摘要
- 我们公布了一项针对 1260 名社会科学家的调查结果，该调查涉及人工智能与编程智能体的使用情况，于 2026 年 2 月和 3 月进行。
- 绝大多数受访者（81%）曾在研究中使用过 AI 聊天机器人，主要用于编写代码和润色文章。但只有 20% 的人将编程智能体——例如 Claude Code 这类能自主编写并执行分析代码的工具——应用到工作中。
- 编程智能体的使用存在显著差异。使用编程智能体的研究人员中，拥有典型男性名字的人数是拥有女性名字人数的两倍。顶尖大学的研究人员使用编程智能体的可能性比其他研究人员高出 40%。
- 编程智能体的使用者比同一学科和职业阶段的其他研究人员发布了更多的工作论文和基金申请，但这可能反映了早期采用者之间已有的差异。
- 研究人员对 AI 帮助撰写可发表论文的乐观程度，高于对 AI 对整个社会科学领域影响的乐观程度。
AI 编程智能体正在如何改变我们研究经济和社会的方式？
人文科学正在发生转变：核心研究任务首次可以交由机器完成。AI 聊天机器人越来越多地助力科学研究，包括在最负盛名的出版物和社会科学领域。这激发了人们对 AI 可能提升研究生产力的乐观情绪，同时也引发了关于同行评审负担过重以及学术 AI 垃圾内容泛滥的担忧。
然而，虽然轮流对话式 AI 聊天机器人主要用于写作辅助，但编程智能体可能更彻底地重构社会科学研究。像 Claude Code 和 Codex 这样的智能体编程平台可以接收研究思路和数据集，编写并运行分析程序，解读输出结果，并自主进行迭代。经验研究中那些曾经不可简化的人类步骤，首次得以实现自动化。在极端情况下，研究人员已经构建了多智能体流水线，用于自动化计算机科学研究并自主执行社会科学研究思路。
这些工具能够加速科学进步，并使其更具大胆探索精神：快速的研究执行意味着廉价且丰富的发现。它们也可能加剧研究资源的不平等，并加大学术记录的拥堵程度。更深远地看，随着人工智能处理越来越多的研究任务，其独特的分析选择可能会塑造我们对经济、社会以及人类自身的集体认知。
在这篇文章中，我们基于2026年初对1260名定量社会科学研究者进行的一项调查，提供初步观察。该调查是一项更大规模、持续进行的研究的基线轮次，该研究旨在探究编码智能体如何影响研究生产力，其中包含一项随机实验，为研究人员提供Claude Code的使用权限。我们将在未来公布这项实验的结果。目前，我们报告基线调查所揭示的内容：谁在使用这些工具、用于何种目的；使用者和非使用者之间的产出差异；以及研究人员对日益普及所带来的影响有何预期。
一项关于定量社会科学研究者中AI编码智能体使用情况的新调查
我们于2026年2月底至3月开展了这项调查，目标人群为活跃的定量社会科学研究者。这并非一个具有代表性的样本——受访者是被招募来参与一项提供Claude Max账户使用权限的研究，因此样本的选择可能偏向于对AI工具好奇的研究人员。然而，受访者与之前收到更通用邀请函的样本群体相当相似（见附录表A2）。
受访者在经济学、政治学和社会学三个领域分布均匀，各约占样本的五分之一，管理科学和心理学紧随其后（见表A1）。我们还收到了来自公共卫生、教育和传播学领域研究人员的少量回复。大约40%的受访者是正教授或副教授，25%是助理教授，约30%是博士生。
编码智能体尚未覆盖大多数社会科学家
我们通过两种方式衡量了AI的总体使用情况。首先，我们询问：“您之前是否使用过生成式AI模型来辅助您的研究过程？”81%的受访者回答“是”。
但那些真正将日益强大的编码智能体融入工作流程的人又如何呢？我们询问了“你是否经常（每周超过一次）使用集成到命令行中的 AI 编码助手（例如 Codex、Cursor 或 Claude Code）？”在后续问题中，我们确认了他们确实使用了这些工具之一（或 Google Antigravity）。
只有 20% 的受访者使用编码智能体。我们的调查是在 2025 年 12 月底引发关于 Claude Code 和 Opus 4.6 的广泛讨论大约两个月后进行的。然而，即使在主动选择参与我们调查的感兴趣受访者中，也只有五分之一的人将智能体融入了工作流程。Claude Code 是最常被报告的编码智能体工具，86% 的用户表示使用 Claude Code（31% 的用户表示使用 Codex，这是第二常见的工具）。
采用率高度不均衡
图 1 显示总体采用率存在巨大差异，从经济学家的 39% 和政治科学家的 25%，到公共卫生（6%）、教育学（4%）和传播学（6%）的个位数。这种梯度大致反映了不同领域在整体 AI 使用上的差异，但编码智能体采用率的差异平均而言更为显著。
