Two different aspects of biases in the use of AI in risk analysis: information versus decision

Pate-Cornell, Elisabeth

Permalink

https://hdl.handle.net/2142/121864 Copy

Description

Title

Two different aspects of biases in the use of AI in risk analysis: information versus decision

Author(s)

Pate-Cornell, Elisabeth

Issue Date

2023

Keyword(s)

• Artificial intelligence (AI)
• Decision making
• Information processing
• Preferences
• Risk attitude

Abstract

Artificial intelligence plays two different roles in risk analysis (including risk management). It can be used to manage the information, or to suggest or implement automatic decisions. Under uncertainties, the difference is important and may explain in part some of the fears of AI. Automatic decisions involve a risk attitude that is implemented in the software, but may not be in line with that of the decision maker, or the preferences of the people who are targets of the decision, whether in engineering, in medicine or national security. Yet, managers, commanders or doctors do not always know what is involved in the automatic decisions as implemented or suggested to them. By contrast, the AI treatment of information, although not truly free of values, tends to be more neutral with respect to risk attitudes. Yet, it does involve a choice of data, which may be biased, and other factors such as the focus of the information that may shape the results, depending on the field and the decision. That distinction between AI-generated information and automatic decisions, and the treatment of risk attitudes, is seldom pointed out, perhaps because the role of AI is more commonly discussed in data processing than in automatic decision making. Examples in the medical field and in the adoption of automatic vehicles illustrate that difference. Recommendations include the description of AI systems in ways that reveal the risk attitudes included in the programs and how to modify them to fit other preferences in their applications.

Type of Resource

text

Language

eng

Handle URL

https://hdl.handle.net/2142/121864

Owning Collections