This document explores the strengths and limitations of human and AI decision-making, using fraud detection as a case study, and examines how augmented intelligence can combine the best of both.
This document examines the interplay between human and AI decision-making, focusing on fraud detection, confidence curves, cognitive bias, and the benefits of augmented intelligence that combines both human judgment and AI recommendations.
When a decision must be made, who should make it—a human or an artificial intelligence (AI)? While humans outperform AI at some tasks, AI statistically excels at others. The answer is not always clear-cut and often involves a nuanced combination of performance curves and human bias.
Consider a fraud detection system that generates thousands of alerts daily. Financial analysts review each alert, but 90% are false positives, overwhelming the analysts. An AI system can help by handling some alerts, but which should be processed by AI and which by skilled analysts?
To answer this, imagine a graph with the y-axis as success rate and the x-axis as confidence score. A confidence score of 0% means the prediction is definitely not a real alert; 100% means it is certainly real. A typical AI performance curve shows high success rates at very low and very high confidence scores, but lower success when the AI is unsure—effectively, the AI is saying “I don’t know.”
Human performance curves are generally flatter. Humans may not match the accuracy of a highly confident AI, but they often outperform AI when the AI is uncertain, especially around the 50% confidence level. Humans can bring in additional information, context, or consult colleagues, while AI sticks to its programmed logic and data.
When an AI assigns a high or low confidence level, it will statistically outperform a human in determining if an alert is real or a false positive. However, for cases where the AI is unsure, humans can make better decisions by leveraging context and external knowledge. This is not a zero-sum game; the best results often come from combining both approaches.
Augmented intelligence merges human decision-making with AI recommendations. Its performance curve typically falls between those of AI and humans alone, especially for cases with moderate confidence scores. This approach can yield the highest success rates for a significant number of predictions.
For augmented intelligence to be most effective, human cognitive bias must be considered. The way AI recommendations are presented to humans significantly influences decision quality. For example, forced display (showing the AI’s recommendation alongside each case) can lead to automation bias, where humans overly trust the AI and ignore their own judgment.
Optional display, where the AI recommendation is only shown when requested, helps overcome automation bias by allowing humans to form their own initial impressions. Trust is also affected by how AI confidence is communicated; if an AI recommendation is accompanied by an accuracy percentage, humans may be less likely to trust it, regardless of the actual accuracy.
The optimal approach to decision-making often involves a blend of human and AI strengths. By understanding confidence curves, cognitive bias, and the dynamics of augmented intelligence, organizations can design systems that maximize accuracy and effectiveness in complex tasks like fraud detection.
(2) When AI is unsure, humans can make better decisions by using additional information and context.
| Display Method | Impact |
|---|---|
| Forced Display | 1. Can lead to automation bias |
| Optional Display | 2. Allows independent human judgment first |
Forced Display-1, Optional Display-2.
Augmented intelligence, which combines human and AI decision-making, can achieve higher success rates than either alone, especially for cases with moderate confidence scores.
True. Augmented intelligence leverages the strengths of both humans and AI for optimal decision-making.