略超过四分之一的博士生和博士后至少每周使用编码智能体；在终身教授中，这一比例下降了一半以上。采用编码智能体的研究人员是资历较浅者——他们技术更娴熟，更可能直接处理代码和数据，并且面临更强的职业压力以产出研究成果。
采用差异不仅体现在学科和职业阶段上。我们根据姓名对研究人员进行性别分类，发现拥有典型男性姓名的受访者采用编程智能体的比例是拥有典型女性姓名的受访者的两倍以上。高地位大学和私立大学的采用率也显著更高。所有这些差异在 p<0.05 水平上具有统计显著性。这些差异比整体 AI 使用中的差异更为显著，表明至少在编程智能体采用的早期阶段，不平等程度更高。
编程智能体使用中的性别差异不仅仅反映了尝试 AI 的比率差距。在尝试过将 AI 用于研究的受访者中，经常使用编程智能体的性别差异甚至比整体样本中的差异略大。即使在相同学科和职业阶段内进行跨性别比较时，这些差异依然存在。
研究人员主要使用 AI 进行编码和编辑，而非写作
在使用 AI（无论是通过编程智能体还是聊天机器人）的研究人员中，他们实际用 AI 来做什么？关于 AI 在学术研究中应用的争论主要集中在写作上：幻觉生成的文献综述、公式化引言中充斥的“不是 X，而是 Y”句式，以及完全自动化论文写作的可能性。
但图4显示，无论是编程智能体用户还是其他用户，最常见的用途都是对定量数据进行编码分析：97%的编程智能体用户和77%的其他AI用户表示会用它来生成代码。其次常见的是编辑文稿，然后是寻求方法建议和了解先前研究的背景。综合编程智能体用户和其他用户来看，所有AI用户中只有三分之一曾用它起草过文稿。这些模式在不同学科中基本一致，只有经济学和管理学研究者普遍使用AI来起草文稿。
编程智能体用户发布了更多工作论文，提交了更多资助申请，但向期刊的投稿数量并未增加
编程智能体是否让研究人员更高效？这正是本调查所启动的更广泛研究要回答的问题。我们针对这一问题开展的实验仍在进行中。但基线调查让我们能够在研究过程中的多个节点上，将编程智能体用户与其他用户进行比较。这种比较纯粹是描述性的：我们将选择使用编程智能体的研究人员与未使用的研究人员进行对比，并预期这两组人在我们无法调整的多个方面存在差异。这些差异不应被解读为因果关系，而应被视为使用编程智能体的研究人员与未使用的研究人员之间的初步比较。
图 5 展示了调查前六个月中，从项目启动到论文提交等不同研究阶段的自报产出。经调整后的估计值将编码智能体使用者与其他研究者进行了比较，并控制了职业阶段、学科领域以及完成调查的周次。编码智能体使用者启动了更多项目，发布了更多工作论文，提交了更多基金申请，并且可能也提交了更多会议论文。
那么，编码智能体使用者是否撰写了更多论文？首先，考虑编码智能体使用者与其他研究者在早期流程产出上的差异。与非智能体使用者相比，编码智能体使用者启动项目的速度大约多出四分之一篇论文，发布工作论文的速度大约多出半篇。从百分比来看，编码智能体使用者在同一学科和职业阶段中，其产出效率比其他研究者高出约 10%（启动的实证项目）到 75%（发布的工作论文）。
然而，这种产出差异仅体现在这些早期流程指标上。我们没有发现证据表明编码智能体使用者向期刊提交了更多新论文，或者能更快地重新提交论文。这可能反映了论文从撰写到提交的时间周期，因为编码智能体的使用是近期才出现的现象。但也可能表明，编码智能体在启动和运行项目方面比在完善论文以提交期刊的“最后一公里”上更有用。
研究人员期望 AI 工具能提高生产力，但不太确信它们能整体改善社会科学。
我们还询问了研究人员对 AI 工具的期望。AI 是否能让社会科学家在撰写可发表论文方面更具生产力？以及他们认为 AI 会让社会科学变得更好还是更差？
研究人员对 AI 提高论文写作生产力持乐观态度。在 1 到 10 分的评分中，88% 的受访者评分高于 5 分，一半的受访者评分达到 8 分或以上。图 6 显示，这些评分与 AI 使用情况密切相关。图左侧显示，在更多类型任务中使用 AI 的研究人员更为乐观。图右侧显示，编码智能体使用者比其他研究者更为乐观。
该调查的受访者是对尝试这些工具感兴趣的人群，因此看到一些对生产力的乐观态度并不意外。但即便在这些乐观者中，关于 AI 在狭义上帮助撰写可发表论文与广义上影响社会科学之间，仍存在显著差距。70% 的受访者对论文生产力的乐观程度高于对更广泛领域影响的乐观程度。很少有研究人员对领域影响的乐观程度高于对论文生产力的乐观程度，而许多人对领域影响更为悲观。
Summary
- We present results from a survey of 1,260 social scientists about AI and coding agent use, fielded in February and March 2026.
