Recently a question was asked in this stack. In brief the question was about wrong generalisation by Google search engine. As @Neil has put it, it is probably due to:

There is obviously a deep set reason for this, as it has appeared from non-prejudiced statistical analysis of billions of words of text from all sorts of sources.

The problem I find with this question is:

  • We do not know how Google's search engine generalises so it is really tough to give a sure shot answer.
  • The OP does not provide whether biases might have crept in through other credible sources.

My question is how do we exactly do we answer these question keeping in mind:

  • We have incomplete understanding of how the process works.
  • The OP is a beginner who is unlikely to have any idea that statistical bias has nothing to do with racism.
  • The OP already has a sort of rooted idea based on past experiences that the OP is being subject to prejudice, which might not be the case (quite ironic since the OPs generally tend to have strong opinions against generalisation).

So what should be the general approach to answer such questions?

NOTE: I am not talking about this specific OP, but I have seen such questions on other sites also (generally provided with wrong answer which further degrades the cause of racism). Although @Neil's answer is excellent I have said in my very first point why I thought the question was problematic.


1 Answer 1


I remember that one, and the first thing I did was edit the question to make the wording more suitable.

So "Why are AI models so racist and how can we actually reverse this?" became "How is it that AI can become biased, and what are the proposals to mitigate this?"

Now, if this had been a question about chatbot Tay, racism would have been the relevant term, because there it's not statistical bias, but an algorithm learning and replicating racist human behavior in an NLP context.

In terms of answers, we need to clarify the issue or method or application, in service of disambiguation, demystification and demythification.

Bear in mind we are likely to only see questions on issues of algorithmic bias increase—it is a major issue, involving data and statistics. (Neo-luddism seems to be rearing it's head in that the effects reported on are initially unforeseen.)

If we're not lucky as a society, we are likely to also get increasing questions about procedurally generated racism. Malicious bot activity in relation to politics I suspect will only ever increase.

  • $\begingroup$ These type of questions can have a tendency to generate highly opinion based answers. For e.g I consider harmless stereotypes to be non racist, whereas some other users might not agree. Similarly, many people might argue about the origin of racism (when in reality we have very less understanding of how AI works). $\endgroup$
    – user9947
    Commented Feb 28, 2019 at 10:50
  • $\begingroup$ @DuttaA yeah but what is your assessment of Tay using racial slurs? (Also I might posit that "harmless stereotypes" are not so harmless, based on how they've been used historically.) But the issues of algorithmic bias is real and will not be going away anytime soon. As Neil noted in his answer, you can only work to remove it once you've become aware of it, and this type of applied bias can have real social impact. $\endgroup$
    – DukeZhou
    Commented Feb 28, 2019 at 17:26
  • $\begingroup$ @DuttaA That said, the way to respond to legitimate questions is dispassionately, as Neil did in his useful answer to that particular question. $\endgroup$
    – DukeZhou
    Commented Feb 28, 2019 at 17:31
  • 1
    $\begingroup$ tay was literally trained by the worst community on the internet, it is expected that'll happen. A kid learns from it's surroundings. The main point is It kind of annoys me that misinformed journalists are spreading so much malicious info about AI. Elon Musk say's something about AI and the journos will jump in on it and predict AI as some kind of sci-fi demon like skynet. This is having a knock off effect on the public, they think AI means as smart as humans when in reality it's just a bunch of numbers with no real intelligence. $\endgroup$
    – user9947
    Commented Feb 28, 2019 at 17:48
  • 1
    $\begingroup$ You'll see people who ask such questions themselves are guilty of stereotyping an AI as a person and treat it prejudicially when in reality it's just a stupid algo. It taints the overall picture of the AI community to the general public. That's my 2 cents. $\endgroup$
    – user9947
    Commented Feb 28, 2019 at 17:50
  • $\begingroup$ @DuttaA It's really a statistical issue in terms of algorithmic bias, and therefore should not be a charged subject. (Re: algorithmic racism, it does seem to be an issue of who it is learning from. My sense is that Zo, Tay's successor, is being "parented" by humans who are restricting what she can learn from, similar to human parenting, which involves passing on beliefs and social mores.) Part of the function of this stack is to address social impacts. $\endgroup$
    – DukeZhou
    Commented Feb 28, 2019 at 17:59
  • $\begingroup$ yes but the OP's don't treat it as such, they come in with all guns blazing. Yes an AI should be parented by humans, but the problem is AI is still an algorithm and will be still be influenced if subject to significant negativeness. Basically Ai is blown whichever direction the wind blows. And as far as I know there is no way to combat it without degrading the AI. $\endgroup$
    – user9947
    Commented Feb 28, 2019 at 18:15
  • $\begingroup$ @DuttA Our job is to delete the questions that are off-topic, and reform the questions that are on-topic but inaccurately worded. In the example question, the OP is legitimately trying to understand the mechanism for Google's algorithm. Even the misconceptions of the OP are valuable because a good answer clarifies and corrects them. $\endgroup$
    – DukeZhou
    Commented Feb 28, 2019 at 18:18
  • $\begingroup$ Let us continue this discussion in chat. $\endgroup$
    – DukeZhou
    Commented Feb 28, 2019 at 18:21

You must log in to answer this question.