Right, but 538 did account for such an error in coming up with their ~75/25% outcome chances. Had polling errors just been random, the overall prediction would have been more like 95/5%.

But their key insight is that polling errors are often correlated, and not just statistical sampling errors that can randomly offset in either direction.