Before a model can learn to recognise something, examples usually must be labelled, marking which image contains what, or which message is spam. This annotation is painstaking human work, and its quality sets a ceiling on what the model can learn.
Inconsistent or biased labels teach the model to be inconsistent or biased in turn. Because of this, careful guidelines, agreement checks between labellers and spot audits matter as much to the final result as the choice of model.