Fair point but the article does mention that the model trains on reported crimes, and also suggests that Black homes tend to report crimes that White homes may not. So the training data has a bias, which is reflected in predictions.
Also, more police presence means more crimes are tracked. That doesn't mean they aren't happening in other areas, but without equal measurement, it may seem like areas have different counts.
So, worth considering if the training data is biased, and adjusting for that in the build and validation phases.