Looks like exactly what you described. Make no assumptions. The model doesn't have a clue about the implied tactical firearms that a human might assume. It doesn't understand anything about culturally understood contexts.
"hallucinations" happens in artificial intelligence systems. Similarly, when you toss a coin, it will usually end either tail or head ... but it can also end standing on its ridge.