Beliefs about accuracy shape confidence attributions to humans and artificial agents

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Effective communication requires understanding others' cognitive and metacognitive states.
• Metacognitive abilities are developing in AI, prompting research into human attribution of confidence to AI.

Purpose of the Study:

• To investigate how humans attribute confidence to AI agents compared to human agents.
• To explore the factors influencing these attributions and the robustness of any observed effects.

Main Methods:

• Seven pre-registered experiments were conducted where participants observed human and AI agents making perceptual choices.
• Participants reported their perception of the observed agent's confidence.
• Behavioral cues (task difficulty, accuracy, response times) and agent type were analyzed.

Main Results:

• Human observers consistently overestimated AI confidence relative to human confidence, even with identical behavior.
• This overestimation persisted across different tasks and agent descriptions, except in more subjective decision-making.
• Prior beliefs about agent capabilities were shown to influence these confidence illusions.

Conclusions:

• A significant "illusion of confidence" exists in human attributions to AI systems.
• Metacognitive attributions play a crucial role in both human-human and human-AI interactions.
• Understanding these illusions is key for effective human-AI collaboration.