The Moral LLM 

Photos © Aditya Mohan.

In legal terms, a person's moral character is defined by their adherence to societal norms and laws, reflecting their ability to distinguish right from wrong within the context of their community. Can this concept be extended to a Large Language Model (LLM) to make them better? The concept of including morality into LLMs has been tried with Anthropic's Claude LLM, which incorporates a unique approach known as Constitutional AI to embed moral principles into its framework. Constitutional AI is designed to imbue Claude with explicit values through a constitution-like document, rather than the implicit values derived from extensive human feedback. This approach aims to create a model that is not only safer and more helpful but also possesses a clear set of adjustable values. The constitution guides the LLM in making judgments about its outputs, encouraging behavior that avoids toxicity, discrimination, and illegal or unethical assistance, while promoting helpfulness, honesty, and harmlessness. However, integrating morality into LLMs presents challenges, as moral values are not universal and can vary significantly across different cultures and societies. For example, the U.S. Constitution itself, while a cornerstone of American legal and moral standards, reflects this variation in morality.

In October 1798. John Adams wrote:

Our constitution was made only for a moral and religious people. It is wholly inadequate to the government of any other.” 

Morality and virtue are crucial for a country to remain free and for society to function well. Virtue means having a deep personal commitment to doing what's right and choosing to follow moral principles voluntarily. It is important to note that Adams’ reference to religion is not about a single religious belief. Adams was strongly opposed to setting up an official religion and supported the freedom of religion. He and others who thought like him were Deists, meaning they believed that a moral system must be based on belief in God. However, the details of that belief or how it showed up in any specific religion didn't really matter much to them.

John Adams

This highlights the complexity of creating universally acceptable moral guidelines for AI systems. Even if a LLM doesn't need to be fine tuned for a religious belief, it still needs to follow a set of moral values and virtues, which are different across cultures and societies.


Further read