AI models designed for structured output have a significantly lower potential for misbehavior than general-purpose models