Lisp was more associated with the classical symbolic and logic-based AI approach which doesn't share much in common with generative AI. It questionable whether the later should even be called "AI" but that battle has been lost years ago.

Python is just a really good glue language and was good enough for the task.

A lot of Lisp was funded by AI research. That research dried up in the 90s (see "AI Winter"). Regular computers of the time were underpowered for Lisp, which originated on mainframes and then moved onto purpose-built hardware. So the new generation of programmers learned Pascal, C, and C++ instead and were mostly disconnected from the AI research that had come before.

Modern AI focuses on numeric methods, whereas classic AI was heavily symbolic. That makes Lisp less well suited for it, since symbolic computation is where Lisp shines. Classic AI does have some useful bits to it, but insofar as AI development goes it's mostly considered a dead end.

Lisp's market share was declining 30 years ago and only continued to decline. Python's has consistently risen in that same time. Also, Lisp offers little benefit, if any, when all the ANN implementations rely on C++ and CUDA code. You can write fast numerical code in Lisp, but it's not as straightforward and it's certainly not an idiomatic or common way to use Lisp. That could have let it compete with the C++ libraries, but wouldn't help with the GPU programming part. Lisp could have been the glue language like Python, but again, Python's popularity was on the rise and Lisp's on the decline.

Because worse is better. You cannot persuade dummies to use a better language, they'll always fallback to the currently worst language. If VB or currently python.