Why is everyone suddenly talking about DuckDuckGo

Contributions to «All of a sudden, for computers, 'Mexicans' and 'illegally' belonged together"

Debiasing is and will remain for long manual labor. What is described in the interview with the word closeness is "unsupervised learning" namely learning based on examples without knowing what they mean.

A simple example: we can take 10,000 pictures of cats and 10,000 pictures of dogs and label them all with "dog" and "cat". This is how the AI ​​learns what dogs and cats look like. If we also see that there are dachshunds and huskies etc. in there and not just the same Border Collie from the neighbor, the network actually generalizes pretty well if we then show pictures of cats and dogs that it has not yet seen. This is "supervised learning" because we labeled the pictures beforehand.

Labeling is not always possible because of the large amount of data, as it would be manual work. That is why "unsupervised learning" is often used. So we show the network unlabelled pictures and it finds out the differences on its own.
It may be that a dachshund and a husky are more different than cats and dogs. So there could simply be 2 groups, one with dachshunds and cats and one with huskies and other large dogs.
Or worse: A group of cats and dogs that are petted by a human hand and those that are not petted.

These were all harmless examples, but it seems obvious that systemic discrimination very easily finds its way into such systems. You then have to manually build in a "penalty" for the network if it finds a connection between negative terms and the gender, appearance, origin of a person. You have to find these connections first, then classify them as undesirable, and then punish them. It can also be automated, but BigTech is apparently not that advanced yet.