How AI Researcher Dylan Baker Uses Technical Communication to Reduce Algorithmic Harm

Dylan Baker is an engineer and researcher currently advocating for and building toward more just algorithmic futures at the Distributed AI Research Institute. In this interview, they sketch out alternative pathways for AI development that weave together different ways of knowing and that value building with the communities that stand to be impacted by emerging technologies. Read on to learn more from Dylan about their strategies for effectively and transparently communicating technical ideas and for reducing harms propagated by biased AI systems.

Hannah Scott: Can you describe your journey through the fields of engineering and artificial intelligence to your current work around building toward and advocating for fairer, more just technological futures?

Dylan Baker: I’m Dylan, and I’m a research engineer at the Distributed AI Research Institute. I went to an engineering school where I studied machine learning and computer science. I ended up at Google working on a team that did experimental machine learning product development. The team was half research, half engineering. I did a lot of data work, model evaluation, and integrating models into different experimental products. But I really wanted to get involved in ML fairness. After a few years, I joined the Ethical AI team to do data engineering, which is when I got involved in research. The team was amazing, they were incredible to work with and I learned so much. But then, in 2022, Timnit Gebru was fired, and then Margaret Mitchell was fired in this really awful, shady way, which really had a chilling effect on everyone. Even though the team was amazing, it was also really hard for us to get recognized for our work both inside the company and sometimes externally. Even when the Ethical AI team had really huge impacts, it was often sidelined. Timnit and Margaret had been fighting tirelessly for our team’s work to get recognized and integrated into the company. It became harder to see a career for myself there, especially as a non-researcher trying to advocate for this kind of ethics work. When Timnit reached out a year later that she was starting DAIR I immediately said, “Absolutely!”

HS: DAIR is setting a prime example of what developing technology outside the incentives of industry and academia can look like. Can you speak to the importance of developing alternative infrastructures through which to conduct research? What are some successful strategies for making such work possible? How do decentralized organizations like DAIR make room for ways of knowing that fall outside strictly technical ones?

DB: DAIR has been really unique compared to the places I’ve worked before, because having independent funding means that we can set priorities and evaluate successes in a really different way. That allows us to take stances that we might not be able to take if we were tied to other forms of funding. It takes the pressure off from needing our impact to look a particular way and opens us up to be able to critique AI from a wide range of perspectives and angles. Ultimately, when we think about what a more ethical development of this technology could look like, I think it looks a lot like machine learning before 2015, or before the last 10 years. It looks like building systems for specific purposes with specific scopes and with specific goals. Then you can start to ask questions like “What values do we want the system to reinforce?” “What data makes sense to use here?” It’s the opposite of the “move fast and break things” philosophy. It’s really hard to advocate for that kind of work because it’s not flashy. It’s frustrating because I think none of this was inevitable. We’ve been talking about it in the field for ages, these issues with general or universal AI systems. I think one of the biggest arguments that we make at DAIR is that we don’t need to be building systems this way, we do not need to be making general purpose AI, we don’t need to be making these kinds of generative AI systems. It’s really hard to go down this line of work without independent funding, because that’s not where the money is right now.

HS: I want to zoom in on the narrowly-scoped applications of AI versus claims to universal or general human intelligence, which in itself is a bit of a misnomer. I think there’s a lot of correspondence here with some of your work around dispelling this kind of veneer of objectivity. In your interactive piece “Datasets Have Worldviews”, you point out that machine learning datasets are necessarily specifically situated in certain cultures and values. What might it look like to build artificially intelligent systems that acknowledge the situated-ness and inherent assumptions of data, rather than making claims to objectivity and universality?

DB: Some of it definitely just comes down to building things for specific purposes. On the harm reduction side, we can also build avenues toward transparency. A lot of people are advocating for making training data accessible, transparent, and searchable, or establishing guardrails on how large models are used. There are lots of harms that we can and are anticipating, and have anticipated. When we’re able to provide transparency and insight into the histories, the contexts of these models, it’s so much easier for not just researchers in this niche, but also the general public to reason about what these models will be useful for and how they might be harmful. With the right information, I think people can make good, informed decisions. But it’s really hard. There’s no incentive for the companies developing these models to make them transparent and open to the regulation and scrutiny that I think they need to be subjected to. Right now, we want to be making the argument that these extremely large models don’t need to exist, that we could regulate them out of existence. But also, if that isn’t happening, here are guardrails and initiatives for transparency and explainability that can help mitigate some of the harms.

HS: A lot of your work focuses on communicating the complicated nature of computing, which helps make these systems, which are increasingly infiltrating our day-to-day lives, more legible and transparent. What strategies for making complex systems more readily understandable have you used and found successful?

