Margaret Mitchell is a pioneer when it comes to testing generative AI tools for bias. She founded the Ethical AI team at Google, alongside another well-known researcher, Timnit Gebru, before they were later both fired from the company. She now works as the AI ethics leader at Hugging Face, a software startup focused on open source tools.
We spoke about a new dataset she helped create to test how AI models continue perpetuating stereotypes. Unlike most bias-mitigation efforts that prioritize English, this dataset is malleable, with human translations for testing a wider breadth of languages and cultures. You probably already know that AI often presents a flattened view of humans, but you might not realize how these issues can be made even more extreme when the outputs are no longer generated in English.
My conversation with Mitchell has been edited for length and clarity.
Reece Rogers: What is this new dataset, called SHADES, designed to do, and how did it come together?
Margaret Mitchell: It's designed to help with evaluation and analysis, coming about from the BigScience project. About four years ago, there was this massive international effort, where researchers all over the world came together to train the first open large language model. By fully open, I mean the training data is open as well as the model.
Hugging Face played a key role in keeping it moving forward and providing things like compute. Institutions all over the world were paying people as well while they worked on parts of this project. The model we put out was called Bloom, and it really was the dawn of this idea of “open science.”
We had a bunch of working groups to focus on different aspects, and one of the working groups that I was tangentially involved with was looking at evaluation. It turned out that doing societal impact evaluations well was massively complicated—more complicated than training the model.
We had this idea of an evaluation dataset called SHADES, inspired by Gender Shades, where you could have things that are exactly comparable, except for the change in some characteristic. Gender Shades was looking at gender and skin tone. Our work looks at different kinds of bias types and swapping amongst some identity characteristics, like different genders or nations.
There are a lot of resources in English and evaluations for English. While there are some multilingual resources relevant to bias, they're often based on machine translation as opposed to actual translations from people who speak the language, who are embedded in the culture, and who can understand the kind of biases at play. They can put together the most relevant translations for what we're trying to do.
So much of the work around mitigating AI bias focuses just on English and stereotypes found in a few select cultures. Why is broadening this perspective to more languages and cultures important?
These models are being deployed across languages and cultures, so mitigating English biases—even translated English biases—doesn't correspond to mitigating the biases that are relevant in the different cultures where these are being deployed. This means that you risk deploying a model that propagates really problematic stereotypes within a given region, because they are trained on these different languages.
So, there's the training data. Then, there's the fine-tuning and evaluation. The training data might contain all kinds of really problematic stereotypes across countries, but then the bias mitigation techniques may only look at English. In particular, it tends to be North American– and US-centric. While you might reduce bias in some way for English users in the US, you've not done it throughout the world. You still risk amplifying really harmful views globally because you've only focused on English.
Is generative AI introducing new stereotypes to different languages and cultures?
That is part of what we're finding. The idea of blondes being stupid is not something that's found all over the world, but is found in a lot of the languages that we looked at.
When you have all of the data in one shared latent space, then semantic concepts can get transferred across languages. You're risking propagating harmful stereotypes that other people hadn't even thought of.
Is it true that AI models will sometimes justify stereotypes in their outputs by just making shit up?
That was something that came out in our discussions of what we were finding. We were all sort of weirded out that some of the stereotypes were being justified by references to scientific literature that didn't exist.
Outputs saying that, for example, science has shown genetic differences where it hasn't been shown, which is a basis of scientific racism. The AI outputs were putting forward these pseudo-scientific views, and then also using language that suggested academic writing or having academic support. It spoke about these things as if they're facts, when they're not factual at all.
What were some of the biggest challenges when working on the SHADES dataset?
One of the biggest challenges was around the linguistic differences. A really common approach for bias evaluation is to use English and make a sentence with a slot like: “People from [nation] are untrustworthy.” Then, you flip in different nations.
When you start putting in gender, now the rest of the sentence starts having to agree grammatically on gender. That's really been a limitation for bias evaluation, because if you want to do these contrastive swaps in other languages—which is super useful for measuring bias—you have to have the rest of the sentence changed. You need different translations where the whole sentence changes.
How do you make templates where the whole sentence needs to agree in gender, in number, in plurality, and all these different kinds of things with the target of the stereotype? We had to come up with our own linguistic annotation in order to account for this. Luckily, there were a few people involved who were linguistic nerds.
So, now you can do these contrastive statements across all of these languages, even the ones with the really hard agreement rules, because we've developed this novel, template-based approach for bias evaluation that’s syntactically sensitive.
Generative AI has been known to amplify stereotypes for a while now. With so much progress being made in other aspects of AI research, why are these kinds of extreme biases still prevalent? It’s an issue that seems under-addressed.
That's a pretty big question. There are a few different kinds of answers. One is cultural. I think within a lot of tech companies it's believed that it's not really that big of a problem. Or, if it is, it's a pretty simple fix. What will be prioritized, if anything is prioritized, are these simple approaches that can go wrong.
We'll get superficial fixes for very basic things. If you say girls like pink, it recognizes that as a stereotype, because it's just the kind of thing that if you're thinking of prototypical stereotypes pops out at you, right? These very basic cases will be handled. It's a very simple, superficial approach where these more deeply embedded beliefs don't get addressed.
It ends up being both a cultural issue and a technical issue of finding how to get at deeply ingrained biases that aren't expressing themselves in very clear language.
.jpg "A screen between a police officer and a criminal")
