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
Source: https://arxiv.org/abs/2601.21650 ↗
Full text: arXiv preprint ↗
Erlei and colleagues apply Akerlof's classic 'market for lemons' framework to AI system adoption, addressing a critical gap in understanding why organizations struggle to evaluate AI capabilities.
The information asymmetry problem — where AI vendors know more about system limitations than buyers — creates adverse selection dynamics that explain both AI hype cycles and adoption failures.
For product directors, this provides a theoretical foundation for understanding why AI procurement often disappoints and why signaling credible capability becomes so important.
The work bridges classic information economics with contemporary technology adoption, offering both diagnostic tools and strategic insight into how markets actually evaluate complex technical capabilities.
Central argument
Through a controlled experiment simulating an AI product marketplace, Erlei et al. demonstrate that information asymmetries about AI system quality produce 'market for lemons' dynamics: when users cannot distinguish capable from defective AI systems, they systematically miscalibrate reliance — under-delegating when lemons are rare, over-delegating when they dominate. Crucially, partial disclosure (revealing only accuracy scores) significantly improves decision quality without increasing overall delegation rates, effectively offsetting a doubling of low-quality systems in the pool. Yet full disclosure paradoxically underperforms: even when users have complete information, they delegate only 58% of tasks to high-quality systems, losing ~20% of achievable performance — suggesting risk aversion and bounded rationality constrain AI adoption independently of information availability.
Critique
The study operationalizes AI quality through just two signals — accuracy and data quality scores — while explicitly setting aside fairness, robustness, and uncertainty, which are precisely the dimensions most likely to be strategically obscured by real-world providers. This principled simplification means the findings on partial disclosure effectiveness may overstate how well users can act on disclosed information in authentic markets, where the relevant quality dimensions are noisier, more contested, and harder to reduce to a single hoverable metric. The gap between laboratory disclosure design and the fragmented, provider-gamed transparency landscape the authors themselves describe in the introduction remains unresolved by the experimental results.
Why it matters for product
For a CPO deciding how to surface AI capabilities to end users — whether in a B2C product or an internal tool — this research argues that disclosure design is a first-order product decision, not a compliance afterthought: the specific signals you expose (and their cognitive load) determine whether users mis-rely on AI features, independent of how good those features actually are. The finding that partial disclosure matches full disclosure in outcome quality also has direct implications for roadmap prioritization: investing in cleaner, actionable single-metric transparency (e.g., a confidence or reliability indicator) may deliver more adoption efficiency than building comprehensive model documentation that users cannot act on under real task pressure.