That’s actually not a good summary at all. You could have quality in, but apply to a different group than was trained on, and get garbage out (eg face recognition trained on one set of races or genders, then try to apply to different races to genders). Or you train on data that’s high quality but is the result of inequity (eg police profiling, or credit worthiness), and then you get inequity out of the model.

I think it's fair to say that one component of AI ethics is identifying these subtle forms of "garbage out" and the subtle forms of "garbage in" that cause them.

Garbage in means you have too much noise over signal coming in. You might be able to call current society garbage if you wish, but from an ML perspective it’s not accurate to say arbitrary input is garbage if it indeed is high quality data that reflects society as it is, or if you only have data on a certain set of classes but what you have is noiseless.

If you are studying criminality and using conviction record data, I think it's fair to say that the biased introduced by biased policing, biased laws, etc. is a form of "garbage-ness" in the data, in that it might not reflect the reality that you were trying to model.

Biased data is garbage if your aim is to have unbiased results...