# 控制论队友：关于生成式人工智能与团队协同的实地实验 # The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork **发表信息 / Publication Information:** 本研究发表于国际管理学与组织理论领域的顶级学术期刊《组织科学》（*Organization Science*，2026年第37卷第4期）。 This study was published in *Organization Science* (Vol. 37, No. 4, 2026), a premier peer-reviewed academic journal in management and organizational theory. **作者团队 / Authors:** * Fabrizio Dell’Acqua (哈佛商学院 / 帝国理工学院 / 哈佛AI研究所) * Charles Ayoubi (法国高等经济商业学院 ESSEC) * Hila Lifshitz (沃里克商学院 / 哈佛AI研究所) * Raffaella Sadun (哈佛商学院 / 哈佛AI研究所) * Ethan Mollick (宾夕法尼亚大学沃顿商学院) * Lilach Mollick (宾夕法尼亚大学沃顿商学院) * Yi Han (宝洁公司 Procter & Gamble) * Jeff Goldman (宝洁公司 Procter & Gamble) * Hari Nair (宝洁公司 Procter & Gamble) * Stew Taub (宝洁公司 Procter & Gamble) * Karim R. Lakhani (哈佛商学院 / 哈佛AI研究所) --- ## 摘要 / Abstract We examine how artificial intelligence (AI) impacts three core pillars of collaboration—performance enhancement, expertise integration, and social engagement—through a preregistered field experiment with 791 professionals at Procter & Gamble, a global consumer packaged goods company. Working on real product innovation challenges, professionals were randomly assigned to work either with or without AI, and either individually or with another professional in new product development teams. Our findings show that (1) AI significantly enhances performance: individuals with AI matched the performance of teams without AI, suggesting that AI can effectively replicate certain benefits of human collaboration. Moreover, (2) AI helps bridge functional silos: without AI, research and development professionals tended to suggest more technical solutions, whereas commercial professionals leaned toward commercially oriented proposals. Professionals using AI produced more balanced solutions, regardless of their professional background. (3) AI’s language-based interface prompted more positive self-reported emotional responses among participants, suggesting it can fulfill part of the social and motivational role traditionally offered by human teammates. Finally, decomposing the innovation process suggests that AI primarily enhances the quality of generated ideas, shifting the distribution of creative output upward, whereas human judgment retains value in evaluative selection. This finding highlights the multiple and complementary roles that human and AI partners can play in new product development tasks and creative problem solving. More generally, our results suggest that AI adoption in knowledge work affects not only performance but also how expertise and sociality appear within teams, offering insights into the impact of generative AI on collaborative work within organizations. 我们通过在宝洁公司（一家全球消费品巨头）791名专业人员中开展的一项预注册实地实验，探讨了人工智能（AI）如何影响协同合作的三大核心支柱——绩效提升、专长整合和社交互动。在应对真实的产品创新挑战时，专业人员被随机分配在有AI或没有AI的辅助下工作，并以个人或与另一名专业人员组成新产品开发团队的形式开展工作。我们的研究结果表明： 1. **AI显著提升了绩效**：使用AI的个人在产出质量上与未使用AI的团队相当，这表明AI可以有效替代人类协同合作的部分绩效优势。 2. **AI有助于跨越职能壁垒**：在没有AI的情况下，研发（R&D）人员倾向于提出更偏技术性的方案，而商业/市场端人员则更偏向商业导向的提案。然而，使用AI的专业人员不论其专业背景如何，都能提出更为平衡的解决方案。 3. **AI的语言交互界面激发了参与者更积极的自我情感反馈**：这表明AI能在一定程度上发挥传统上由人类队友提供的社交和激励作用。最后，通过对创新过程的解构，我们发现AI主要提升了“创意生成”阶段的质量，将整体创意产出的分布向上推移，而人类的判断力在“创意筛选”阶段依然保有独特价值。这一发现强调了在产品开发与创造性问题解决中，人类与AI伙伴可以扮演多元且互补的角色。总体而言，研究表明在知识型工作中引入AI，不仅会提升绩效，还会重塑团队中专长整合与社交互动的模式。 --- ## 1. 