The Challenge
Contemporary AI deployments predominantly position human users as passive recipients of algorithmic decisions, constrained to post-hoc explainability (XAI) or basic contestability. Despite the rise of data-intensive systems, AI evaluation remains heavily technocentric, optimizing purely for model accuracy rather than collaborative utility or agency. This creates a critical gap where workers and domain experts cannot meaningfully intervene in, modify, or co-design underlying algorithmic mechanics. Without intentional, interactive mechanisms, organizations encounter severe adoption barriers, deskilling, and ethical risks in high-stakes operational environments where shared human-machine responsibility is vital.
Core Findings
Conducting a comprehensive systematic literature review across Human-Centered AI (HCAI), Interactive Machine Learning (IML), Contestable AI, and Hybrid Intelligence, Raees et al. evaluate human-AI interaction across three dimensions: users, implementations, and system goals. The authors find that current empirical research disproportionately focuses on low-risk domains like leisure and education, frequently neglecting mission-critical sectors such as healthcare and security. Furthermore, most systems merely pursue superficial user experience enhancements to drive passive acceptance, rather than granting algorithmic control. Only a minimal fraction of studies allow direct model adaptation or interactive training loops. Crucially, the literature remains dominated by theoretical frameworks rather than validated socio-technical deployments.
Strategic Takeaway
For enterprise leaders and socio-technical researchers, this review underscores that explainability alone cannot ensure trustworthy AI integration. Sustainable digital transformation requires shifting from passive transparency interfaces toward bidirectional Interactive AI frameworks. Leaders must redesign workflow architectures to incorporate Interactive Machine Learning, enabling front-line knowledge workers to steer parameters, provide causal corrections, and co-design decision logic. Treating users as active collaborators rather than evaluators mitigates algorithmic drift, curbs cognitive disengagement, and ensures accountability across complex enterprise operations.