- The vast majority of respondents (81%) have tried using AI chatbots in research, particularly for writing code and editing prose. But only 20% have adopted coding agents—tools like Claude Code that autonomously write and execute analysis code—into their work.
- There are sharp disparities in use of coding agents. Twice as many researchers with typically male names use coding agents as those with female names. Researchers at top universities are 40% more likely than others to use coding agents.
- Users of coding agents post more working papers and grant proposals than others in the same discipline and career stage, but this could reflect pre-existing differences among early adopters.
- Researchers are more optimistic about AI helping write publishable papers than about the effects of AI on the social sciences as a whole.
How are AI coding agents changing how we study the economy and society?
The human sciences are shifting: for the first time, core research tasks can be handed off to machines. AI chatbots increasingly contribute to scientific research, including in the most prestigious publications and in the social sciences. This has spurred optimism that AI could boost research productivity—while also stoking fears about overloaded peer review and a deluge of academic AI slop.
But while turn-taking AI chatbots have primarily been used for writing assistance, coding agents could restructure social science research more radically. Agentic coding platforms like Claude Code and Codex can take a research idea and a dataset, write and run an analysis, interpret the output, and iterate autonomously. What had been irreducibly human steps in empirical research can, for the first time, be automated. At the extreme, researchers have built multi-agent pipelines to automate computer science research and autonomously execute social science research ideas.
These tools could accelerate science and make it more daring: fast research execution should mean cheap and plentiful discovery. They could also amplify disparities in research resources and exacerbate congestion in the scholarly record. More deeply, as AI handles a broadening swath of research tasks, its distinctive analytical choices could stamp our collective understanding of our economy, our society, and ourselves.
In this post, we offer a first look, drawing on a survey of 1,260 quantitative social scientists fielded in early 2026. The survey is the baseline wave of a larger ongoing study of how coding agents affect research productivity, including a randomized experiment providing researchers with access to Claude Code. We will publish results from this experiment in the future. For now, we report what the baseline survey reveals about who is using these tools and for what; how output differs between users and non-users; and what researchers expect about the implications of growing adoption.
A new survey on AI coding agent use among quantitative social scientists
We fielded the survey in late February and March 2026, targeting active quantitative social scientists. This was not a representative sample—respondents were recruited for a study that offered access to Claude Max accounts, so selection into the sample could tilt toward researchers curious about AI tools. However, the respondents were fairly similar to an earlier sample that received a more generic invitation (see Table A2 in the Appendix).
Respondents were evenly split between economics, political science and sociology, each around a fifth of the sample, with management sciences and psychology close behind (see Table A1). We also received a smaller number of responses from public health, education and communications researchers. Roughly 40% were full or associate professors, 25% were assistant professors, and about 30% were doctoral students.
Coding agents haven't reached most social scientists
We measured overall AI use in two ways. First, we asked “Have you previously used genAI models to aid your research process?” 81% of respondents said yes.
But what about those who have actually adopted increasingly capable coding agents into their workflow? Here, we asked “Do you regularly (more than once a week) use an AI coding assistant integrated into your command line (such as Codex, Cursor, or Claude Code)?” In a follow-up question, we verified that they used one of those tools (or Google Antigravity).1
Only 20% of respondents use coding agents. Our survey came around two months after a flurry of discussion about Claude Code and Opus 4.6 that kicked off in late December of 2025. Yet even among interested respondents who self-selected into our survey, only ⅕ had adopted agents into their workflow. Claude Code is the most common coding agent tool reported, with 86% of users reporting Claude Code use (31% report using Codex, the next most common tool).
