"The best achieved accuracy of over 99.8% across all demographics".
As with any technology, not all systems are created equal. In the studies conducted by NIST, while some of the algorithms tested only had a 70% accuracy rate, or had problems identifying women or those from ethnic minorities, the best achieved accuracy of over 99.8% across all demographics. The institute anticipates that this will improve yet further as researchers expand and refine the datasets used by the deep-learning technology.
It’s worth mentioning, however, that the success and accuracy of any facial recognition system are dependent on how it is used and where it is located. The camera will need to clearly see the face presented for verification, so good lighting is important. Similarly, it should be positioned in such a way that accidental verification does not take place from someone passing behind. While the set-up of a facial recognition access control system may take a little work, the benefits are more than worth it.