It's my understanding that the error rates are so high with these systems even in good conditions that they are next to useless.
This is more well known in the context of medical testing. Suppose you are screening for a rare disease. You are testing a group of 10 000 people, of which 100 have the disease and 9 900 do not.
Let's say you have a test that is right 99% of the time.
Of the 100 people you test who actually have the disease, your test will catch 99 of them.
Of the 9 900 people you test who do not have it, your test will falsely report that 99 of them have the disease.
The result is you get 198 people reported as having the disease, but only 50% of them actually have the disease. The other 50% are false positives.
If the test were right 90% of the time rather than 99% of the time, the results would be 90 positives who have the disease, and 990 positives who do not. Only 8.3% of the people who test positive actually have the disease!
If you used that 90% test on a population where 5 000 of 10 000 have the disease, things are dramatically different. 90% of the people who test positive actually have the disease.
Same thing with facial recognition. Set up a facial recognition system at a big football game and look for people who have outstanding arrest warrants, and you are in a situation like the medical test for the rare disease. A 90% accurate system might give you 10 times as many false positives as true positives.
Set up a system at a border that has a lot of one-day tourist traffic both ways, and use it to try to match up people coming back at the end of the day with people who left earlier in the day, and it might be more like the 50/50 disease. A 90% accurate test might only have a 10% false positive rate.