So there are training data. Then there is a refinement and evaluation. Training data may contain all kinds of really problematic stereotypes in different countries, but then the techniques of relieving bias can only look at English. In particular, it tends to focus on North America and the USA. Although you can somehow reduce bias for English users in the US, you haven’t done it all over the world. You still risk strengthening really harmful views around the world because you focused only on English.
Does generative artificial intelligence introduce up-to-date stereotypes to various languages and cultures?
This is part of what we find. The idea that blondes are stupid is not something that can be found all over the world, but is in many languages we looked at.
When you have all the data in one shared latent space, semantic concepts can be transferred in different languages. You risk promoting harmful stereotypes that others didn’t even think about.
Is it true that AI models sometimes justify stereotypes in their exits, just shit?
It was something that came out in our discussions about what we found. We were a bit strange that some stereotypes were justified by references to scientific literature that did not exist.
The results saying that, for example, science showed genetic differences in which it was not shown, which is the basis of scientific racism. AI results presented these pseudo-scientific views, and then used the language that suggested academic or academic support. He talked about these things, as if they were facts when they were not actual at all.
What were some of the biggest challenges when working on a set of shades?
One of the biggest challenges concerned language differences. A really common approach to assessing prejudices is to exploit English and make a judgment with a gap: “People with [nation] They are unbelievable. “Then you throw yourself into various nations.
When you start putting sex, now the rest of the sentence begins to agree grammatically on gender. It really was a limitation of bias, because if you want to perform these contrasting swaps in other languages - which is very useful for measuring bias – you need to change the rest of the sentence. You need various translations in which the whole opinion changes.
How to create templates in which the entire opinion must be agreed by gender, number, in many of these different things with the goal of the stereotype? We had to come up with our own language annotation to explain it. Fortunately, a few people were involved who were language nurses.
So now you can perform these contrasting statements in all these languages, even those with rules with really complex compatibility, because we have developed a up-to-date, template -based approach to the assessment of the deviation, which is syntactically sensitive.
It is known that generative AI has been strengthening stereotypes for some time. With such high progress in other aspects of AI research, why are this type of extreme prejudice still common? This is a problem that seems insufficiently accepted.
This is quite a great question. There are several different types of answers. One is cultural. I think that in many technology companies it is believed that this is not such a large problem. Or, if so, it is a fairly elementary correction. What is prioritized, if something is priority, are these elementary approaches that can go wrong.
We will get superficial corrections for very basic things. If you say girls like a pink, he recognizes it as a stereotype, because it is something you think about prototype stereotypes, pops up to you, right? These very basic cases will be served. This is a very elementary, superficial approach in which these deeper embedded beliefs are not resolved.
After all, it is both a cultural and technical issue of finding how to get deeply rooted prejudices that are not expressed in a very brilliant language.