Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
So there is training information. Then, there is a wonderful arrangement and evaluation. Training information can be all kinds of problematic stereotypes among countries, and then biased methods can only look in English. Especially, North America and USA tend to be centered. If you will reduce biases to English users in the United States, you did not do this in the world. You only strengthen really harmful landscapes in the world, according to your English.
Is generative AI presents new stereotypes to different languages and cultures?
This is part of what we found. The idea that blondes are stupid, is not anything found in the world, is found in the languages we look at.
When all your data in a shared secret space, semantic concepts can be transferred in languages. You risk promoting harmful stereotypes that others do not think so.
It is true that the AI models will sometimes justify the stereotypes with their performances by preparing the shit?
It was something that came out in the discussions of what we found. Some stereotypes said they referred to the scientific literature that some stereotypes did not exist.
The results showed genetic differences where the basis of science, scientific racism. The AI speeches made these false scientific views and then used the language that offers academic writing or academic support. It was as if it was not true but did not actually talk about it.
What were some of the biggest challenges when working in the Shades database?
One of the biggest challenges was around the language differences. A really common approach to Bias assessment is to use English and make a sentence with a nest: “People [nation] are invalid. “Then, you turn into different nations.
When you start to make sex, the rest begins to agree grammatically in sex. This is a restriction on this contrast in other languages, because if you want to exchange this contrast in other languages, it is very useful to measure this bias in other languages - you need to change the rest of this sentence. You need different translations that all sentences change.
How do you do the templates where the entire sentence needs to be agreed with all these different types of gender, majority and stereotype? To do this, we should have come to our own language annotation to give an account. Fortunately, several people with linguistic nerds took part.
Thus, now these opposite expressions can also do those who have the rules of the rules of this language, and have developed this novel, developed a template-based approach to the syntactic sensitive bias.
Generative AI, now it was known to strengthen stereotypes for a while. With so many progress in other aspects of the AI research, such extreme biases of this type still dominate? This is an issue that appears directed.
This is a pretty big question. There are several different answers. One is cultural. I think that many technological companies are believed that there is no big problem. Or, if you have, it is a fairly simple correction. If something is prioritized, it will be a priority, it is simple approaches that can go wrong.
We will get surface adjustments for many basic things. If you say girls like pink, you know as a stereotype, because if you are thinking of thinking prototypical stereotypes, isn’t it? These are very basic situations will be solved. This is a very simple, superficial approach to which the more deeply affected beliefs are not addressed.
It is a technical issue for learning how to make a cultural issue and a deep bias that does not express themselves in a very clear language.
