生成式人工智能应用中的角色迁移与期望重新校准：对某州交通运输部的纵向研究
Persona Migration and Expectation Recalibration in Generative AI Adoption: A Longitudinal Study at a State Department of Transportation
浏览论文内容
中文总结 AI 辅助
研究某州交通运输部对Microsoft 365 Copilot的采用，通过匹配两波调查等方法，发现使用后感知有用性下降，确定三种基线角色及迁移情况，提出公共部门人工智能采用应动态监测并提供针对性支持，给出员工异质性跟踪框架。
中文摘要 AI 辅助
生成式人工智能工具在公共机构中的试点日益增多，但关于员工实际使用后接受度变化的证据有限。本研究考察了某州交通运输部在为期八周的试点中对Microsoft 365 Copilot的采用情况。通过匹配的两波调查测量了参与前后的感知有用性、感知易用性、行为意图和信任度。样本包括124名员工，通过非参数检验评估总体变化，k均值聚类确定基线接受角色，固定质心分配跟踪迁移。开放式回答通过基于关键词的内容映射进行检查。使用后感知有用性显著下降，表明期望重新校准，而感知易用性、行为意图和信任度变化较小且不显著。出现了三种基线角色：怀疑者、谨慎积极使用者和拥护者。尽管角色数量变化不大，但个体迁移幅度较大。向上迁移与有用性、行为意图和信任度的提高相关，向下迁移则相反。沟通和总结仍然是稳定的用例，而数据、图表和演示任务减少。准确性和隐私担忧减少，但工作和技能担忧增加。公共部门人工智能的采用应通过特定角色的培训、工作流程示例、验证程序和信任校准保障措施进行动态监测和支持。该研究提供了一个在企业生成式人工智能实施过程中跟踪员工异质性的框架。
英文摘要
Generative AI tools are increasingly being piloted in public agencies, but limited evidence explains how employee acceptance changes after hands-on use. This study examines Microsoft 365 Copilot adoption during an eight-week pilot at a state Department of Transportation. A matched two-wave survey measured perceived usefulness, perceived ease of use, behavioral intention, and trust before and after participation. After matching and response-quality screening, the sample included 124 employees. Nonparametric tests assessed aggregate changes, k-means clustering identified baseline acceptance personas, and fixed-centroid assignment tracked migration. Open-ended responses were examined using keyword-based content mapping. Perceived usefulness declined significantly after use, suggesting recalibration of expectations, while perceived ease of use, behavioral intention, and trust showed only small, nonsignificant changes. Three baseline personas emerged: Skeptics, Cautiously Positive users, and Champions. Although persona counts changed modestly, individual movement was substantial: 40 percent of Skeptics moved to Cautiously Positive, while 68 percent of Champions moved to less enthusiastic personas. Upward movement was associated with gains in usefulness, behavioral intention, and trust; downward movement was associated with declines in usefulness and trust. Communication and summarization remained stable use cases, while data, chart, and presentation tasks declined. Accuracy and privacy concerns decreased, but job and skills concerns increased. Public-sector AI adoption should be monitored dynamically and supported through persona-specific training, workflow examples, verification routines, and trust-calibration safeguards. The study offers a framework for tracking workforce heterogeneity during enterprise generative AI implementation.
