cathexis - Thanks for sharing the video. Initially, it's amusing to see Gemini "chicken out". Further reflection brings the realization (as noted in a number of comments) that this example shows the limitation of using a Large Language Model AI in an area where its training is inadequate.

I believe a few comments mentioned hallucinations, which reminded me of an October 2025 interview of ChessBase's Matthias Wüllenweber, where he says AI "struggles with the chess middlegame due to its hallucinations. Chess is simply too complex."

https://en.chessbase.com/post/in-chess-ai-fails-due-to-its-hallucinations-an-int...

Imagine my surprise when, a few weeks later, the latest version of his company's flagship product, ChessBase 26, is announced to have an AI feature that will take a middlegame position and suggest a plan. In describing the features of ChessBase 26 in a November interview, Wüllenweber presented the AI feature in a tone of voice that suggested to me that it was added partly for entertainment value. In the example he showed, the AI did present a reasonable-looking plan, although at one of point Wüllenweber warned about the possibility of hallucinations.

https://en.chessbase.com/post/matthias-wuellenweber-on-all-new-functions-for-che...