Machine learning in all its glory...

I still was surprised about how accurate it can get when you design and train your model well (or, like me, copy stuff from randomly selected ML papers).
I once had a model that detected preprocessed cards taken from screenshots of a competetive (and undisclosed) game of chance with 99.999% accuracy.
At the end i had the samples preprocessed down to 5X5 grayscale matrices. I wouldn't have been able to identify anything in the resulting blobs. But two convolutional and one fully connected neural network layers where enough to get it done. That thing was the opposite of deep and therefore fast to train on the CPU.

Machine learning works pretty well for classifying similar things to the things used in training.
If you want to detect non-matches and outliers, you have to train on non-matches and outliers.
If you need to detect unknown things, you have to train pretty extensively on all known things and then detect unknown things by low or multiple high confidence scores.