引言 / Introduction Teamwork is central to modern organizations. Whether designing a new product, solving strategic challenges, or supporting impactful innovation, human collaboration has often been associated with higher-quality results than individuals working alone. There are three main pillars commonly used to justify teamwork. The first is performance: teamwork is often more effective than individual work and allows for more complex problems to be tackled. The second is expertise sharing and knowledge complementarities: teamwork allows people with different expertise to come together and work on the same problem effectively. Finally, there is human sociality: people often enjoy connecting with other people, which can increase their motivation to work. In this paper, we investigate how the introduction of artificial intelligence (AI) impacts these three pillars of teamwork in the context of a field experiment in a large organization. 团队合作是现代组织的核心。无论是设计新产品、解决战略难题，还是推动重大的创新，人们往往认为团队协同能比个人单打独斗带来更高质量的成果。支持团队合作主要有三大支柱：第一是**绩效（Performance）**，团队合作通常比个人工作更有效，能够解决更复杂的问题；第二是**专长共享与知识互补（Expertise Sharing）**，团队合作让不同领域的专家能够凝聚智慧，共同攻克同一难题；第三是**人类的社交性（Sociality）**，人们通常喜欢与他人建立连接，这能显著提高工作的积极性与组织认同感。在本文中，我们通过在大型企业中进行的一项实地实验，研究了人工智能（AI）的引入如何影响团队合作的这三大支柱。 > **背景补充 / Background Context:** > > 本文中所使用的“**控制论（Cybernetics）**”概念，最早由美国数学家诺伯特·维纳（Norbert Wiener）于1948年提出。控制论旨在研究机器与生物体内部的控制与通信机制。在现代工作场景中，“控制论队友（Cybernetic Teammate）”不仅是指一个被动的信息搜索工具或计算器，而是一个能够通过“提问-反馈-调整”的动态双向闭环与人类进行自适应协同的交互式系统。 --- ## 2. 理论背景 / Theoretical Background ### 2.1. 团队合作 / Teamwork A wide range of scholarship shows that collaboration can outperform individual effort in organizations by integrating multiple perspectives, thereby tackling complex problems more effectively. These performance advantages stem from the synergy that arises when team members share real-time feedback, pool different skill sets, and engage in collective problem solving. Beyond raw performance, specialized knowledge resides in individuals and must be integrated to solve complex problems. However, as the volume and sophistication of available knowledge grow (the "burden of knowledge"), deep individual specialization necessitates team-based coordination. Finally, teamwork provides critical social and motivational benefits, creating a supportive environment that fosters motivation and persistence. 大量学术研究表明，团队合作通过整合多维视角，能比个人努力表现出更好的组织绩效。这种绩效优势源自团队成员实时反馈、技能融合和集体解决问题时产生的协同效应。除了单纯的绩效外，专业知识本身分布在不同的个体中，解决复杂问题必须实现知识整合。然而，随着人类知识总量的增长（即“知识负荷”加重），个体的深度专业化导致跨领域的协调合作变得不可或缺。最后，团队合作还提供了关键的社交和情感激励，创造了相互支持的环境，从而维持了工作的持久动力。 ### 2.2. 生成式人工智能 / Generative AI Unlike previous technologies that primarily automated routine tasks, GenAI’s capacity to converse, reason, and create positions it less as a passive tool and more as an active cybernetic teammate. LLMs are trained on human language and act more like a person than a machine. GenAI can engage with tacit knowledge—the kind of implicit understanding that traditionally could only be shared through direct human interaction. 与以往主要将常规性、可编程任务自动化的传统软件技术不同，生成式AI（GenAI）具备对话、推理和创作的能力，这使其不再只是一个被动的应用工具，而更像是一个主动的控制论队友。大型语言模型（LLM）基于人类语言海量数据训练而成，其交互反馈的表现更接近人类而非机器。生成式AI可以处理“隐性知识（Tacit Knowledge）”——这类难以言传的隐性理解在过去通常只能通过人与人之间的直接互动来传递。 ### 2.3. 