That's terrible logic. I'm reminded of the Dijkstra quote:

"Program testing can be used to show the presence of bugs, but never to show their absence."

If it gets lots of similar examples right then that would be a good sign.

Best would be to use some kind of a debugging tool to understand what representations it's using. We don't have access for GPT-4, unfortunately.