Not long ago someone use the Google “Photos” app to use artificial intelligence to go through a large number of photos and classify them automatically. Anyone who has taken hundreds of photos can appreciate the temptation to let someone (or something) else look after the sorting-out.
Most of the classifications were fairly predictable. One of them caused Google to issue a sincere apology for the unintended racism of the category applied to one photo.
“We’re appalled and genuinely sorry that this happened”
Google apologises for Photos app’s racist blunder – BBC News
Elsewhere in the story we learn the app sometimes classified dogs as horses. No doubt other errors also appear on a regular basis.
Anyone familiar with writing computer programs or the idea of fuzzy logic can appreciate the general problem.
Consider this example. On a scale of “1” to “4” for “birdness” (i.e. how good an example of a bird is this creature?) rate: chickens, bats, eagles, puffins, emus, ….
Bats?
Bats are not birds but a significant number of people will rank them with a “5”. Just outside the “bird” boundary. Why? Since they fly with wings and so do most birds the ranking illustrates the proximity of “bats” to the general category of “birds” along the crucial dimension of being able to fly.
The dimensions used in the classification process plus the precision with which each dimension is represented is the basis of being “correct” or “incorrect”.
In this case the software got the “primate” part correct but after that it did not get the type of primate correct.
What is really interesting here is not whether the person who wrote the software should get at least a B+ (but obviously not an A+) for the effort.
It’s the apology. It’s about “being sorry” for something which neither involves any kind of intentionality on the part of the person making the statement nor even any realistic chance the person offering the apology had any way of averting the event. It’s even about how to classify something as “racism” — a socially constructed term with ethical connotations — instead of a “classification error” as it relates to image processing software. In other words: “causality”.
It’s also about identity, individuality, group membership, and collective guilt.
And luck.
Everyone presently alive exists because their ancestors survived. Some of those ancestors were possibly saints while others were possibly psychopathic monsters. Each of the ancestors shaped the trajectory of history in such a way as to allow all of us to be here today. In legal doctrine there is a concept known as “the fruit of a poisoned tree”. It means evidence which is collected illegally cannot be used in court.
But what about the seeds of the fruit of the seeds of the fruit of the poisoned tree?
Are the children of war criminals guilty of war crimes? The grandchildren? Do we have to wait for ten generations?
And what about the rest of the people: those whose psychopathic monster ancestors have not yet been identified by the descendants of their victims?
It’s very difficult to discuss these ideas with examples which directly relate to present-day allegations of collective guilt, collective entitlement, collective identity, and each individual’s sense of self.
These matters often arise in the discussion of matters concerning preferential policies for hiring, promotion, and other forms of “restorative justice” when various social engineering devices are introduced in order to correct in the present injustices which happened in the past.
The counter-arguments normally involve claiming that those who get the benefits did not suffer the original injustice while those paying the price did not commit the original misdeeds. Victimhood gets inherited by one individual through the means of cultural transmission while guilt is inherited by others. In these cases it is clear that the concept of “culture” is assumed to be very real indeed.
It is interesting to notice how “cultural inheritance” arguments are considered gauche or invalid when discussing jihadis.
Which brings up the underlying feature of this episode. It yields yet more examples for “locally rational” or “partially rational” worldviews.
The major aspects of the worldviews will be the point of view of the individual who holds it, the dimensions and boundaries for the worldview, and the various statuses and classifications for those entities and processes which make the worldview function.
The narrator of the worldview may be speaking for individual or collective advantage, the general purpose of the narrative is either to enhance or diminish an idea or the identity of another, and the means for doing this can be emotional, logical, or the introduction of things purporting to be historical facts.
Recalling Aristotle’s categories for Rhetoric, these are appeals to emotion, appeals to tribal or ethnic values or advantage, and appeals to the more objective ideals of “logic and evidence”.
Thus we can have a Google executive classify a software glitch as “racism”.
Or we can have 25-year-old members of Culture A demanding apologies from 25-year-old members of Culture B for events which took place 75 years ago. And maybe they need to do this in order to maintain the respective traditions of both of their cultures.
Maybe it’s not “rational” for some cultures to mingle with others.
Maybe it’s easier for us to discuss these topics when the substantive details are furnished by software errors and science fiction stories.
Unjustified Margins 