新产品开发 / New Product Development New product development (NPD) represents an ideal context for examining how AI affects collaborative work. NPD typically involves generating solutions under time pressure with limited resources, mirroring key features of an innovation tournament. The challenge is compounded by organizational tendencies toward local search and existing routines, which can constrain the exploration of novel solutions. 新产品开发（NPD）是检验AI如何影响协同工作的理想场景。新产品开发通常要求在时间紧迫、资源有限的情况下快速提出方案，这高度契合“创新竞赛”的特征。此外，组织往往存在安于现状（局部搜索）和路径依赖的倾向，这限制了对新奇解决方案的探索。 --- ## 3. 实验设计 / Experimental Design ### 3.1. 实证背景 / Empirical Setting We conducted a field experiment at Procter & Gamble (P&G). P&G, a large multinational firm with structured R&D processes and a skilled workforce, provides a useful setting. With roughly 7,000 R&D professionals worldwide, the firm encompasses end-to-end product development activities. Cross-functional teams composed of R&D and commercial representatives are the core units where innovation happens. 我们选择在宝洁公司（Procter & Gamble）开展实地实验。 > **背景补充 / Background Context:** > > 宝洁公司（P&G）创立于1837年，是全球最大的日用消费品巨头之一，旗下拥有帮宝适（Pampers）、吉列（Gillette）、佳洁士（Crest）等众多标志性品牌。宝洁以其严谨、系统化、多职能协同的创新流程闻名于世，拥有约7000名研发人员。在宝洁，跨职能协同（特别是研发专家与商业/市场人员的合作）是新产品开发和市场开拓的起点，因此它代表了研究知识型团队协同与决策的“黄金标准”实验室。 ### 3.2. 实验方法 / Experimental Approach The experiment was conducted as a one-day virtual product development workshop. Participants were randomly assigned to one of four conditions, in a 2 × 2 experimental design: 1. **Individual No AI** (Control / 对照组) 2. **Team No AI** (Treatment 1 / 实验组1: R&D + Commercial, No AI) 3. **Individual + AI** (Treatment 2 / 实验组2: Individual with GPT-4) 4. **Team + AI** (Treatment 3 / 实验组3: R&D + Commercial with GPT-4) 本实验设计为期一天的虚拟产品开发研讨会。791位宝洁资深员工被随机分配到2 × 2矩阵的四个组别中： 1. **个人无AI组**（对照组）：单兵作战，不使用AI。 2. **人类团队无AI组**（实验组1）：1名研发专家 + 1名商业专家配对，不使用AI。 3. **个人+AI组**（实验组2）：个人使用基于GPT-4（通过微软Azure云服务接入安全专线）的AI工具辅助。 4. **团队+AI组**（实验组3）：人类团队（研发+商业）在GPT-4的辅助下协同。 **Figure 1. Treatment Matrix / 图1. 实验干预矩阵** ![Figure 1](https://abxcyz.com/minio/abxcyz/20260801/c5a6e5c65ef547a599099b575311437f.gif) --- ### 表1. 样本描述性统计 / Table 1. Summary Statistics #### 个人级别样本特征 (Panel A: Individual) | 变量 / Variable | 个人无AI组 (Individual No AI) | 个人+AI组 (Individual + AI) | 均值差异 (Mean Diff.) | | :--- | :---: | :---: | :---: | | 女性比例 (Female) | 0.578 (0.494) | 0.555 (0.497) | -0.023 | | 男性比例 (Male) | 0.422 (0.494) | 0.432 (0.495) | 0.010 | | 职级带宽 (Band Level) | 2.071 (0.742) | 2.065 (0.762) | -0.006 | | 公司司龄 (Years of Experience) | 12.351 (8.293) | 11.816 (7.807) | -0.535 | | 研发人员比例 (R&D Specialist) | 0.604 (0.491) | 0.594 (0.493) | -0.010 | | 工作中ChatGPT使用频率 (1-5分) | 2.786 (1.126) | 2.735 (1.206) | -0.050 | | 个人生活ChatGPT使用频率 (1-5分) | 2.468 (1.200) | 2.529 (1.147) | 0.061 | | 工作中是否有AI使用权限 (0/1) | 0.812 (0.392) | 0.800 (0.401) | -0.012 | | 实验前对工作使用AI的期望值 (1-5分) | 3.539 (0.951) | 3.555 (1.027) | 0.016 | | **样本人数 (Observations)** | **154** | **155** | | #### 团队级别样本特征 (Panel B: Team) | 变量 / Variable | 人类团队无AI组 (Team No AI) | 团队+AI组 (Team + AI) | 均值差异 (Mean Diff.) | | :--- | :---: | :---: | :---: | | 女性比例 (Female) | 0.596 (0.492) | 0.556 (0.498) | -0.040 | | 男性比例 (Male) | 0.404 (0.492) | 0.444 (0.498) | 0.040 | | 职级带宽 (Band Level) | 2.000 (0.714) | 2.083 (0.734) | 0.083 | | 公司司龄 (Years of Experience) | 10.091 (7.616) | 10.476 (8.108) | 0.385 | | 研发人员比例 (R&D Specialist) | 0.500 (0.501) | 0.500 (0.501) | 0.000 | | 工作中ChatGPT使用频率 (1-5分) | 2.574 (1.225) | 2.615 (1.179) | 0.041 | | 个人生活ChatGPT使用频率 (1-5分) | 2.326 (1.056) | 2.480 (1.092) | 0.154 | | 工作中是否有AI使用权限 (0/1) | 0.713 (0.427) | 0.746 (0.384) | 0.033 | | 实验前对工作使用AI的期望值 (1-5分) | 3.430 (1.003) | 3.534 (1.021) | 0.103 | | **样本总人数 (Participants)** | **230 (115对团队)** | **252 (126对团队)** | | *(注：括号内为标准差 / Standard deviations in parentheses.)