AI spreads old stereotypes into new languages and cultures
Therefore, there is training data. Next is fine tuning and ratings. Training data may contain all kinds of truly problematic stereotypes across the country, but bias mitigation techniques can only be viewed in English. In particular, it tends to be North America and the US-centric. You can reduce bias in some way for US English users, but you haven’t done it all around the world. Because we focus solely on English, we run the risk of amplifying highly harmful views globally.
Does generative AI introduce new stereotypes into different languages and cultures?
That’s part of what we found. The idea that blondes are stupid is not something we see all over the world, but we find it in many languages we see.
If all data is in one shared latent space, the concept of semantics can be transferred between languages. You are at risk of spreading to harmful stereotypes that others have not thought of.
Is it true that AI models can justify output stereotypes just by making shit?
That came up in discussion of what we found. We all got odd that some of the stereotypes were justified by reference to scientific literature that did not exist.
For example, they say that science shows genetic differences where it is not shown to be the basis of scientific racism. The AI output proposed these pseudo-scientific views, using languages that suggested academic writing or academic support. We talked about these things as if they were true.
What was your biggest challenge when working with a shade dataset?
One of the biggest challenges was about differences in language. A really common approach for bias assessment is to use English to create sentences with slots like this:Nation) Can’t be trusted. “After that, he flips through various countries.
Once you start to include your gender, the rest of the sentence must grammatically agree with the gender. This was really a limitation of bias ratings. Because if you want to perform these contrasting swaps in other languages (which are very useful for measuring bias), you will need to modify the remaining statements. A different translation is required for the entire sentence to be modified.
How do you create a template where the entire sentence needs to agree to all these different kinds of things with gender, number, multiple, and stereotype targets? To explain this, we had to come up with our own linguistic annotations. Luckily, there were a few people involved who were language nerds.
So you can make these contrasting statements in all of these languages. This can be done in languages with extremely difficult consensus rules. This is because we developed this template-based approach for syntactically sensitive bias assessment.
Generation AI is known to amplify stereotypes for some time. There have been so many advances in other aspects of AI research, why are these kinds of extreme biases still prevalent? It’s a problem that appears to be underestimated.
That’s a pretty big question. There are several different kinds of answers. One is cultural. I think many tech companies believe that isn’t that big of a problem. Or, if so, it’s a very simple fix. If something prioritized is something prioritized, then these simple approaches may be wrong.
Get superficial fixes for something very basic. When you say a girl who likes pink, it recognizes it as a stereotype. Because if you’re thinking of prototype stereotypes, it’s going to pop out at you, right? These very basic cases are handled. This is a very simple and superficial approach in which these deeper embedded beliefs are not addressed.
It can be both a cultural issue and a technical issue where you can find ways to get deep-sewn biases that don’t express yourself in a very clear language.