Credit: AP Illustration/Jenni Song
“Scaling up” is the catchphrase for the artificial intelligence industry, as tech companies race to use ever-increasing amounts of internet data to improve their AI systems.
This is also a red flag for Mozilla's Abeba Birhane, an AI expert who has long questioned the values and practices of her field and its impact on the world.
Her latest research finds that the scale of online data used to train popular AI image-generating tools is disproportionately increasing racist output, particularly against Black men.
Birhane is Senior Advisor for AI Accountability at the Mozilla Foundation, the nonprofit parent organization of the free software company that runs the Firefox web browser. Raised in Ethiopia and living in Ireland, he is also an Adjunct Assistant Professor at Trinity College Dublin.
Her interview with The Associated Press has been edited for length and clarity.
Q: What inspired you to get into the AI field?
A: I'm a cognitive scientist by training. Cognitive science doesn't have its own department wherever you study it. Where I studied it was in a computer science department. I was placed in a lab full of machine learning people. They were doing a lot of really cool stuff, but nobody was paying attention to the data. I found that really interesting and really fascinating, because I thought data was one of the most important factors for the success of a model. But I thought it was weird that people didn't pay that much attention or spend that much time on, “What's in my dataset?” That's how I got interested in the field. And eventually, I started auditing large datasets.
Q: Can you talk about your work on the ethical foundations of AI?
A: Everyone has an opinion about what machine learning is. Machine learners, i.e. people in the AI community, say machine learning is value-free — it's just math, objective, neutral, etc. Meanwhile, social science scholars say that like any technology, machine learning encodes the values of the people who drive it. So we systematically studied the 100 most influential machine learning papers to really find out what the field values, and we did it in a very rigorous way.
A: Was one of those values to be scaled up?
Q: Scale is thought to be the secret to success. Researchers at big companies like DeepMind, Google, and Meta claim that scale trumps and cancels out noise. The idea is that as you scale, everything in your dataset becomes even and balanced. And eventually you'll end up with something like a normal distribution, or something close to the truth. That's the idea.
Q: But your research explores how scale can cause harm. What are those harms?
A: As you scale these datasets, the issues they contain also scale, at least when it comes to hateful content, toxicity, etc. More specifically, in the context of our research, as you scale the datasets, the hateful content in them also scales. We measured the amount of hateful content in two datasets. As the datasets scaled from 400 million to 2 billion, there was an increase in hateful content, targeted content, and offensive content. This was a very conclusive finding that shows that the laws of scaling don't really hold when it comes to training data. We found (in a separate paper) that women and especially men with darker skin tend to be assigned suspicious and criminal labels at a much higher rate.
Q: How hopeful or confident are you that the AI industry will make the changes you propose?
A: These are not just mathematical and technical achievements. They are also tools to shape and influence society. In our recommendations, we also incentivize and pay attention to values such as justice, fairness, and privacy. To answer honestly, I am not at all confident that the industry will accept our recommendations. The industry has never accepted such recommendations that actually encourage it to take these social issues seriously, and probably never will. Companies and large corporations tend to act when they are legally required to do so. We need very strong and enforceable regulations. Companies also react to public outrage and public perception. They tend to make changes when they are in a situation where their reputation is damaged.
©2024 The Associated Press. All rights reserved. This material may not be published, broadcast, rewritten or redistributed without permission.
Quote: The AI industry is impacting the world, Mozilla adviser Abeba Birhane challenges its core values (July 22, 2024) Retrieved July 22, 2024 from https://techxplore.com/news/2024-07-ai-industry-world-mozilla-abeba.html
This document is subject to copyright. It may not be reproduced without written permission, except for fair dealing for the purposes of personal study or research. The content is provided for informational purposes only.
