For most machine learning systems, "performance equivalent to an expert human" is the best you can expect - simply because it's learned from training data labelled by expert humans.

So if a person looks much younger than their age in the judgement of an expert human, I wouldn't expect the model to do any better than that.

You should also know a lot of work in this area relies on photos of celebrities scraped from the internet - which is much easier than getting loads of labelled images of normal people in normal situations, which would be a total hassle practically and legally.

Of course that has some benefits - if you know the celebrity's date of birth and the date of the photo, you don't need to rely on human labelling to know the age of the person in the photo! But it has the major disadvantage that if your application doesn't involve professionally made up people with movie star looks in evening wear on red carpets - you might find real world performance falls short of the benchmark claims.

It depends on the quality of the image. It's most accurate for a square head-on photo, like a passport photo, and can be wildly off for occluded or angled photos. It's more accurate for people under 30, for older people it often underestimates age.

i would assume it CAN Be accurate if people 'look their age' but of course you can easily have someone who looks younger or older than their age that skews the overall etimate

Check out the README. They comment on the accuracy.