Really, your response should be that the video game use case is easier to detect going off track. It's a lot more feasible to detect when Random Peasant #2154 in Skyrim is breaking the fourth wall than a generic chatbot.

The exact same scenario as the article could happen with an NPC in a game if there's no/poor guardrails. An LLM-powered NPC could definitely start insisting that it's a real person that's in love with you, with a real address you should come visit right now, because there's not necessarily an inherent difference in capability when the same chatbot is in a video game context.