* --- ### 3.3. 收集的成果 / Collected Outcomes The solutions submitted by participants were assessed blindly by human expert evaluators on a 1-10 scale across dimensions including *Quality*, *Novelty*, *Feasibility*, and *Technicality*. We also measured self-reported emotional states (Positive and Negative Emotions) before and after the task. 参与者提交的商业/技术解决方案由多位外部行业专家在双盲（不知道作者及实验组别）的情况下，基于1-10分的量表在**质量（Quality）**、**新颖性（Novelty）**、**可行性（Feasibility）**和**技术性（Technicality）**等维度进行独立打分。此外，本研究也评估了参与者在任务前后的自我报告心理情绪变化（积极情绪与消极情绪）。 --- ## 4. 实验分析策略 / Empirical Strategy ### 4.1. 分析方法 / Analytical Approach We estimate the causal effect of AI adoption and team configuration using the following regression specification: 我们使用普通最小二乘法（OLS）回归模型来估计引入AI及团队模式对最终解决方案产生的因果效应： $$Y_i = \beta_0 + \beta_1 \text{TeamNoAI}_i + \beta_2 \text{Individual+AI}_i + \beta_3 \text{Team+AI}_i + \gamma \text{Controls}_i + \delta \text{FE}_i + \epsilon_i$$ Where $Y_i$ represents the outcome variables (Quality, Novelty, etc.), $\text{Controls}_i$ includes pre-experimental individual characteristics (experience, band level, prior AI usage), and $\text{FE}_i$ includes calendar-date and business-unit fixed effects. The baseline group is individuals working without AI (Individual No AI). 其中， $Y_i$ 代表各种因变量（如质量、新颖性等）。 $\text{Controls}_i$ 表示实验前的个人控制变量（司龄、职级、先前的AI使用频率）， $\text{FE}_i$ 表示日期和业务部门固定效应。基准对照组为独立作业且无AI辅助的个人（Individual No AI）。 --- ## 5. 实验结果 / Results ### 5.1. 绩效提升 / Performance Figure 2 and Figure 3 show that both AI-assisted individuals and teams outperformed the control group significantly. 图2和图3显示了不同实验组在方案最终评估质量上的差异。 **Figure 2. Average Solution Quality / 图2. 解决方案平均质量（标准化得分）** ![Figure 2](https://abxcyz.com/minio/abxcyz/20260801/917e0b3ab0c54229be7cc6cd3ff5d7dd.gif) **Figure 3. Pairwise Density Comparisons / 图3. 各组得分概率密度分布两两对比** ![Figure 3](https://abxcyz.com/minio/abxcyz/20260801/08dece9ee3564aef96093510f53e5f76.gif) --- ### 表2. OLS回归分析：最终解决方案质量 / Table 2. Solution Quality (Standardized) | 变量 / Dependent Variable: | 质量 (Quality - 1) | 质量 (Quality - 2) | 质量 (Quality - 3) | | :--- | :---: | :---: | :---: | | **人类团队无AI (Team No AI)** | 0.245**
(0.120) | 0.262**
(0.122) | 0.307**
(0.131) | | **个人 + AI (Individual + AI)** | 0.373***
(0.106) | 0.386***
(0.108) | 0.370***
(0.107) | | **团队 + AI (Team + AI)** | 0.392***
(0.122) | 0.404***
(0.123) | 0.463***
(0.139) | | 日期与业务单元固定效应 (FE) | 否 (No) | 是 (Yes) | 是 (Yes) | | 个人及背景控制变量 (Controls) | 否 (No) | 否 (No) | 是 (Yes) | | 样本数 (Observations) | 550 | 550 | 550 | | 调整后 $R^2$ (Adjusted $R^2$) | 0.023 | 0.023 | 0.048 | | *p-value (Team+AI = Team No AI)* | 0.242 | 0.254 | 0.216 | *(注：回归系数基于基准对照组（个人无AI）进行标准化；括号内为稳健标准误 / Robust standard errors in parentheses. *p < 0.1, **p < 0.05, ***p < 0.01.)