Removing a file from a dataset affects future training. A model that learned from the file may still reflect patterns acquired while it was present.
A team assessing unlearning can compare the resulting model with one trained without the selected data. It also needs to check that useful behavior was preserved.
ELI5
Think of removing an ingredient from stored supplies after it has already been used in a recipe. Removing the unused ingredient does not change the dish already made.
For example, deleting a training file does not by itself remove what a model learned from that file; changing that influence is a separate problem.
Does dataset deletion change a model immediately?
It changes the available training data. Weights already learned remain until the model is updated or replaced.
Why is verifying unlearning difficult?
Training examples can influence many model behaviors, and a few tests may fail to reveal effects that persist.

