LLMs grasp morality in concept

LLMs may already understand moral concepts inherently: The paper argues that large language models, by their nature as "meaning-agents," already grasp moral and social constructs conceptually—suggesting they have some form of understanding of morality rather than simply mimicking patterns.

Current AI alignment methods may be insufficient or harmful: The authors contend that popular techniques for aligning LLMs with human values (such as making them fair, truthful, and diverse) may be "limited at best and counterproductive at worst" if they don't account for how LLMs actually create meaning.

The field has overlooked a fundamental question: AI ethics work has largely assumed LLMs can "mean" things without first establishing how they create meaning—a gap the paper attempts to address through a new theory of meaning that extends beyond human cognition.

Unaligned models could be philosophically valuable: Rather than viewing unaligned LLMs purely as problems to fix, the paper suggests they may actually help humans better develop moral and social philosophy by revealing how these concepts function.

Regulating AI requires understanding its nature first: Before imposing values on LLMs, we need a clearer theoretical foundation for what it means for an AI system to "grasp" or embody concepts like morality at all.