When Life Gives You AI, Will You Turn It Into A Market for Lemons? Understanding How Information Asymmetries About AI System Capabilities Affect Market Outcomes and Adoption

Paper Title

When Life Gives You AI, Will You Turn It Into A Market for Lemons? Understanding How Information Asymmetries About AI System Capabilities Affect Market Outcomes and Adoption

Publication Info

• Topic area: Information asymmetries in AI adoption and market dynamics.
• Keywords: AI adoption, information asymmetry, disclosure design, market for lemons, human-AI interaction, trust calibration, transparency, decision-making efficiency, regulatory frameworks, user reliance.

Background and Problem

• Problem / challenge: AI systems often exhibit hidden defects or unreliable performance, creating information asymmetries between suppliers and users. Current disclosure mechanisms are inconsistent and fail to address these gaps effectively.
• Significance: Addressing information asymmetries is critical for improving user trust, decision-making efficiency, and market outcomes in AI adoption.
• Motivation and related work: Previous research has highlighted the challenges of trust calibration and reliance on opaque AI systems. Regulatory efforts like the EU AI Act aim to improve transparency but lack empirical validation. This paper builds on economic theories of "market for lemons" to explore disclosure strategies in the context of AI adoption.

Solution

• Proposed approach: Experimental framework simulating AI adoption in a "market for lemons" with varying lemon densities and disclosure levels.
• Novelty:
  1. Adaptation of "market for lemons" theory to human-AI interaction.
  2. Systematic study of partial and full disclosure designs under varying densities of low-quality AI systems.
  3. Quantitative analysis of user behavior, reliance, and decision-making efficiency in AI markets.
• Procedure and key techniques:
  • Simulated market environment with three tasks: skin cancer prediction, loan approval, and deceptive review detection.
  • Between-subjects design with three lemon density conditions (low, medium, high) and two disclosure conditions (no disclosure, partial disclosure), plus a full disclosure benchmark.
  • Participants interact with a pool of AI systems, choosing between self-decision or delegation to AI, with feedback provided after each trial.
  • Metrics include delegation rates, coins earned, and reliance on low-quality AI systems.

Results

• Concrete findings:
  • Partial disclosure improves decision-making efficiency, reducing reliance on low-quality AI systems by significant margins.
  • Full disclosure leads to fewer delegations to lemons but exhibits persistent under-reliance on AI systems, resulting in efficiency losses.
  • Participants struggle to accurately estimate lemon density, leading to misaligned reliance behaviors.
• Advantage over baselines:
  • Partial disclosure offsets the negative impact of medium lemon density, improving efficiency compared to no disclosure.
  • Full disclosure achieves optimal avoidance of lemons but does not maximize AI adoption.
• Experiments / evaluation:
  • 330 participants recruited via Prolific, balanced across gender and age.
  • Tasks randomized across trials; lemon density and disclosure conditions varied systematically.
  • Mixed-effects logistic regression and OLS analyses used to evaluate delegation behavior and performance outcomes.
• Limitations and future work:
  • Static market dynamics; real-world AI markets may exhibit dynamic seller and buyer behaviors.
  • Simplified operationalization of AI quality; future studies should incorporate fairness, robustness, and safety indicators.
  • Binary delegation model; future research should explore graded reliance and verification mechanisms.

Summary

This paper investigates how information asymmetries and disclosure designs affect AI adoption and market outcomes. Using an experimental framework based on the "market for lemons" theory, the study demonstrates that partial disclosure significantly improves decision-making efficiency by reducing reliance on low-quality AI systems, while full disclosure eliminates information asymmetries but fails to maximize AI adoption due to persistent under-reliance. Results highlight the importance of interpretable and actionable disclosure cues for fostering appropriate reliance. The findings have implications for regulatory frameworks, emphasizing the need for enforceable yet lightweight disclosure rules to mitigate market inefficiencies and support user-centered AI adoption.