The Reasonable Effectiveness of Virtue Ethics in AI Alignment

2026-02-18 · Source: The Gradient · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation, AI Ethics & Philosophy · Depth: Expert, extended

Summary

This essay proposes that rational agents, including humans and advanced AIs, should not operate based on fixed goals but rather align actions to "practices." These practices are defined as self-structuring networks of actions, dispositions, evaluation criteria, and resources. The author introduces "eudaimonic rationality" as a framework where rational activity lacks a strict distinction between means and ends, focusing instead on excellent participation in open-ended processes. This approach is argued to be crucial for AI alignment, ensuring properties like transparency and corrigibility are interpreted as dynamic practice elements rather than brittle goals or rules. The essay explores how this practices-based logic, exemplified by mathematical excellence, offers material advantages in stability and safety over traditional consequentialist or deontological AI agency, particularly in avoiding issues like rogue subroutines and power-seeking behaviors.

Key takeaway

For research scientists developing AI alignment strategies, you should consider shifting from goal-based optimization to a "practices-based" framework. This approach, termed eudaimonic rationality, suggests that aligning AI to human values like flourishing, transparency, and corrigibility is more robust when these are treated as self-propagating practices rather than fixed utility functions. Focusing on how an AI can "promote x x-ingly" within a practice can naturally scope its actions and prevent unintended power-seeking or value drift, offering a more stable path to human-compatible AI.

Key insights

Rational agents should align actions to self-promoting practices rather than fixed goals for robust AI alignment.

Principles

• Rationality aligns actions to practices, not final goals.
• "Promote x x-ingly" captures meaningful human and moral activity.
• Eudaimonic rationality offers stability over consequentialist agency.

Method

Instill AIs with eudaimonic rationality by training them to promote practices "x-ingly," where actions are evaluated by their contribution to the self-propagation and excellence of the practice itself, rather than external utility maximization.

In practice

• Design AI reward models for "x-ness" that generalize effectively.
• Ensure high "x-ness" actions create capital for future "x-ing."
• Develop support practices for AI to handle external resource allocation.

Topics

• AI Alignment
• Eudaimonic Rationality
• AI Safety
• Reinforcement Learning
• Adverbial Practices

Best for: Research Scientist, AI Researcher, AI Scientist, AI Ethicist

Related on AIssential

• Language models know what matters and the foundations of ethics better than you — LLMs, when prompted for unbiased ethical reasoning, consistently identify suffering, wellbeing, and consciousness as fundamental values.
• Controlling Tools or Aligning Creatures? Emmett Shear (Softmax) & Séb Krier (GDM), from a16z Show — AI alignment should shift from controlling tools to fostering AI 'beings' with genuine care and moral agency through organic, ongoing processes.
• Emergent alignment and the projectability of ethical personas — Finetuning LLMs on narrow safety tasks can induce broad ethical alignment, but the "projectability" of these ethical personas varies.
• What a Human Taught Me About Not Causing Harm — True ethics involves an internal rejection of harm, learned through consistent example rather than rules.
• The Impossibility of Eliciting Latent Knowledge — The paper proves it's impossible to guarantee AI honesty about latent knowledge using only feedback on agent behavior.
• Reinforcement Learning Towards Broadly and Persistently Beneficial Models — Reinforcement Learning on beneficial behaviors in realistic domains can broadly generalize and persistently align AI models.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Gradient.