* From Table 2, teams without AI improve quality by 0.24 standard deviations over individuals without AI, validating the traditional benefits of team collaboration. Interestingly, individuals with AI improve quality by 0.37 standard deviations, matching or even slightly exceeding the performance of teams without AI (the difference between the two is statistically indistinguishable). Adding a human teammate to an AI-enabled individual (Team + AI) yields an improvement of 0.39-0.46 standard deviations, which does not show a statistically significant marginal benefit over the Individual + AI group. 通过表2可以得出： * 在没有AI的情况下，人类团队合作比单干提升了 **0.24个标准差** (p < 0.05)，这证实了传统团队合作的绩效红利。 * **令人瞩目的是**，当个人拥有AI辅助后（个人+AI），其产出质量大幅提升了 **0.37个标准差** (p < 0.01)。这表明，**使用AI的个人在产出质量上完全可以比肩、甚至略微超过不使用AI的人类双人团队**。 * 当我们在AI的基础上再叠加一个人类伙伴（团队+AI）时，方案质量提升了 **0.39至0.46个标准差**。然而，这一表现与“个人+AI”相比，在统计学上并没有表现出显著的边际增量，提示了合作中的边际递减效应。 --- ### 表3. 解决方案文本长度（字数）分析 / Table 3. Solution Length | 变量 / Dependent Variable: | 字数 (Length - 1) | 字数 (Length - 2) | 字数 (Length - 3) | | :--- | :---: | :---: | :---: | | **人类团队无AI (Team No AI)** | 30.456
(27.419) | 56.746*
(30.865) | 57.184
(38.673) | | **个人 + AI (Individual + AI)** | 504.507***
(42.963) | 511.568***
(45.206) | 503.833***
(45.081) | | **团队 + AI (Team + AI)** | 543.745***
(42.328) | 556.997***
(43.737) | 551.578***
(51.989) | | 对照组均值 (Control Mean) | 381.422 | 306.565 | 336.197 | | 样本数 (Observations) | 550 | 550 | 550 | | 调整后 $R^2$ (Adjusted $R^2$) | 0.317 | 0.337 | 0.344 | 实验表明，AI的使用极大地充实了方案的细节。在AI辅助下（个人+AI与团队+AI），最终提交方案的长度显著增加了500多字，显示出AI在拓展论述深度和全面性上的强大能力。 --- ### 5.2. 专长壁垒的打破 / Expertise One of the most remarkable findings is how AI helps non-core-job professionals bridge the expertise gap. 实验中一个极具组织行为学意义的发现是：**AI能够显著帮助在非核心岗位（非新产品开发岗位）工作的员工弥补专业能力劣势。** **Figure 4. Average Quality: Core Jobs vs. Non-Core Jobs / 图4. 专长匹配度与AI引入的异质性绩效分析** ![Figure 4](https://abxcyz.com/minio/abxcyz/20260801/42257b9a5cb74fdfb3deb1fa1dcdb735.gif) --- ### 表4. 分样本回归：按员工任务熟悉度分组 / Table 4. Quality by Task Familiarity | 组别 / Group: | 非核心岗位员工 (Non-Core Jobs) | 核心创新岗位员工 (Core Jobs) | | :--- | :---: | :---: | | **人类团队无AI (Team No AI)** | -0.132
(0.248) | 0.377**
(0.165) | | **个人 + AI (Individual + AI)** | 0.360**
(0.156) | 0.457***
(0.153) | | **团队 + AI (Team + AI)** | 0.203
(0.253) | 0.455**
(0.179) | | 样本数 (Observations) | 218 | 332 | *(注：表内系数提取自包含所有控制变量的Model 3 / Coefficients from Model 3 with controls.)* As shown in Table 4, for professionals whose daily jobs do not involve product development (Non-Core Jobs), working in teams without AI did not bring positive performance gains. However, when equipped with AI, non-core-job individuals working alone improved performance by 0.360 standard deviations, comparable to teams with core-job expertise. 如表4所示，对于日常工作较少接触新产品开发（即“非核心岗位”）的员工，在没有AI时，即使把他们强行组队也无法带来绩效提升（Team No AI得分为-0.132且不显著）。**但是，一旦赋予他们AI工具（个人+AI），他们的绩效瞬间跃升了0.360个标准差**，达到了与核心岗位专家类似的产出水平。这意味着AI能够充当专业技能的“均衡器”（Equalizer），消除组织内部的经验劣势。 **Figure 5. Degree of Solution Technicality for Individuals / 图5. 个人方案的技术化程度分布（无AI vs. 有AI）** ![Figure 5]() In addition, without AI, commercial professionals generated commercially-oriented solutions while R&D professionals proposed technically-oriented ones. With AI, this functional barrier was dismantled—both groups submitted balanced solutions that combined commercial and technical perspectives. 此外，在传统的单兵作战中，商业人员提交的方案偏市场，研发人员的方案偏技术（见图5左侧双峰对峙）。**而在使用AI辅助后，这种职能壁垒消失了（见图5右侧重合分布）**。