AI-TAM: a model to investigate user acceptance and collaborative intention in human-in-the-loop AI applications

More and more frequently, digital applications make use of Artificial Intelligence (AI) capabilities<br> to provide advanced features; on the other hand, human-in-the-loop approaches are on the<br> rise to involve people in AI-powered pipelines for data collection, results validation and decision making.<br> Does the introduction of AI features affect user acceptance? Does the AI result quality<br> affect people’s willingness to use such applications? Does the additional user effort required in<br> human-in-the-loop mechanisms change the application adoption and use?<br> This study aims to provide a reference approach to answer those questions. We propose a model<br> that extends the Technology Acceptance Model (TAM) with further constructs explicitly related to<br> AI – user trust in AI and perceived quality of AI output, from explainable AI (XAI) literature – and<br> collaborative intention – willingness to contribute to AI pipelines.<br> We tested the proposed model with an application for car damage claim reporting with AI-powered<br> damage estimation for insurance customers. The results showed that the XAI related factors have<br> a strong and positive effect on behavioral intention, perceived usefulness, and ease of use of the<br> application. Moreover, there is a strong link between behavioral intention and collaborative intention,<br> indicating that indeed human-in-the-loop approaches can be successfully adopted in final user<br> applications. Users were invited to test the interactive prototype of the BumpOut application and to report the given car accident from start to finish. These are the two interactive prototypes experienced by users: FlawlessAI-Group prototype FailingAI-Group prototype This study is shared as a research object adopting the RO-Crate specification.越来越多的数字应用正借助人工智能（Artificial Intelligence，AI）能力以提供高级功能；与此同时，人在回路（human-in-the-loop）方法正逐渐兴起，让人类参与到人工智能驱动的流程中，承担数据采集、结果验证与决策制定的工作。 人工智能功能的引入是否会影响用户接受度？人工智能的结果质量是否会影响人们使用此类应用的意愿？人在回路机制中额外所需的用户工作量，是否会改变应用的采纳与使用情况？ 本研究旨在提供一种可用于解答上述问题的参考范式。我们提出了一个扩展版技术接受模型（Technology Acceptance Model，TAM），新增了与人工智能明确相关的构念：源自可解释人工智能（explainable AI，XAI）相关文献的用户对人工智能的信任度、人工智能输出感知质量，以及协作意愿——即参与人工智能流程的贡献意愿。 我们依托一款面向保险客户、搭载人工智能驱动的损伤评估功能的汽车理赔申报应用，对所提出的模型进行了实证检验。结果显示，与可解释人工智能相关的因素，对该应用的行为意向、感知有用性与感知易用性均具有显著正向影响。此外，行为意向与协作意愿之间存在强关联，这表明人在回路方法确实可在终端用户应用中成功落地。 研究邀请用户测试BumpOut应用的交互式原型，并从始至终上报所遭遇的交通事故。用户共体验过两款交互式原型：FlawlessAI-Group原型与FailingAI-Group原型。 本研究以RO-Crate规范作为研究对象的共享标准，现已公开共享。