When noise mitigates bias in human-algorithm decision-making: An agent-based model.

Area of Science:

• Decision Sciences
• Computational Social Science
• Human-Computer Interaction

Background:

• Algorithmic systems are increasingly used in critical decision-making processes.
• Algorithms offer low variability (noise) compared to human judgment, but their interaction with human noise is poorly understood.
• The impact of biased algorithmic advice on noisy human judgment requires investigation.

Purpose of the Study:

• To investigate how biased algorithmic advice interacts with noisy human judgment.
• To determine if human noise mitigates or exacerbates algorithmic bias.
• To explore the implications of noise in human-algorithm decision-making systems.

Main Methods:

• Agent-based modeling was employed to simulate decision-makers' judgments.
• The model incorporated scenarios with biased algorithms and biased/noisy human advisors.
• Simulations analyzed the effects of varying levels of bias and noise on decision outcomes.

Main Results:

• Contrary to expectations, human noise can mitigate algorithmic bias by reducing reliance on algorithmic advice.
• Noise in human advice increases reliance on prior beliefs, impacting belief updating.
• Polarized prior beliefs lead to asymmetric responses to advice, favoring interventionist cues.

Conclusions:

• The absence of noise in algorithmic advice is not universally desirable; human noise can be beneficial.
• Human noise can serve as a buffer against algorithmic bias.
• Population-level variability in decision-making may stem from environmental noise, challenging traditional noise audit assumptions.