Even with extremely noisy input, given a consistent asymmetry, the true signal can still stand out.

The random noise tends to cancel out with enough scale, and the true signal–however faint–is all that's left.

How much scale you need depends on how clearly the signal stands out from the status quo.

Very clear benefit: works at low scale.

Very weak benefit: needs massive scale.

Related: the smaller the dataset, the more precise and principled your extraction algorithm has to be.

If you have a massive dataset, then something that has a small asymmetry is sufficient.