Explaining AI Decisions to Mitigate Cognitive Biases in
Human-AI Collaboration
Miswan Gumanti1,* Citra Dewi2
1 Institut Bakti Nusantara, Lampung, Indonesia
2 Universitas Lampung, Indonesia
Emails: mgumanti0205@gmail.com · citra.dewi@eng.unila.ac.id
Received: January 24, 2026 Revised: February 27, 2026 Accepted: March 28, 2026 ⋆ Corresponding author
ABSTRACT
Human-AI collaboration can improve decision quality only when users know when to rely on an AI recommendation
and when to resist it. Explanations are often proposed as a remedy, but explanation content can also intensify
automation bias or reinforce a user’s initial belief. This paper presents a cognitive explanation-selection model
for mitigating over-reliance and under-reliance in AI-assisted decision tasks. The study compares no explanation,
feature-based, contrastive, example-driven, and hybrid explanations across simulated novice, intermediate, and
expert decision makers using a public medical decision dataset as the task substrate. The analysis focuses on
reliance behaviour rather than on model accuracy alone. The proposed model estimates when the user is likely to
accept a wrong recommendation, reject a correct recommendation, or accept advice simply because it confirms
an initial judgment. The results indicate that contrastive and hybrid explanations are more effective for reducing
automation bias, while example-driven explanations preserve trust for lower-expertise users. The paper concludes
with a transparent interface loop for high-stakes environments in which explanation style is selected according to
user expertise, AI confidence, and human-AI agreement.
Keywords: Human-AI collaboration Explainable AI Automation bias Confirmation bias Appropriate reliance
1. INTRODUCTION
AI-assisted decision-making is now common in domains
where human judgment remains legally, ethically, or operationally
central. Medical monitoring, risk triage, air traffic
supervision, financial screening, and emergency response all
involve situations in which an AI model may recommend
an action while a human operator retains responsibility for
the final decision. In these settings, better model accuracy
is not sufficient. A highly accurate model can still harm performance
when users accept its incorrect recommendations
too readily, ignore its correct recommendations, or treat its
explanations as rhetorical persuasion rather than as evidence
to be inspected.
This paper addresses that problem by studying explanation
formats as cognitive interventions. The central question is not
simply whether explanations increase trust, but whether they
improve reliance calibration. Appropriate reliance requires
two complementary behaviours: accepting correct AI advice
when it improves the user’s decision, and rejecting incorrect
AI advice when the user’s own judgment is better. This
distinction is critical because many explanation interfaces
raise acceptance of AI advice in general. When acceptance
rises for both correct and incorrect advice, the interface may
create automation bias rather than reliable collaboration.
A second difficulty is confirmation bias. Users rarely encounter
AI recommendations as blank slates. They bring
initial beliefs, expertise, previous cases, and domain expectations
into the decision process. When an AI recommendation
agrees with the user’s first impression, the recommendation
may be accepted with little scrutiny. When it disagrees, it
may be dismissed defensively. Explanation formats therefore
need to do more than justify the AI output; they need to cre-