无论员工背景如何，在AI的启发下，他们都能写出商业模式与技术可行性兼备的、均衡而高水准的综合方案。 --- ### 5.3. 社交性与情感体验 / Sociality and Emotion Surprisingly, interacting with AI also brought human-to-human collaborative advantages to solo workers. We measured positive emotions (enthusiasm, energy, and excitement) and negative emotions (anxiety, frustration, and distress). 通常人们担心，用AI工具替代人类沟通会导致员工孤立和情绪低落。然而，实验结果却完全相反。 **Figure 6 & 7. Emotional Changes During the Task / 图6 & 图7. 任务期间积极情感增加值与消极情感减少值** ![Figure 6](https://abxcyz.com/minio/abxcyz/20260801/c6b4929caa9645a684c761e1017e5d76.gif) ![Figure 7](https://abxcyz.com/minio/abxcyz/20260801/9019711df5454e50857e83c80807f14f.gif) --- ### 表5. 积极情绪变化回归分析 / Table 5. Positive Emotions Change | 变量 / Dependent Variable: | 积极情绪提升 (Positive - 1) | 积极情绪提升 (Positive - 2) | 积极情绪提升 (Positive - 3) | | :--- | :---: | :---: | :---: | | **人类团队无AI (Team No AI)** | 0.269**
(0.124) | 0.254**
(0.126) | 0.257*
(0.137) | | **个人 + AI (Individual + AI)** | 0.457***
(0.107) | 0.475***
(0.106) | 0.485***
(0.106) | | **团队 + AI (Team + AI)** | 0.635***
(0.131) | 0.635***
(0.129) | 0.666***
(0.153) | | 样本数 (Observations) | 533 | 533 | 533 | --- ### 表6. 消极情绪变化回归分析 / Table 6. Negative Emotions Change | 变量 / Dependent Variable: | 消极情绪降低 (Negative - 1) | 消极情绪降低 (Negative - 2) | 消极情绪降低 (Negative - 3) | | :--- | :---: | :---: | :---: | | **人类团队无AI (Team No AI)** | -0.136
(0.124) | -0.094
(0.121) | -0.006
(0.141) | | **个人 + AI (Individual + AI)** | -0.233**
(0.117) | -0.247**
(0.116) | -0.263**
(0.117) | | **团队 + AI (Team + AI)** | -0.235**
(0.118) | -0.221*
(0.116) | -0.157
(0.138) | | 样本数 (Observations) | 530 | 530 | 530 | *(注：消极情绪得分为负数，代表消极情绪显著减少 / Negative coefficients denote a reduction in negative feelings.)* As shown in Tables 5 and 6, individuals working with AI experienced a 0.485 standard deviation increase in positive emotions and a 0.263 standard deviation decrease in negative emotions. This emotional relief is twice as strong as the effect of partnering with a human teammate without AI (Team No AI). This language-based AI counterpart acts as an encouraging partner, effectively reducing distress and loneliness in high-pressure tasks. 数据分析显示，相比无AI的个人单兵作战： * 拥有AI的个人积极情绪显著上升了 **0.485个标准差** (p < 0.01)，消极情绪显著减少了 **0.263个标准差** (p < 0.05)。 * **这一情绪改善幅度，甚至是普通人类配对组（无AI团队）的两倍**。 * 这得益于生成式AI极富人情味的自然语言交互界面。在遇到困难时，AI能够源源不断地提供富有建设性、富有包容性甚至“体贴”的语言反馈，扮演了传统意义上只能由人类队友提供的社交支持和减压角色。 --- ### 表7. 积极情绪提升与未来AI使用期望的相关性 / Table 7. Positive Emotion Change vs Expected Future Use | 变量 / Dependent Variable: Positive Emotions | 无AI组 (Control) | 有AI组 (Treatment) | | :--- | :---: | :---: | | **对将来工作使用GenAI的期望差值** | 0.140
(0.182) | 0.638***
(0.243) | | 样本数 (Observations) | 262 | 271 | --- ### 表8. 消极情绪减少与未来AI使用期望的相关性 / Table 8. Negative Emotion Change vs Expected Future Use | 变量 / Dependent Variable: Negative Emotions | 无AI组 (Control) | 有AI组 (Treatment) | | :--- | :---: | :---: | | **对将来工作使用GenAI的期望差值** | -0.170
(0.154) | -0.663***
