Let people fail! Exploring the influence of explainable virtual and robotic agents in learning-by-doing tasks

AI explanations may undermine learning outcomes: Participants who completed the learning-by-doing task independently acquired superior knowledge compared to those assisted by explainable AI agents, suggesting that relying on AI explanations—even when justified—can reduce the depth of human learning and problem-solving skills.

The physical form of AI influences human behavior differently: When interacting with a humanoid robot, participants were more likely to follow its suggestions compared to a computer interface, yet this increased compliance did not translate to better task performance, highlighting how the embodiment of AI can affect human trust and deference in ways that may not be beneficial.

Explanation style matters, but inconsistently: "Partner-aware explanations" (tailored to the human's perspective) produced different behavioral effects depending on whether participants interacted with a computer or robot, suggesting that the effectiveness of AI communication strategies is not universal and depends on contextual factors.

Human autonomy may be essential for skill development: The superior knowledge acquisition among participants without AI assistance implies that struggling through challenges independently—rather than receiving AI guidance—is crucial for developing robust understanding and competence in learning tasks.

Explainable AI requires careful implementation in educational contexts: These findings challenge the assumption that more transparent, justified AI recommendations automatically improve human-AI collaboration, particularly in tutoring and training scenarios where human learning and growth should be primary objectives.