# Trust in AI: progress, challenges, and future directions > [!info]+ <center>Metadata</center> > > |<div style="width: 5em">Key</div>|Value| > |--:|:--| > |文献类型|journalArticle| > |标题|Trust in AI: progress, challenges, and future directions| > |短标题|对人工智能的信任:进步、挑战和未来方向| > |作者|[[Saleh Afroogh]]、 [[Ali Akbari]]、 [[Emmie Malone]]、 [[Mohammadali Kargar]]、 [[Hananeh Alambeigi]]| > |期刊名称|[[Humanities and Social Sciences Communications]]| > |DOI|[10.1057/s41599-024-04044-8](https://doi.org/10.1057/s41599-024-04044-8)| > |存档位置|| > |文库编目|DOI.org| > |索书号|| > |版权|| > |分类|[[Nature系列]]| > |条目链接|[My Library](zotero://select/library/items/UL6VUERT)| > |PDF 附件|[Afroogh 等 - 2024 - Trust in AI progress, challenges, and future directions.pdf](zotero://open-pdf/library/items/WC9FQHNB)| > |关联文献|| > ^Metadata > [!example]- <center>本文标签</center> > > `$=dv.current().file.tags` > [!quote]- <center>Abstract</center> > > We conducted an inclusive and systematic review of academic papers, reports, case studies, and trust frameworks in AI, written in English. Given that there is not a specific database on trust in AI in particular, we used the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework to develop a protocol in this review (Fig. 1). In order to conduct a comprehensive review of the relevant studies, we followed two approaches. First, we manually searched for the most related papers on trust in AI: 19 papers were identified through the online search after the removal of duplicate files. Secondly, we fulfilled a keyword-based search (using the http://scholar.google.com search engine) to collect all relevant papers on the topic. This search was accomplished using the following keyword phrases: (1) “trust + AI” which provided 19 relevant result pages of Google Scholar, (2) “trust + Artificial + Intelligence” for which the first five result pages were reviewed, (3) “trustworthy + AI,” for which the first 15 result pages were reviewed; and (4) “trustworthy + Artificial + Intelligence,” for which the first 13 result pages of Google Scholar were reviewed. > [!tldr]- <center>隐藏信息</center> > > itemType:: journalArticle > title:: Trust in AI: progress, challenges, and future directions > shortTitle:: 对人工智能的信任:进步、挑战和未来方向 > creators:: [[Saleh Afroogh]]、 [[Ali Akbari]]、 [[Emmie Malone]]、 [[Mohammadali Kargar]]、 [[Hananeh Alambeigi]] > publicationTitle:: [[Humanities and Social Sciences Communications]] > journalAbbreviation:: Humanit Soc Sci Commun > volume:: 11 > issue:: 1 > pages:: 1568 > series:: > language:: en > DOI:: [10.1057/s41599-024-04044-8](https://doi.org/10.1057/s41599-024-04044-8) > ISSN:: 2662-9992 > url:: [https://www.nature.com/articles/s41599-024-04044-8](https://www.nature.com/articles/s41599-024-04044-8) > archive:: > archiveLocation:: > libraryCatalog:: DOI.org > callNumber:: > rights:: > extra:: 🏷️ /unread、📒 > collection:: [[Nature系列]] > tags:: #unread > related:: > itemLink:: [My Library](zotero://select/library/items/UL6VUERT) > pdfLink:: [Afroogh 等 - 2024 - Trust in AI progress, challenges, and future directions.pdf](zotero://open-pdf/library/items/WC9FQHNB) > qnkey:: Afroogh 等 - 2024 - Trust in AI:progress, challenges, and future directions > date:: 2024-11-18 > dateY:: 2024 > dateAdded:: 2025-04-13 > datetimeAdded:: 2025-04-13 14:27:26 > dateModified:: 2025-04-13 > datetimeModified:: 2025-04-13 15:00:39 > > abstract:: We conducted an inclusive and systematic review of academic papers, reports, case studies, and trust frameworks in AI, written in English. Given that there is not a specific database on trust in AI in particular, we used the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework to develop a protocol in this review (Fig. 1). In order to conduct a comprehensive review of the relevant studies, we followed two approaches. First, we manually searched for the most related papers on trust in AI:19 papers were identified through the online search after the removal of duplicate files. Secondly, we fulfilled a keyword-based search (using the http://scholar.google.com search engine) to collect all relevant papers on the topic. This search was accomplished using the following keyword phrases:(1) “trust + AI” which provided 19 relevant result pages of Google Scholar, (2) “trust + Artificial + Intelligence” for which the first five result pages were reviewed, (3) “trustworthy + AI,” for which the first 15 result pages were reviewed; and (4) “trustworthy + Artificial + Intelligence,” for which the first 13 result pages of Google Scholar were reviewed. %--------------ω--------------% 以下是《人文与社会科学通讯》中《对AI的信任:进展、挑战与未来方向》文献综述的详细总结: --- ### **研究背景与意义** - AI技术已渗透医疗、交通、金融、军事等核心领域,成为社会基础设施的一部分。然而,用户对AI的信任度直接影响技术采用率,信任缺失可能阻碍AI的广泛应用。 - AI系统具有自主学习、不可预测性和“黑箱”特性,导致其决策透明度和可解释性不足,进一步加剧信任危机。 - 需要系统性界定AI信任的定义、范畴及影响因素,探索构建可信AI的技术与非技术路径,推动负责任的AI发展。 --- ### **方法论** - 采用PRISMA框架进行系统文献综述,通过Google Scholar等平台检索关键词(如“Trust + AI”“Trustworthy AI”),筛选329篇相关文献。 - 纳入标准:聚焦AI信任的学术文章,涵盖技术、伦理、法律等多维度分析,排除重复及主题不符文献。 --- ### **核心研究发现** #### **1. AI信任的模型与类型** - **信任定义特殊性**:不同于人际信任(基于善意与诚实),AI信任围绕技术能力(准确性、可靠性)、透明度(可解释性)及伦理合规性。用户接受脆弱性并以预期行为匹配度为信任基础。 - **信任类型**: - **人机交互**:如医生信任医疗AI诊断、用户依赖自动驾驶决策。 - **机机交互**:物联网设备间的信任协作,需防对抗攻击(如区块链用于智能合约验证)。 - **AI与对象交互**:如自动驾驶系统识别可信交通标志或社交媒体数据的真实性。 #### **2. 可信AI的评估指标** - **技术指标**: - **安全性**:系统抵御攻击与错误的能力(如医疗数据隐私保护)。 - **准确性**:预测或决策的正确率(如金融风险评估模型)。 - **鲁棒性**:在噪声或数据偏差下的稳定性(如自动驾驶应对极端天气)。 - **价值指标**: - **伦理**:避免偏见(如招聘AI的公平性)、保障隐私(如用户数据匿名化)。 - **法律合规**:符合GDPR等监管要求(如算法可审计性)。 - **社会效益**:促进可持续发展(如环保能源管理AI)。 #### **3. 不信任AI的根源** - **技术缺陷**:模型错误(金融AI的错误建议)、数据偏见(医疗诊断中的种族偏差)。 - **伦理与法律威胁**:如人脸识别侵犯隐私、自动化武器威胁人类安全。 - **自主性威胁**:AI替代人类决策引发的失控风险(如自动驾驶紧急避让的伦理困境)。 - **尊严威胁**:情感交互型AI(如护理机器人)削弱人际关系的真实性。 #### **4. 信任建立策略** - **技术增强**:开发可解释AI(XAI)、提高模型鲁棒性(对抗训练)、实时不确定性量化。 - **透明度措施**:公开算法逻辑(如开源代码)、提供决策追溯(医疗诊断依据可视化)。 - **伦理与治理**:建立行业标准(如IEEE伦理指南)、跨学科监管框架(技术+法律专家协作)。 - **用户教育**:提升公众对AI能力的合理认知(避免过度依赖或排斥)。 --- ### **讨论与挑战** - **价值冲突**:如透明度增加可能降低模型性能(加密影响计算效率),需权衡实际需求。 - **动态信任校准**:用户信任随时间变化(如自动驾驶初期高信任→事故后信任崩塌),需自适应调节机制。 - **文化差异**:不同地区对隐私、公平性的定义不一(如亚洲集体主义vs西方个人主义),需本土化策略。 --- ### **结论与未来方向** - **研究重点**:需融合技术、伦理、法律等多学科,制定动态信任模型与统一评估体系。 - **技术革新**:开发实时信任监测工具、个性化信任校准算法(如基于用户性格特征)。 - **政策推动**:政府与企业协作建立认证机制(如可信AI标签),加强国际标准互认。 - **社会参与**:通过公众咨询、参与式设计提升AI系统的社会接受度。 --- ### **典型案例** - **医疗领域**:病理AI需结合医生经验,解释性功能(如热图标记)可增加信任;但过度依赖AI可能导致误诊。 - **金融领域**:聊天机器人因中立性获得信任,但复杂产品仍需人类顾问(信任的“算法厌恶”现象)。 - **自动驾驶**:用户对技术可行性的怀疑(如紧急制动可靠性)需通过模拟测试与透明度提升缓解。 --- 该综述系统梳理了AI信任的全景,为技术开发者、政策制定者及用户提供了实践指南,并指出现有研究的空白(如跨文化信任差异),为未来研究指明方向。 ## ✏️ 笔记区 > [!WARNING]+ <center>🐣 总结</center> > >🎯 一句话总结:: > [!inbox]- <center>📫 导入时间</center> > >⏰ importDate:: 2025-04-13 >⏰ importDateTime:: 2025-04-13 15:00:09 https://www.jianguoyun.com/p/DXLzUfoQk6_XChi68_QFIAA %--------------ω--------------%