Scholarlink

<p>AI agents that autonomously plan and execute multi-step tasks represent a&#xa0;fundamental shift from conventional chatbots. They book travel, compose emails, or analyze financial data increasingly without human approval for each individual step. However, this autonomy introduces novel risks absent from traditional software: agents may hallucinate non-existent interfaces, permanently embed erroneous information in their memory, or escalate minor errors across chained workflows into consequences that are difficult to reverse. Conventional risk assessments, designed for static models, fall short in addressing these challenges. Trust thus becomes not merely a&#xa0;matter of acceptance but a&#xa0;critical factor for safety and quality. This article combines architectural risk analyses of agentic systems with established trust models and insights from social explanation research. Following a&#xa0;design science research approach, five design recommendations are derived: contrastive decision transparency, graduated action autonomy, proactive uncertainty communication, auditable and correctable memory, and domain-specific competence boundaries. The recommendations are illustrated through a&#xa0;consistent practical scenario, an AI travel assistant, and are aimed at UX designers and system architects. The overarching insight: trustworthiness does not emerge from maximizing performance claims but from a&#xa0;calibrated approach to the system’s own limitations.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Vertrauen ist kein Feature: Fünf Gestaltungsempfehlungen für autonome KI-Agenten

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月29日

查看全文
  • Patrick Hiske,
  • Varinia Wittholz,
  • Susanne Robra-Bissantz

摘要

AI agents that autonomously plan and execute multi-step tasks represent a fundamental shift from conventional chatbots. They book travel, compose emails, or analyze financial data increasingly without human approval for each individual step. However, this autonomy introduces novel risks absent from traditional software: agents may hallucinate non-existent interfaces, permanently embed erroneous information in their memory, or escalate minor errors across chained workflows into consequences that are difficult to reverse. Conventional risk assessments, designed for static models, fall short in addressing these challenges. Trust thus becomes not merely a matter of acceptance but a critical factor for safety and quality. This article combines architectural risk analyses of agentic systems with established trust models and insights from social explanation research. Following a design science research approach, five design recommendations are derived: contrastive decision transparency, graduated action autonomy, proactive uncertainty communication, auditable and correctable memory, and domain-specific competence boundaries. The recommendations are illustrated through a consistent practical scenario, an AI travel assistant, and are aimed at UX designers and system architects. The overarching insight: trustworthiness does not emerge from maximizing performance claims but from a calibrated approach to the system’s own limitations.