Or even without calling it 'AI', perceptual hashing like you typically see in applications like CSAM is pretty damn close to ML techniques. The normal thought

Or even without calling it 'AI', perceptual hashing like you typically see in applications like CSAM is pretty damn close to ML techniques. The normal thought process is "we don't want child predators to sneak through just by cropping, or sticking a water mark on the image, or some other way slightly modifying the image like the color balance. Can we come up with something that'll hash the same even with minor modifications to the image?". And you basically end up with intentionally overfit AI.