C-046Value Alignment and AI EthicsConfidence: Medium

Micah Carroll, Matija Franklin & Hal Ashton, Beyond Preferences in AI Alignment

Tan Zhi-Xuan (2025)

One-Sentence Thesis

The authors challenge the dominant preferentist approach to AI alignment, which assumes that human preferences are an adequate representation of human values, and argue for a reframing of AI alignment to focus on normative standards and social roles.

Argument Outline

  1. 1Introduction to the preferentist approach to AI alignment
  2. 2Critique of rational choice theory as a descriptive model of human decision-making
  3. 3Critique of expected utility theory as a normative standard of rationality
  4. 4Alternative approaches to AI alignment, including alignment with normative standards and social roles

Key Distinctions

Preferentist approach vs. alternative approaches to AI alignment
Descriptive vs. normative accounts of human decision-making

Key Terms

Preferentist approach
An approach to AI alignment that assumes human preferences are an adequate representation of human values
Rational choice theory
A theoretical framework that assumes human behavior can be modeled as the maximization of expected utility
Expected utility theory
A normative standard of rationality that assumes rational agents maximize expected utility

Flashcards

17 cards

Related Questions

4

Which of the following does Tan Zhi-Xuan criticizes in "Micah Carroll, Matija Franklin & Hal Ashton, Beyond Preferences in AI Alignment"?

3

In Tan Zhi-Xuan's "Micah Carroll, Matija Franklin & Hal Ashton, Beyond Preferences in AI Alignment", Garrabrant criticizes which of the following?

4

In Tan Zhi-Xuan's "Micah Carroll, Matija Franklin & Hal Ashton, Beyond Preferences in AI Alignment", London, A. J. defines which of the following?

3

In Tan Zhi-Xuan's "Micah Carroll, Matija Franklin & Hal Ashton, Beyond Preferences in AI Alignment", Baker et al. develops which of the following?

3

In Tan Zhi-Xuan's "Micah Carroll, Matija Franklin & Hal Ashton, Beyond Preferences in AI Alignment", van Rooij et al. provides which of the following?

4

What is the main critique of the preferentist approach to AI alignment?