I think you're asking some very different questions. Questions about what is racist and why are very different from questions about why ai designers don't understand all of the internals of their ais.

Racism has multiple definitions. Some involve intent and others don't. That complicates your question. l'm sure I've seen racism in ML models discussed before.

Who decided that race should be an input to the model? And who decided what "race" each individual is? There is a fairly large component of social construct in race.

@JonathanZ : it's more a matter of not having excluded race from the data than of including it. If the sample data is biased, which it is going to be if it is an unfiltered snapshot of the web, the results will be. This does cause error is both directions: someone recently got a racially diverse group when they asked a system to generate a photo of Nazis, as a result of an attempt to correct for this. The whole point of these systems is to generalize, and they are subject to (and subjected to) the problems of over generalization.

Notice that we just assume that race is there in the data. And that's because race is deeply embedded in our society. And, as you point out, there are lots of noxious racist beliefs embedded in there with them. It feels like the question is coming from an assumption that "the AI" is coming from a place of pure deduction, and doesn't realize how much of recent models are built on stuff people have produced, and plenty of it is rather dumb and prejudiced.

I often wondered about positive attribution given to a race... as an example "I love the Japanese, such a studious, calm, wise people"... (hypothetical example, of no claimed accuracy)... am I being "racist"? Am I implying other races don't meet those criteria, and by raising one, lowering others by unspoken default? I haven't quite decided in my mind, where I stand on such. Again noting I have nothing for, or against the Japanese or any other race, and was offering only a hypothetical, so there may be no confusion!

"If a human tries to generalize of a bad action based on a race, it's called racism." Not necessarily. If you can establish via observation a pattern of people of Lithuanian ancestry jaywalking 75% more frequently than people descended from Ecuadorians, that might be a statistically valid enough to be considered factual, but it isn't racism. Taking action on that information by profiling and targeting Lithuanians for random stops would likely be considered racist. They are two very different things, and machines are so far incapable of the second action.

@AlistairRiddoch, Matt Walsh's film "Am I Racist?" may help your perspective if you are confused on what is racist and what isn't. (Spoiler alert: The answer depends on your race!)

It very often is considered a form of racism or institutional racism when that happens.

Seems most answers answer "if it's racist for humans, why isn't it for machines?". Sadly the more interesting question is:"If it's not racist for machines, then why is it for humans?". Also sadly the question is not clear enough to point to the core of the matter.

Who says it's not called racism? Go and look at the case where the Google started identifying pictures of black people as "gorillas".

@DJClayworth, that would better be described as a "mistake"...

@JimmyJames, A book cannot be racist. An author might be racist, but the physical or electronic medium by which their racist themes might be published cannot be.

I think legalized racism is one of the top selling points not only of AI but other related information technologies. Consider "social credit" schemes, which essentially read the tracks of historical racism to reconstruct a parallel network of denigration that can be used wherever and whenever traditional racism would have been, for equivalent effect. A naive view of capitalism leads some to think that is useless, but merchants know that if they collectively bully a group of people (car dealers selling to women) they can force them to accept bad trades for lack of alternatives.

Consider the true meaning of the word "discrimination": to tell one thing from another. We use the word to refer to what we consider to be an immoral and act of biased discrimination (racism, sexism, ageism, ...). However, that's not the only meaning. If I tell you to find a needle in a haystack, and you find it, you discriminated the needle from the hay. Machine learning discriminates good choices from bad choices (based on what it is told is good/bad output), but it does not inject a bias on its own. Any bias comes from the provided measure of what is good/bad output that it learns from.

@JimmyJames "to be racist" is not the same as "to carry a racist message". The paper and ink are not to blame for the message that the author constructed using said paper and ink. It might seem silly for me to imply that you're saying that paper is sentient and is racist; but AI similarly has no sentience and cannot be attributed racism - it's just easier to forget that distinction when AI sometimes appears to behave lifelike, which a piece of paper never does.

I agree with the idea that we should be putting the "blame" (although I would use "responsibility") for racism on the creator and/or user of an AI/ML process rather than the algorithm itself... However, the sentence "to carry a racist message" has already imbued "message" with the quality "racist". I think it's clear that an inanimate thing can be called "racist" when you are attributing its creator/user with that sentiment, intentionally or otherwise.

A definition of “racism” is severely missing from this question. The word is used in many contexts with vastly different meanings. Sometimes it’s anything that ties a specific property to a specific group of people, i.e. racial stereotyping (“Mexicans love spicy”). Other times it specifically refers to purposeful negative discrimination regarding a specific group (“blacks are not allowed in this coffee”). Occasional incorrect usage even considers personal preferences racist (“black girls are not my type, I find white blondes more pretty”.) It should be defined what level of usage do you mean.

@Flater I would disagree: sentience is not required for a system to be racist, and carry out racist actions. POSIWID, so any system that leads to discriminatory outcomes is discriminatory. To avoid AI confusion: A flowchart can similarly be racist if step one is "if black, deny them"