Honest Ethics & AI – Part 1: The origins of morality
Why transformer-based LLMs can't handle moral reasoning, and how value alignment misses the mark.
In a multi-part essay series titled 'Honest Ethics & AI,' independent researcher Jesper L. challenges the prevailing approach to AI safety and ethics. The first part, 'The origins of morality,' argues that current AI systems, particularly transformer-based LLMs, are fundamentally unfit for making decisions with moral consequences. Jesper L. begins by deconstructing the circular definitions of 'moral' and 'ethics' found in dictionaries, tracing the origins of human morality back to the simple, pragmatic concepts of 'good' and 'bad.' He contends that these concepts are relative and grounded in real-world experience, something AI systems lack.
Jesper L. positions himself as an outsider—a biologist, not an ML scientist—offering a fresh perspective. He argues that value alignment is the wrong target for creating more ethical AI, as it focuses on aligning systems with human preferences rather than enabling genuine moral reasoning. The series will continue with parts on ethical reasoning, metaethics, and a new alignment paradigm. Jesper L. emphasizes the importance of moral vigilance and reality-grounded reasoning, suggesting that AI developers are already moving in the right direction by focusing on interpretability and robustness. This essay serves as a call for clearer moral thinking in AI development, starting with understanding the origins of human morality.
- Jesper L. argues transformer-based LLMs lack the capacity for moral judgment due to their absence of reality-grounded reasoning.
- The series defines morality as rooted in pragmatic concepts of 'good' and 'bad,' which are relative and experiential.
- Jesper L. suggests value alignment is the wrong target, advocating for moral vigilance and a new alignment paradigm.
Why It Matters
This perspective challenges AI developers to rethink ethics, focusing on human moral clarity over system alignment.