度最小的两个群体是科学和法律领域的从业者。部分律师对AI遵循精确指令的能力表示担忧。例如：“我给出了非常具体的规则，说明什么内容放在哪里、如何解读法律文件、我希望它做什么……但它每次都会偏离指令。”
随着人工智能在经济中扩散，一个关键问题是其收益将流向何处——是流向劳动者、他们的管理者、消费者，还是企业。受访者在大约四分之一的访谈中指出了这些收益的获得者。总体而言，这些人中的大多数提到自己受益，表现为任务完成更快、工作范围扩大以及时间得到释放。但在指明受益方的受访者中，有10%表示雇主或客户正在要求并获得了更多的工作量。较小比例的受访者提到了人工智能公司受益，而更小比例的人则认为人工智能将是净负面影响。这取决于职业阶段：只有60%的早期职业工作者表示他们个人从人工智能中受益，而资深专业人士的这一比例为80%。
范围与速度
受访者还分享了他们在哪些方面获得了生产力提升。我们将其分为范围、速度、质量和成本。例如，许多使用人工智能进行编码任务的人会说这样的话：“我不是技术背景，但现在我成了一名全栈开发者。”这是工作范围的扩展；人工智能为他们解锁了新的能力。相比之下，一些用户加快了他们已经从事的任务的速度，比如那位会计师所说：“我构建了一个工具，帮助我在15分钟内完成一项过去需要2小时的融资任务。”质量提升通常来自于对代码、合同及其他文件进行更彻底的检查。还有一小部分受访者提到了使用人工智能的低成本：“如果我雇一个社交媒体经理，那会超出我的预算。”
我们发现最常见的生产力提升体现在范围上，在明确提及生产力影响的用户中，有48%提到了这一点。在提及生产力的用户中，40%强调了速度。
人们与 Claude 的互动体验也可能影响他们对 AI 的担忧程度。为了评估这一点，我们测量了受访者报告的工作提速情况，具体方法是提取他们的工作是否变得慢得多（我们将其编码为 1）、速度没有变化（编码为 4）、还是变得快得多（编码为 7）。
我们发现，工作提速与感知到的职业威胁之间的关系呈 U 型曲线（见图 6）。最左侧的柱状图代表那些表示 AI 拖慢了其工作速度的受访者。这些受访者更倾向于认为 AI 对他们的生计构成了重大威胁。例如，一些创意工作者，如纯艺术家和作家，发现 AI 过于僵化和死板，无法帮助他们完成自己的工作。与此同时，他们担心 AI 向创意领域的扩散会使他们更难找到工作。
对于其余受访者，感知到的职业威胁随着其回答所隐含的提速水平提高而持续增加。这在经济学上具有一定道理：如果完成某项任务所需的时间正在迅速缩短，那么该岗位未来的存续能力可能会面临更多不确定性。
Key findings:
- Our recent survey of 81,000 Claude users shows that people who work in roles that are more exposed to AI have more concerns about AI-driven job displacement. These concerns are also higher among early-career respondents.
- Those in the highest- and lowest-paid occupations report the largest productivity gains, most commonly from increases in scope (doing new tasks).
- Respondents experiencing the largest speedups from AI express higher concern about job displacement.
In order to inform the public about the economic changes we’re observing with AI, our Economic Index shares what work Claude is being asked to do, and in which jobs Claude is doing the largest share of tasks. To date, however, we’ve lacked information on how these usage patterns map onto people’s thoughts and impressions of AI.
Our recent survey study with 81,000 Claude users provides a way to connect people’s economic concerns with what we’ve quantified in Claude traffic.
The survey asked people about their visions and fears around advances in AI. Many of the thoughts that people shared touched on economic topics. We learned that many people fear job displacement—though they also feel more productive and empowered at work. In some cases, AI has enabled them to start businesses, or given them time for more important things; in others, AI feels stifling, or imposed on them by their employers.
The survey’s results provide initial evidence that observed exposure (our measure of AI displacement risk) is correlated with economic concern around AI. People in highly exposed occupations—as defined by the tasks Claude is observed performing—were more nervous about economic displacement. This is consistent with people being broadly aware of AI’s diffusion and potential impacts. We expand on our findings below.
Who worries about job displacement?
“Well like anyone who has a white collar job these days I'm 100% concerned, pretty much 24/7 concerned about losing my job eventually to A.I.”—Software engineer.1
One fifth of the respondents in our survey voiced concern about economic displacement. Some worried about this in the abstract: one software developer cautioned about “the possibility of AI in its current state being used to replace junior positions.” Others lamented that their jobs, or aspects of their jobs, were being automated away. One market researcher said, “In terms of improving my capability, it's no doubt. [B]ut in the future AI may replace my work.” In some jobs, people felt it made their work harder. One software developer observed that “when AI arrived, the project managers started giving harder and harder tickets and bugs to solve.”
Throughout this report, we use Claude-powered classifiers to infer people’s attributes and sentiments from their responses. For example, many participants mention their line of work in passing or give informative details about their work life, which allows us to infer their occupation. Similarly, we quantify concerns about job loss by prompting Claude to identify and interpret direct quotes in which respondents indicate that their own role is at risk of AI-driven displacement. We give example prompts in the Appendix.
