Effects of knowledge and importance on responsibility in human-AI decision making

Area of Science:

• Human-Computer Interaction
• Artificial Intelligence Ethics
• Cognitive Psychology

Background:

• Accountability for AI failures is critical as AI systems increasingly support human decision-making.
• Existing research on responsibility attribution focuses on system autonomy, transparency, or anthropomorphism.
• The joint influence of cognitive framing (prior agent knowledge) and contextual framing (task importance) on responsibility judgments is understudied.

Purpose of the Study:

• To investigate how cognitive and contextual framing jointly influence responsibility attribution in human-agent interactions.
• To examine how prior knowledge of an AI agent and the perceived importance of a task affect judgments of responsibility.
• To identify factors shaping accountability gaps in agent-assisted decision-making.

Main Methods:

• A three-factor mixed-design experiment involving 588 participants.
• Participants assessed responsibility for users, agents, and developers following simulated failed agent-assisted interactions.
• Manipulation of factors included prior knowledge of the agent and perceived task importance.

Main Results:

• Prior knowledge of the AI agent significantly shifted responsibility away from the user and towards the agent and its developer.
• Increased perceived task importance substantially elevated responsibility attributed to the developer or provider.
• Responsibility attribution was found to be dynamic, influenced by user expectations and situational seriousness.

Conclusions:

• Responsibility in human-agent interactions is not static but dynamically modulated by user familiarity with the agent and the perceived importance of the task.
• Findings have implications for designing AI systems, educating users, and informing legal policies to address accountability gaps.
• Understanding these framing effects is crucial for the responsible deployment of socially embedded AI agents.