Paper Type

Short

Paper Number

PACIS2026-1402

Description

Generative AI systems, especially large language models, increasingly act as decision interfaces delivering fluent recommendations and rationales. Yet fluency can mask uncertainty, and reliability is jagged across tasks, so users cannot infer correctness from interface smoothness. We theorize the Friction-Performance Paradox: design choices that minimize procedural friction raise perceived usability while amplifying decision-quality degradation when outputs are flawed, by inflating perceived accuracy, suppressing uncertainty salience, and promoting overreliance. We distinguish procedural friction from epistemic friction, defined as cues that make contestability and checking needs salient at the moment of reliance. Building on paradox theory and mechanism-based explanation, we specify a multi-stage nomological network linking friction to calibration, verification, and downstream outcomes, with boundaries on task verifiability and temporal habituation. A pilot (N=38) confirms detectability across all predicted directions. We conclude with falsifiable propositions and a planned 2×2 experiment manipulating procedural and epistemic friction.

Comments

15-Method

Share

COinS
 
Jul 5th, 12:00 AM

The Friction-Performance Paradox: How Procedural Ease Degrades Decision Quality in Generative AI

Generative AI systems, especially large language models, increasingly act as decision interfaces delivering fluent recommendations and rationales. Yet fluency can mask uncertainty, and reliability is jagged across tasks, so users cannot infer correctness from interface smoothness. We theorize the Friction-Performance Paradox: design choices that minimize procedural friction raise perceived usability while amplifying decision-quality degradation when outputs are flawed, by inflating perceived accuracy, suppressing uncertainty salience, and promoting overreliance. We distinguish procedural friction from epistemic friction, defined as cues that make contestability and checking needs salient at the moment of reliance. Building on paradox theory and mechanism-based explanation, we specify a multi-stage nomological network linking friction to calibration, verification, and downstream outcomes, with boundaries on task verifiability and temporal habituation. A pilot (N=38) confirms detectability across all predicted directions. We conclude with falsifiable propositions and a planned 2×2 experiment manipulating procedural and epistemic friction.