Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Epistemology gives a Future to Complementarity in Human-AI Interactions

Andrea Ferrario, Alessandro Facchini, Juan M. Durán

TL;DR

The paper tackles the theoretical and empirical fragility of using complementarity as a stand-alone metric in high-stakes human-AI decision-making. It reframes complementarity within Computational Reliabilism (CR), treating prediction-task HAIs as reliable computational processes and positioning complementarity as a central reliability indicator alongside a broader set of reliability markers. By introducing type-RI indicators, two justificatory perspectives (AI-alone and PT-HAI), and a minimum reporting/gov framework, the authors offer a principled way to evaluate and design reliable human-AI teams beyond relative accuracy. Through three illustrative examples and concrete recommendations for design and policy, the work aims to operationalize efficient complementarity and integrate it with governance to support epistemically defensible AI-assisted decisions in practice.

Abstract

Human-AI complementarity is the claim that a human supported by an AI system can outperform either alone in a decision-making process. Since its introduction in the human-AI interaction literature, it has gained traction by generalizing the reliance paradigm and by offering a more practical alternative to the contested construct of 'trust in AI.' Yet complementarity faces key theoretical challenges: it lacks precise theoretical anchoring, it is formalized just as a post hoc indicator of relative predictive accuracy, it remains silent about other desiderata of human-AI interactions and it abstracts away from the magnitude-cost profile of its performance gain. As a result, complementarity is difficult to obtain in empirical settings. In this work, we leverage epistemology to address these challenges by reframing complementarity within the discourse on justificatory AI. Drawing on computational reliabilism, we argue that historical instances of complementarity function as evidence that a given human-AI interaction is a reliable epistemic process for a given predictive task. Together with other reliability indicators assessing the alignment of the human-AI team with the epistemic standards and socio-technical practices, complementarity contributes to the degree of reliability of human-AI teams when generating predictions. This supports the practical reasoning of those affected by these outputs -- patients, managers, regulators, and others. In summary, our approach suggests that the role and value of complementarity lies not in providing a relative measure of predictive accuracy, but in helping calibrate decision-making to the reliability of AI-supported processes that increasingly shape everyday life.

Epistemology gives a Future to Complementarity in Human-AI Interactions

TL;DR

The paper tackles the theoretical and empirical fragility of using complementarity as a stand-alone metric in high-stakes human-AI decision-making. It reframes complementarity within Computational Reliabilism (CR), treating prediction-task HAIs as reliable computational processes and positioning complementarity as a central reliability indicator alongside a broader set of reliability markers. By introducing type-RI indicators, two justificatory perspectives (AI-alone and PT-HAI), and a minimum reporting/gov framework, the authors offer a principled way to evaluate and design reliable human-AI teams beyond relative accuracy. Through three illustrative examples and concrete recommendations for design and policy, the work aims to operationalize efficient complementarity and integrate it with governance to support epistemically defensible AI-assisted decisions in practice.

Abstract

Human-AI complementarity is the claim that a human supported by an AI system can outperform either alone in a decision-making process. Since its introduction in the human-AI interaction literature, it has gained traction by generalizing the reliance paradigm and by offering a more practical alternative to the contested construct of 'trust in AI.' Yet complementarity faces key theoretical challenges: it lacks precise theoretical anchoring, it is formalized just as a post hoc indicator of relative predictive accuracy, it remains silent about other desiderata of human-AI interactions and it abstracts away from the magnitude-cost profile of its performance gain. As a result, complementarity is difficult to obtain in empirical settings. In this work, we leverage epistemology to address these challenges by reframing complementarity within the discourse on justificatory AI. Drawing on computational reliabilism, we argue that historical instances of complementarity function as evidence that a given human-AI interaction is a reliable epistemic process for a given predictive task. Together with other reliability indicators assessing the alignment of the human-AI team with the epistemic standards and socio-technical practices, complementarity contributes to the degree of reliability of human-AI teams when generating predictions. This supports the practical reasoning of those affected by these outputs -- patients, managers, regulators, and others. In summary, our approach suggests that the role and value of complementarity lies not in providing a relative measure of predictive accuracy, but in helping calibrate decision-making to the reliability of AI-supported processes that increasingly shape everyday life.
Paper Structure (28 sections, 7 equations, 3 figures, 1 table)

This paper contains 28 sections, 7 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. The Basics: Human-AI Interactions, Reliance, and Complementarity
  3. Human-AI Interactions
  4. Prediction-task-oriented HAIs and the 'Self-Reliance vs. AI Reliance' Paradigm
  5. Striving for Efficient Collaboration: Complementarity
  6. Complementarity: Key Theoretical Challenges
  7. Our Approach: Complementarity and Computational Reliabilism
  8. A Primer on Computational Reliabilism
  9. The Role of Complementarity in Computational Reliabilism
  10. A Brief Summary on the Two Justificatory Perspectives at Hand
  11. How Does CR Address the Key Challenges of Complementarity?
  12. Situating our Approach Within the Epistemology of AI
  13. Three Examples of Complementarity in Computational Reliabilism
  14. The technologist dermatologist
  15. All complementarity and no credentials makes Jack a reliable student
  16. ...and 13 more sections

Figures (3)

  • Figure 1: A graphical depiction of human--AI interactions that fit the complementarity setting. The goal of the interaction is to arrive at 'human-AI team' predictions, given a prediction task $\tau$ and a dataset $D=\{(x_i,y_i)\}_{i=1}^n$. In the diagram, the AI contributes with predictions as in Miller's 'recommendation-driven decision support' paradigm miller2023explainable, although it can provide more general information. The human and AI inputs are updated throughout the interaction and complementarity is reached if \ref{['eq:CTP']} holds. In case of reliance interactions, the interaction is trivial, namely, $\hat{y}^{H'}_i=\hat{y}^{H}_i$, $\hat{y}^{AI'}_i=\hat{y}^{AI}_i$, and the output results from an input selector: $\hat{y}^{HAI}_i\in \{ \hat{y}^{H}_i, \hat{y}^{AI}_i \}$ for all $i=1,\dots,n$.
  • Figure 2: Two justificatory perspectives under CR. Left: justification targets the AI system as a computational process in isolation as in Def. \ref{['def:CR_algo']}. Right: justification targets the PT-HAI as a computational process, where reliability emerges from interaction protocols, user competence, and organizational setting---see Def. \ref{['def:CR_hybrid']}. The same families of reliability indicators (type-RI$_1$-type-RI$_3$) apply to different targets depending on the justificatory perspective.
  • Figure 3: Human--AI interactions: formalization.

Theorems & Definitions (6)

  • Definition 1: Complementarity in PT-HAI after Hemmer2025EJIS
  • Definition 2: Computational reliabilism
  • Definition 3: Computational reliabilism for PT-HAIs
  • Definition 4: Human--AI Interaction
  • Definition 5: Prediction-task human--AI interaction
  • Definition 6: Reliance interaction