[Paper Review] AI, Meet Human: Learning Paradigms for Hybrid Decision Making Systems

Clara Punzi, Roberto Pellungrini+3 more|arXiv (Cornell University)|Feb 9, 2024

Complex Systems and Decision MakingDecision Sciences3 citations

TL;DR

This paper proposes a comprehensive taxonomy of hybrid decision-making systems, categorizing them into three learning paradigms: Human Oversight, Learn to Abstain, and Learn Together. It synthesizes existing literature to provide a conceptual and technical framework for modeling human-AI interaction, emphasizing mutual learning, trust, and performance optimization in high-stakes applications.

ABSTRACT

Everyday we increasingly rely on machine learning models to automate and support high-stake tasks and decisions. This growing presence means that humans are now constantly interacting with machine learning-based systems, training and using models everyday. Several different techniques in computer science literature account for the human interaction with machine learning systems, but their classification is sparse and the goals varied. This survey proposes a taxonomy of Hybrid Decision Making Systems, providing both a conceptual and technical framework for understanding how current computer science literature models interaction between humans and machines.

Motivation & Objective

• To address the lack of a unified classification for human-AI interaction in machine learning systems.
• To identify and formalize key learning paradigms that enable effective collaboration between humans and AI in decision-making.
• To analyze the strengths, limitations, and open challenges in current hybrid systems, particularly regarding communication, human effort, and system adaptability.
• To support future research by providing a structured framework for designing trustworthy, transparent, and effective hybrid decision-making systems.
• To highlight critical gaps in current systems, such as poor communication of uncertainty and lack of adaptability to heterogeneous human users.

Proposed method

• Proposes a formal taxonomy of Hybrid Decision-Making Systems based on the role of human and machine in the learning and decision process.
• Classifies systems into three paradigms: Human Oversight (human verifies AI output), Learn to Abstain (human knowledge is embedded in data for model adaptation), and Learn Together (humans actively correct AI reasoning).
• Analyzes existing works under each paradigm, identifying core techniques such as uncertainty estimation, active learning, and feedback-based fine-tuning.
• Introduces a conceptual framework that models the joint behavior of human and machine agents, including their computational steps and interaction dynamics.
• Evaluates systems through a lens of trust, transparency, fairness, and reliability, emphasizing bidirectional learning and mutual understanding.
• Identifies technical and human-centered challenges, such as lack of interpretability in human feedback and insufficient adaptation to diverse user types.

Experimental results

Research questions

• RQ1 How can human-AI interaction in decision-making systems be systematically categorized and formalized?
• RQ2 What are the key differences and trade-offs between Human Oversight, Learn to Abstain, and Learn Together paradigms in hybrid systems?
• RQ3 How do current hybrid systems handle uncertainty, trust, and mutual learning between humans and machines?
• RQ4 What are the major open challenges in communication, human effort, and adaptability within hybrid systems?
• RQ5 How can hybrid systems be designed to support diverse human users and prevent misuse or degradation from adversarial or inconsistent inputs?

Key findings

• The taxonomy of three learning paradigms—Human Oversight, Learn to Abstain, and Learn Together—provides a foundational framework for understanding and designing hybrid decision-making systems.
• Current systems often fail to communicate uncertainty or reasoning to humans, limiting trust and effective collaboration, especially in Human Oversight and Learn to Abstain settings.
• Human-AI collaboration is hampered by high human labor costs, particularly in data annotation, with limited mechanisms to reduce this burden.
• Existing systems are largely monolingual and inflexible, lacking the ability to adapt to heterogeneous human users or handle multiple experts effectively.
• Malicious or conflicting human inputs are not well-modeled or protected against, exposing systems to poisoning and degradation risks.
• Bidirectional learning—where both humans and machines learn from each other—is essential for building trustworthy, reliable, and adaptive hybrid systems.