This was well known even when these programs existed in the laboratory stage.

"Alice O’Toole’s group at the University of Texas, Dallas, trained facial discrimination networks similar to Cottrell’s on a training set that contained a relatively large number of oriental faces plus a small number of Caucasian faces. The trained network showed a success level similar to Cottrell’s but it was less able to discriminate among, or determine the sex of, novel faces that were Caucasian as opposed to Oriental. Although it performed just fine on Oriental faces, all Caucasian faces “looked pretty much the same” to it. A second training experiment this time with African faces in the majority over Oriental, yielded a network that performed fine on novel African faces, but poorly on Oriental faces . . . a third experiment, this time with a few African faces scattered on the large number of Caucasians, yielded a similar result (Paul Churchland 1996 p. 52)"

more on this here. https://teedrockwell.medium.com/the-neuroscience-of-racism-why-some-people-think-they-all-look-alike-775f3462094d