Respondents’ perceived threat from AI was correlated with our own measure of observed exposure, which reflects the percentage of a job’s tasks for which Claude is used. A respondent was more concerned about AI when our observed exposure measure for that respondent was higher. Elementary school teachers were less worried about their own displacement than software engineers, for example, consistent with the fact that Claude usage skews toward coding tasks.
We show this in Figure 1 below. The y-axis is the percentage of respondents in a given occupation who said that AI is already replacing their role or is likely to do so soon. The x-axis is observed exposure. The plot shows that, on average, people in more exposed occupations tended to express more concern about their jobs being automated away. For every 10-percentage-point increase in exposure, perceived job threat increased by 1.3 percentage points. People in the top 25% of exposure mentioned the worry three times as often as those in the bottom 25%.
Another important worker characteristic is career stage. In previous research, we reported tentative signs of a slowdown in the hiring of recent graduates and early-career workers in the United States. For about half of respondents in this survey, we were able to infer career stage from their answers.2 We found that early-career respondents were much more likely to express concern about job displacement than senior workers.
Who benefits from AI?
Using Claude to assess the survey responses, we rated the extent of people’s self-reported productivity gains from AI on a 1–7 scale, where 1 is “less productive,” 2 is “no change,” and each subsequent level denotes a larger gain. Responses that scored 7 included testimonials like, “It used to take months to make the website I [made] in 4-5 days”; Claude gave a 5 to statements like, “What might have taken four hours was accomplished in half the time,” and a 2 to ones like, “Personally, I had AI help me fix code on a website. But it took multiple passes to get the result I was after.”3
Overall, people reported meaningful productivity gains on average. The mean productivity rating was 5.1, corresponding to “substantially more productive.” Our respondents were, of course, active Claude users who were willing to take a survey. This could make them more likely to report productivity benefits than the average user. Some 3% reported negative or neutral impacts, and 42% did not give a clear indication on productivity.
This splits somewhat across income lines. The left panel in Figure 3 shows that people in high-paying jobs, like software developers, conveyed the largest productivity gains from AI. This result is not driven only by coding; it holds when we leave out computer and math occupations. It echoes a previous Economic Index finding that also favored higher-paid workers: in tasks requiring greater levels of education, Claude tended to reduce the time taken to complete a task (relative to doing it without AI) by a higher percentage.
Some of the lowest-paid workers describe high productivity gains as well. This included a customer service representative using “AI to save me a lot of time with creating a response based on another one.” And in some cases, people in low-wage jobs were using AI on technical side projects. One delivery driver, for example, was using Claude to start an e-commerce business, and a landscaper was building a music application.
We look at this in more detail in the right panel of Figure 3, showing the inferred productivity gain by major occupational group. At the top are management occupations. These respondents are mostly entrepreneurs using Claude to build a business.4 The next highest category is computer and math, which includes software developers. The two groups exhibiting the mildest productivity improvements were workers in scientific and legal professions. Some lawyers worried about AI’s ability to follow precise instructions. For example: “I have given very specific rules about what is where, how to read a legal document, what I want it to do… but it diverges every time.”
A key question as AI diffuses through the economy is where the benefits will accrue—to workers, their managers, consumers, or corporations. Respondents indicated the recipient of these gains in about a quarter of interviews. Overall, most of these people cited benefits to themselves, through faster tasks, expanded scope, and freed-up time.5 But 10% of respondents who named a recipient said that employers or clients were asking for and getting more work. A smaller share mentioned benefits to AI companies, and an even smaller share said that AI would be a net negative. This depended on career stage: only 60% of early-career workers indicated that they personally benefited from AI, compared to 80% of senior professionals.
Scope and speed
Respondents also shared where they experienced gains in productivity. We separate this into scope, speed, quality, and cost. For example, many people using AI for coding tasks said things like, “I’m a non tech guy but now I’m a full stack developer.” This is an expansion of scope; AI unlocks new abilities for them. In contrast, some users sped up tasks they were already doing, like the accountant who said, “I built a tool that helps me finish a financing task in 15 minutes that used to take 2 hours.” Quality gains often came from more thorough checks of code, contracts, and other paperwork. And a small share of respondents mentioned the low cost of using AI: “[I]f I hire a social media manager it’s over my budget.”
We find that the most common productivity enhancement is in scope, which was cited by 48% of users who explicitly mentioned productivity effects. 40% of users who mentioned productivity emphasized speed.
People’s experience with Claude might also shape their concerns about AI. To assess this, we measured the speedup reported by respondents, by extracting whether their work was now much slower (which we coded as 1), showed no change in speed (4), or had become much faster (7).
We found that the relationship between speedup and perceived job threat is U-shaped (see Figure 6). The leftmost bar shows respondents who reported that AI slowed them down. These respondents were more likely to indicate that AI posed a significant threat to their livelihoods. For example, some creative workers, like fine artists and writers, found AI too stifling and rigid to help them at their own work. At the same time, they feared the diffusion of AI into creative fields would make it harder for them to find work.
For the remaining respondents, perceived job threat increases consistently with the level of speedup implied by their answers. This makes some economic sense: if the time required to do one’s tasks is shrinking quickly, there may be more uncertainty about the future viability of the role.