I had to go to the screenshots to see what you thought the expected answer was (left).

This is one of those questions that could have multiple answers, or require follow up questions, depending on how pedantic the asker wants to be.

Trick question, the island was in a lake, you’re nowhere near the ocean.

Trick question, it’s a small island and the ocean is all around you, not just on the left. How big must an island before this isn’t true? Is it a line of sight question?

You can't start reasoning by assuming the asker is pedantic; that's a defeatist mindset. The prompt is very clear to any reasonable human.

I got the wrong answer with Haiku and Sonnet 4.5

Correct answer with Sonnet 4.6, but this might as well be a coin flip. I've found Sonnet 4.6 to be substantially dumber than 4.5. I'd rate Sonnet 4.5 a 10/10 at creative writing and 4.6 a 3/10.

My ChatGPT just expired and I was about to get Claude instead, but I'm starting to rethink this.

If you project from the centre, as if the observer is trapped in the middle of the earth like a Jules Verne prisoner with x-ray vision watching the world above them, it is to the right (anti-clockwise to the left).

But no one thinks like that.

After testing whis, what strikes me is how stubborn the LLMs are about being wrong. Is that a more important takeaway: that LLMs seem to back down less even when clearly wrong?

This sounds like a classic topological brain teaser. If you’re driving clockwise on an island, the ocean remains on your left (in right-hand drive logic). But the real edge case is the 'island within a lake on an island' scenario. In 2026, with autonomous navigation, these recursive boundary definitions are actually a practical challenge for mapping algorithms. It's a great reminder that local geometry isn't always as simple as a single loop.