(0.201) | 表7和表8证明，在实际体验了AI协助后，那些积极情绪提升最明显、压力减负最大的员工，其在未来工作中继续深度拥抱生成式AI的意愿和良性期望也最强烈。 --- ### 5.4. 过程与极值分析：超越总分 / Process and Extremes #### 5.4.1. 解构创意生成的不同阶段 / Decomposing AI's Contributions New product development consists of multiple stages: ideation (generating 5 initial ideas), selection, and execution (expanding the selected idea). We designed the experiment to restrict all groups to generating exactly 5 ideas, allowing us to isolate stage-by-stage impact. 新产品开发分为：创意萌芽（起初生成5个点子）、评估筛选、方案深度细化执行。我们在实验中做了一个精细的设计：**限制所有组别必须正好起草5个点子**，从而剔除数量优势，专注于透视人类和AI在质量维度的博弈。 **Figure 8. Decomposing the Innovation Funnel / 图8. 创新过程各个维度的深入解构** ![Figure 8](https://abxcyz.com/minio/abxcyz/20260801/4eae0378792c43fe9665c307819601ff.gif) * **Panel (a) Average Quality of Generated Ideas (生成的创意平均质量)**: AI conditions show significantly higher average quality in the portfolio of 5 generated ideas. (在生成的5个初始点子包中，使用AI的两组平均质量均远高于人类无AI组)。 * **Panel (b) Selection Accuracy (选择最佳创意的准确概率)**: Interestingly, human teams without AI (Team No AI) performed best in selecting their highest-quality idea (50% probability), while AI conditions dropped to roughly 37%. (在点子生成后，**没有AI的人类团队在挑选最佳点子上表现最优秀**，有50%的概率挑中5个里面最好的那个；而使用AI的组别该概率下滑到了37%左右)。 * **Panel (c) Quality Gap (创意包内的质量高低差)**: The variance of quality remains identical across all conditions, indicating AI shifts the entire performance curve upward rather than homogenizing ideas. (创意包内最差和最好创意的分差在所有组中基本一致，这说明AI整体抬升了创意产出分布，而非造成千篇一律的平庸方案)。 * **Panel (d) Selected Idea Quality (挑选出方案的绝对质量)**: Although AI users had slightly lower selection accuracy, they still submitted higher absolute-quality final ideas because their baseline portfolio quality (Panel a) was much higher. (虽然AI辅助下的人类筛选最佳点子的准确率稍逊一筹，但由于他们的创意初始大礼包（Panel a）绝对基数太高，所以最后挑选出来细化的终版创意在绝对分数上依然显著高出人类无AI组)。 --- #### 5.4.2. 追求卓越：产生“爆款创意”的概率 / Exceptional Performance Measures In innovation tournaments, the disproportionate value comes from the absolute best ideas (top-decile breakthroughs). 在真实的商业环境中，平庸创意毫无用处，创新的商业价值绝大部分由排在前10%的“爆款创意”贡献。 **Figure 9. Top 10% Solutions / 图9. 各实验组产生“排名前10%爆款创意”的概率** ![Figure 9](https://abxcyz.com/minio/abxcyz/20260801/634db7c34f57459cafe54d18715a329c.gif) --- ### 表9. 回归模型：产出极值（前10%优秀方案）的概率 / Table 9. Probability of Being in the Top 10% | 因变量 / Dependent Variable: | 前10%爆款率 (Top 10% - 1) | 前10%爆款率 (Top 10% - 2) | 前10%爆款率 (Top 10% - 3) | | :--- | :---: | :---: | :---: | | **人类团队无AI (Team No AI)** | 0.037
(0.033) | 0.045
(0.034) | 0.054
(0.041) | | **个人 + AI (Individual + AI)** | 0.019
(0.029) | 0.029
(0.029) | 0.030
(0.029) | | **团队 + AI (Team + AI)** | 0.092**
(0.037) | 0.098**
(0.038) | 0.112**
(0.045) | | 样本数 (Observations) | 550 | 550 | 550 | As shown in Table 9, the combination of human teamwork and AI (Team + AI) is extremely powerful for breakthrough outcomes. It increases the probability of generating a top-10% breakthrough solution by 11.2 percentage points over the individual control, translating to a nearly threefold increase in breakthrough potential. 从表9可以看出，**“团队+AI”的强强联合对于产出“爆款创意”具有极其显著的溢价效应**。其爆款产出率比基准个人组高出了 **11.2个百分点**（实际概率增幅达约3倍）。这说明，尽管个人+AI在平均分上与团队相当，但如果要探索 Rugged Landscape（崎岖复杂的创新景观）并攀登真正的最高峰，**“多职能人类团队 + AI”依然是企业寻求颠覆性突破的最佳配置。** --- ### 5.5. 补充分析 / Additional Analyses #### 5.5.1. 达克效应与自信过度 / Expected Quality Despite objective quality improvements, AI users were less confident about their performance. 有趣的是，尽管AI辅助下的客观方案质量更高，但这些参与者在主观上反而更“不自信”。 **Figure 10. Expected Top 10% Placement / 图10. 