It’s not like instructions are the only way to enforce rules on an LLM. Just have its output pass through a filter that checks it for the rules. If it passes, the response moves on to the user. If it fails the filter, the LLM gets prompted to try again.
LLMs cannot differentiate between data and instructions. Such laws would be no better than the “guardrails” that major companies already try to use.
It’s not like instructions are the only way to enforce rules on an LLM. Just have its output pass through a filter that checks it for the rules. If it passes, the response moves on to the user. If it fails the filter, the LLM gets prompted to try again.
How do we implement said filter?