No I don't think this is a training issue. Light skin physically reflects more light, which gives cameras significantly more data to work with to detect shadows/etc.

The face recognition on a phone gets around that by creating their own light, with dot projection at a light wavelength the human eye can't see, but I've only ever heard of that being done for short range face recognition. CCTV cameras are too far away from the face and are not really accurate enough for anyone (including white people).

AFAIK Amazon's system was mostly intended for their self service retail stores... that's a different scenario entirely since you're only comparing faces to other customers who are in the store at the same time as you. And also the stakes are much lower - if two people appear to be the same person you can just flag both customers as needing to be verified by a staff member. No big deal at all.

Using it as evidence for a crime though, will inevitably result in false convictions.