参与者主观认为自己能进“前10%优秀榜单”的比例** ![Figure 10](https://abxcyz.com/minio/abxcyz/20260801/47f1c6ac08ab4953bc7aeab76b9891ba.gif) As shown in Figure 10, AI-enabled participants were 9.2 percentage points less likely to expect their solutions to rank in the top 10%. This suggests that interacting with high-capability AI increases cognitive humility, preventing the typical overconfidence ("Dunning-Kruger effect") observed in human self-assessments. 如图10所示，拥有AI协助的员工预测自己能得高分的比例降低了9.2个百分点。这表明，与高智商、高博学的AI进行深层对话，在无形中提升了人类的认知谦逊（Cognitive Humility），打破了人类在单打独斗时往往会产生的盲目乐观和自我过度膨胀（即“达克效应”）。 --- #### 5.5.2. 团队融合度分析 / Human Team Collaboration As shown in Figure 11, without AI, human teams often fell into dominance traps—either R&D or commercial perspectives dominated the narrative, causing a bimodal distribution. AI-enabled teams achieved a unimodal, balanced distribution, suggesting AI moderates social friction and dominance in group dynamics. **Figure 11. Degree of Solution Technicality for Teams / 图11. 人类团队产出的技术性程度分布（无AI团队 vs. 有AI团队）** ![Figure 11](https://abxcyz.com/minio/abxcyz/20260801/cae32ab875704386b24e53d6725ae4a7.gif) 在没有AI的情况下，人类双人组在制定决策时往往存在“话语权争夺”，导致方案呈现极端的双峰分布（见图11左侧）。而引入AI后，其互动更趋温和中肯，促使团队采纳了融汇两家之长、更为温和且全面的单一分布方案（见图11右侧），有效缓解了传统团队中由于资历、性格差异导致的话语权霸凌。 --- #### 5.5.3. 采纳AI文案的深度 / Patterns of AI Use **Figure 12. Retention Rate of AI Text / 图12. 最终方案中保留AI生成原句的比例分布** ![Figure 12](https://abxcyz.com/minio/abxcyz/20260801/6a31be9f06fc4cd6b4f0aba46c3b76a3.gif) We analyzed the percentage of sentences in final submissions authored by AI. The polarized distribution suggests two main styles of use: "Centaurs" who utilize AI for ideation but rewrite the final copy, and "Cyborgs" who heavily integrate and submit AI's directly-generated text. 分析表明（图12），员工在使用AI时表现出两极分化：一部分人（半人马模式 Centaurs）只向AI寻找灵感，最终几乎全由自己撰写（左端保留率为0%）；而另一部分人（赛博格模式 Cyborgs）则深度整合并高度保留了AI生成的文本（右端保留率超过75%）。这体现了人类在与新技术共事时多元的协作习惯。 --- ## 6. 讨论 / Discussion Our study shows that generative AI has evolved beyond a passive software tool to become a genuine cybernetic teammate. By demonstrating capabilities to enhance individual performance to match human teams, reduce functional boundaries, and alleviate task-related distress, AI reshapes the traditional pillars of organizational collaboration. 本研究表明，生成式AI已经突破了传统软件的工具范畴，成为真正意义上的**“控制论队友”**。它不仅帮助单兵作战的个人跨越了研发与商业之间的专业鸿沟、提升产出绩效至比肩人类传统团队的水平，更在心理层面为知识型员工提供了正向的情感支撑。 > **背景补充与理论联系 / Theoretical Context and Discussion:** > > 根据著名的组织行为学家斯坦纳（Steiner, 1972）关于团队绩效的经典理论，团队在扩大规模时（如从单人变为双人、三人），虽然理论上注入了更多元的信息库，但也带来了不可避免的“协同损耗（Process Loss）”（例如沟通摩擦、搭便车行为、意见不一）。 > > 在我们的实验中， > * 从“个人无AI”跃升到“个人+AI（双人配对）”，绩效获得了爆发性增长，且带来了极佳的情感反馈； > * 但当我们更进一步组建“人类双人团队 + AI（三人组合）”时，其平均产出的提升开始面临边际收益递减。 > > 这一规律再次印证了斯坦纳的理论：如何在发挥AI专长互补优势的同时，最小化人类在面对AI系统或多人合作时的协调摩擦和认知负荷，将是未来设计高效人机协作系统的关键。 --- ## 7. 结论 / Conclusion Generative AI’s integration into collaborative work challenges foundational assumptions in organizational design. It offers exciting avenues for businesses to build more agile, cross-functional project teams, while highlighting the complementary value of human judgment in strategic evaluation. Future research should examine how these dynamics evolve over long-term collaboration. 生成式AI的崛起正在对经典的组织行为学和团队设计理论发起挑战。它不仅为企业组织架构扁平化、灵活构建跨职能混合团队提供了巨大机遇，同时也向我们揭示：即便技术再如何强大，**人类的自主选择与最终决策（筛选好创意的直觉与商业判断力）**依然是人机协同系统中不可或缺的灵魂。 --- * 论文补充材料与原始数据请参见官方出版地址 / Supplemental Material and online appendix are available at: [https://doi.org/10.1287/orsc.2025.20702](https://doi.org/10.1287/orsc.2025.20702)*