Neural networks for data selection in particle physics have been used since before the current AI hype cycle took off. This means there are already some established guidelines on how to mitigate and conservatively estimate pontential biases. I certainly hope the IceCube researchers have followed them.