Morality in AI. A plea to embed morality in LLM architectures and frameworks

Current AI safety approaches may be insufficient: The paper argues that existing methods like fine-tuning and reinforcement learning from human feedback are "bottom-up" approaches that may not adequately address moral reasoning in LLMs, suggesting a need for more fundamental architectural changes to how these systems process moral meaning.

Morality should be built into AI design from the ground up: Rather than trying to patch moral behavior onto existing systems, the authors propose embedding moral processing directly into the core architecture and design principles of transformer-based models, similar to how human cognition is fundamentally structured.

Attention mechanisms are key to moral processing: The paper draws parallels between human attention and LLM attention mechanisms, proposing that how these systems "pay attention" to information is central to how they handle moral questions—suggesting that redesigning attention could improve moral reasoning.

Philosophy can inform AI architecture: The authors apply Iris Murdoch's concept of "loving attention" (careful, compassionate observation) as a philosophical framework for understanding what moral processing should look like in AI systems, bridging humanities and technical design.

This remains an early-stage exploration: The paper acknowledges significant limitations in its technical operationalizations, indicating that embedding morality in LLM architectures is still a developing area requiring further research and refinement.