Human vs. Algorithmic Auditors: The Impact of Entity Type and Ambiguity on Human Dishonesty

TL;DR

This study investigates how human dishonesty varies when monitored by humans versus machines, especially under conditions of transparency and ambiguity, revealing that opacity in algorithms can increase dishonest behavior.

Contribution

It provides experimental evidence on how entity type and ambiguity influence dishonesty, highlighting the behavioral effects of algorithmic opacity in verification processes.

Findings

Cheating magnitude is similar under transparent verification for both humans and machines.

Ambiguous verification increases cheating when monitored by machines, leading to polarization between honesty and maximal dishonesty.

Opacity in algorithms can unintentionally promote severe dishonesty in verification contexts.

Abstract

While most of the existing literature focused on human-machine interactions with algorithmic systems in advisory roles, research on human behavior in monitoring or verification processes that are conducted by automated systems remains largely absent. Our study examines how human dishonesty changes when detection of untrue statements is performed by machines versus humans, and how ambiguity in the verification process influences dishonest behavior. We design an incentivized laboratory experiment using a modified die-roll paradigm where participants privately observe a random draw and report the result, with higher reported numbers yielding greater monetary rewards. A probabilistic verification process introduces risk of detection and punishment, with treatments varying by verification entity (human vs. machine) and degree of ambiguity in the verification process (transparent vs. ambiguous). Our results show that under transparent verification rules, cheating magnitude does not significantly differ between human and machine auditors. However, under ambiguous conditions, cheating magnitude is significantly higher when machines verify participants' reports, reducing the prevalence of partial cheating while leading to behavioral polarization manifested as either complete honesty or maximal overreporting. The same applies when comparing reports to a machine entity under ambiguous and transparent verification rules. These findings emphasize the behavioral implications of algorithmic opacity in verification contexts. While machines can serve as effective and cost-efficient auditors under transparent conditions, their black box nature combined with ambiguous verification processes may unintentionally incentivize more severe dishonesty. These insights have practical implications for designing automated oversight systems in tax audits, compliance, and workplace monitoring.