If an AI’s output could be extrapolated entirely from its training data, we wouldn’t have competing AI systems. One AI algorithm would be deterministic.
If different AI can produce different results from the same training… then the training set cannot be the *entire* explanation for bias.