Adoption is highly uneven
Figure 1 shows there is large variation in the overall adoption rate, from 39% of economists and 25% of political scientists to single digits for public health (6%), education (4%) and communication (6%). This gradient roughly tracks differences across fields in overall AI use, but differences in coding agent adoption are steeper on average.
Just over a quarter of doctoral students and postdocs use coding agents at least weekly; among tenured professors that rate falls by more than half. The researchers adopting coding agents are the juniors—more technologically fluent, more likely to be working directly with code and data, and facing stronger career pressures to produce research.
Adoption differences extend beyond discipline and career stage. We classify researcher names according to gender and find that those with typically male names have adopted coding agents at more than twice the rate of respondents with typically female names. High-status and private universities also see notably higher use. All of these differences are significant at the p<0.05 level. These differences are starker than the differences in overall AI use, and suggest higher inequality, at least in this early period of coding agent adoption.
The gender gap in coding agent use does not just reflect a gap in rates of trying AI. Among respondents who have tried using AI for research, there is even a slightly larger gender gap in regular coding agent use than in the overall sample. These differences also persist when comparing across genders in the same disciplines and career stages.
Researchers mainly use AI to code and edit, not write
Among researchers using AI, whether through coding agents or chatbots, what are they actually using it for? Debate about AI in academic research has focused heavily on writing: hallucinated literature reviews, “it’s not X, it’s Y” strewn across formulaic introductions, and the possibility of fully automated paper writing.
But Figure 4 shows that the most common use, for both coding agent users and others, is for coding up analysis of quantitative data: 97% of coding agent users and 77% of other AI users report using it to generate code. Next most common is editing prose, followed by asking for methods advice and background on prior research. Aggregating across coding agent users and others, only a third of all AI users have used it to draft prose at all. These patterns generally hold across disciplines, with only economists and management researchers commonly using AI to draft prose.
Coding agent users are posting more working papers and sending out more grant proposals, but not submitting more to journals
Are coding agents making researchers more productive? That’s the question motivating the broader study this survey kicks off. The experiment we are running on this question is still ongoing. But the baseline survey lets us compare coding agent users to others across a whole bunch of checkpoints in the research process. This comparison is purely descriptive: we compare researchers who select into coding agent use to those who do not, and expect that the two groups differ in a number of ways that we cannot adjust for. Differences should not be interpreted as causal, but as a first cut comparison between researchers using coding agents and those who are not.
Figure 5 shows self-reported output over the six months before the survey at different stages of the research process, from projects started to papers submitted. The adjusted estimates compare coding agent users to others, controlling for career stage, discipline, and the week they completed the survey. Coding agent users are starting more projects, posting more working papers, submitting more grants, and possibly sending out more conference submissions.
So are coding agent users writing more papers? First, consider the differences in early pipeline output between coding agent users and others. Coding agent users are starting projects at a pace of around a quarter of a paper more and posting around a half of a working paper more than non agent users. In percentage terms, coding agent users look around 10% (empirical projects started) to 75% (working papers posted) more productive than others in their discipline and career stage.
However, this productivity difference only appears for these early pipeline measures. We find no evidence that coding agent users are submitting more new papers to journals or resubmitting papers more quickly. This could reflect the timeline of getting a paper to submission, as coding agent use is a recent phenomenon. But it could also reflect that coding agents are more useful at getting projects up and running than they are at the last mile of perfecting a paper for journal submission.
Researchers expect AI tools to raise productivity, but are less confident that they will improve social science overall
We also asked researchers what they expected of AI tools. Does AI make social scientists more productive, in terms of writing publishable papers? And do they think AI will make the social sciences better or worse?
Researchers are optimistic about AI raising paper-writing productivity. On a 1 to 10 scale, 88% of respondents were above a 5, and half were at 8 or above. Figure 6 shows that these ratings vary strongly with AI use. The left side of the plot shows researchers that use AI for more types of tasks are more optimistic. The right side shows coding agent users are more optimistic than others.
The survey is drawing from people who are interested in trying these tools out, so it should not be surprising to see some optimism about productivity. But even among these optimists, there is a real gap between views about AI helping narrowly with publishable papers and broadly affecting the social sciences. 70% of respondents are more optimistic about paper productivity than about broader field impact. There are few researchers more optimistic about field impacts than about paper productivity, and many who are more pessimistic.