constitution

A constitution is a written set of rules that tells an AI system which questions it's allowed to answer and which ones it needs to block. Think of it like the training manual a company hands to new employees. If the manual only has vague guidelines, employees get confused and end up rejecting even harmless requests just to be safe. AI safety systems work the same way: when the constitution is written poorly, the system starts mistaking normal, safe questions for dangerous ones and blocks them too.

That's why developers go back and rewrite this document, build new training data based on the revised standard, and retrain the program that makes these judgment calls. In effect, they're refining the manual itself so it can tell ambiguous cases apart more accurately.

In the article, it's described as: "they rewrote the constitution — the rule set behind the classifier — rebuilt the training data around it, and retrained the classifier." Here, constitution doesn't mean a national charter; it refers to the internal rulebook that defines what an AI should treat as dangerous.