Introduction
As artificial intelligence becomes increasingly integrated into professional and personal decision-making, human attitudes are shifting from "algorithm aversion" toward "algorithm appreciation." However, this increased reliance brings new risks, particularly when navigating complex, nonlinear choices—decisions where variables interact in ways that make the optimal solution difficult to find. This study investigates how classic cognitive biases, specifically framing and anchoring, influence human performance and effort when collaborating with an AI teammate on such complex tasks.
Methodology
To simulate real-world complexity, researchers employed a rugged-landscape search task based on the NK model. Participants were tasked with "tuning" two dials, each with 24 possible settings, to find the highest "elevation" (score) on a hidden 2D grid. The landscape was intentionally "rugged," featuring multiple local peaks and valleys to prevent simple linear searching.
After two solo rounds to establish a baseline, participants were paired with an AI teammate that took control of one dial. The AI utilized a stochastic simulated annealing algorithm designed to suggest moves and collaborate toward a higher score. The study manipulated two key psychological factors:
- Framing: Participants were placed in either a "Gain" frame (accumulating points) or a "Loss" frame (starting with a high score and attempting to minimize deductions).
- Anchoring: Half of the participants received an "anchor"—a message indicating the maximum possible value achievable on the map.
Results
Analysis of 1,588 tasks from 400 participants revealed a significant performance drop when working with the AI. The average adjusted score with an AI teammate was 3.613, compared to 3.718 when working solo ($p = 0.003$). This decline was primarily driven by a shift in human behavior: participants submitted significantly fewer settings when working with the AI ($M = 30.37$) than they did alone ($M = 43.20$).
The Impact of Framing and Anchoring
The "Loss Frame" proved particularly detrimental. Participants focused on avoiding losses placed excessive faith in the AI, leading them to cease their own search efforts prematurely. This over-reliance resulted in lower final scores compared to those in the "Gain" frame.
However, the introduction of an anchor acted as a stabilizer. By providing an explicit benchmark for the maximum possible score, participants were better able to recognize when the AI had not yet reached the optimal peak, prompting them to continue searching rather than "checking out."
Discussion and Conclusion
The findings suggest that "algorithm appreciation" can easily devolve into counter-productive over-reliance, especially under negative framing. Humans tended to offload the tedious "exploitation" (fine-tuning) work to the AI, often before a true local maximum was reached.
This research highlights the need for AI systems that are "bias-aware." Future AI teammates could be designed to detect when a human partner is in a loss-avoidance mindset or is failing to explore sufficiently, proactively suggesting further search rather than simply providing a recommendation. By understanding these cognitive pitfalls, we can design more effective human-AI collaborative systems that mitigate, rather than amplify, human bias.
Summary updated