DB: Yeah, communication is something I think about a lot, and I’m learning a lot from other fields — science communication, education, design, and journalism are all places where there are a lot of different skill sets for figuring out how to communicate things effectively. At the core, I’m always trying to invest myself in the language, the values, and the goals of the people I’m trying to communicate with. I try to identify where I am, and then see where my ideas fit into my audience’s frameworks. For the “Datasets Have Worldviews” piece, my original idea came from a paper called “Do Datasets Have Politics” by Morgan Klaus Scheuerman, which goes into depth about what dataset documentation tells us about the underlying values of computer vision as a field, which was just so cool. It’s a really rich paper. I wanted to get into all of it in one article. But I realized it’s the kind of thing that just would not have been seen as relevant or important to the pure ML engineering crowd that I had been coming from. So I tried to take the very simplest concept from it — that datasets have politics — and publish it alongside all of these other articles that explain very specific technical papers in ML using the same visual language, using the same format, going for the same audience. But the focus was, “Hey, you are engineers, you care about accuracy, well, the politics of datasets is something that impacts accuracy.” For me, communicating these ideas always starts with letting go of the pressure to get folks 100% on the same page as me, and trying to identify a couple of things that I want to move the needle on. Then it’s much simpler. From there, it comes down to researching your audience, identifying good metaphors and visuals, and getting feedback.

HS: With “Datasets Have Politics,” you’re weaving this thread between social science, critical algorithmic theory, and machine learning research and engineering. In your experience, what possibilities does bringing together these different disciplines, these different ways of knowing, afford in terms of research outcomes?

DB: Incorporating multiple perspectives is something I’ve learned a lot from Timnit and that she’s really fostered at DAIR. We’ve taken a really broad view of whose expertise is considered valuable, which extends all the way to our fellows, who aren’t always people who come from academic or corporate research. Adrienne Williams, for example, is someone who organizes with Amazon drivers. Krystal Kauffman organizes with Amazon Mechanical Turk workers. Meron Estefanos is a self-taught journalist. Having a really wide range of voices in the room is just totally invaluable, and forces a lot of cross-cultural communication that is really helpful. This interdisciplinary perspective opens up other aspects of our work that I think are important, which means that our impact isn’t all going to look like academic papers, because that’s not where all of us come from. It means that we can all develop interdisciplinarity, which is really hard to do in so many spaces because a lot of large companies are not going to foster developing different disparate skill sets in parallel. But we need that sort of interdisciplinary thinking, we need to be able to hold many ways of knowing and many ways of valuing in the same spaces. Valuing and rewarding that kind of work, actually putting it into practice and moving forward with it, is really special.

HS: In your experience in working with these disparate communities, what does it look like to incorporate feedback to prioritize marginalized communities in AI development? What kind of strategies might people employ to involve communities in technological development processes?

DB: That’s a really great question. I think it depends, obviously, on the context and the project, but it’s something I’m learning a lot from community-based researchers who are experts in building and developing relationships with people who are impacted by the technologies that you’re interested in. And not only building those relationships, but seeding space and giving power, ownership, and credit to the people whose lives and experiences you want to draw on. Building in community with those for whom you’re building requires you to treat the work as a true collaboration.

HS: I’ve been seeing a lot of ML research come out in recent weeks focused on generative agents, or AI models that can take action in the digital, and then ultimately, physical world. I’m wondering what you’re thinking about the potential of a digital world that’s increasingly occupied by AI agents, and how that changes how we connect and interact with each other online and form community with others, when we don’t know whether the digital beings we’re interacting with are human or not.

DB: So, I have a younger sister who is 20 now, and I remember talking to her about TikTok. She told me she had just started a new account, and that she had to “break in the algorithm.” The ease with which she picked up the concept of “breaking in an algorithm,” and the comfort and fluidity with which she and her friends navigated this machine learning-managed space really drove home the way that people adapt to this kind of thing. We develop logics about how to move through the world and interact with technology, and those logics shift as technology shifts. One of the reasons that communicating the histories and contexts of these technologies feels so critical is that it informs the way people reason about their relationships to these technologies. These relationships arise through trial and error as people learn about new technologies, feel them out, and figure out what their digital world looks like within this new landscape. But I think we can prevent a lot of harm and help people have better relationships to technology if people have a better understanding of where these tools come from and why they exist. I’m reminded of this piece of art that I encountered called ImageNet Roulette, which Trevor Paglen and Kate Crawford put out, which was this really simple piece where people could upload their own images to a model that was trained exclusively on the person categories from this really foundational computer vision dataset called ImageNet. After you upload a picture of yourself, it might label you “orphan” or “harlot” or “underdog.” This really simple artistic intervention really revealed something important about what a model was “seeing” and then surfaced some of these labels that ultimately were removed from ImageNet. It showed something important about its history and helped people understand the kinds of issues that may have arisen, and the types of issues that still exist.

HS: ImageNet Roulette is such a great example of making aspects of technology more legible. I’m wondering, for artists who are figuring out how to engage with these generative art tools like DALL-E, what are some best practices that you use yourself for engaging with them critically?

DB: It’s a really interesting time. Digging into the history and context of these technologies is both a foundational tool for understanding their harms, and I think it’s what we need to make actually interesting art in this time and about these technologies. It’s stuff that looks at the edge cases and the boundaries of these technologies, and that reveals the human fingerprints on all of it, that really express something interesting. It’s very easy to get excited and lean into all the shiny new toys, and that’s cool. But it also plays into the hype machine that tech companies are proliferating. The most interesting work comes through digging in and finding all of the little pieces of humanity there, the pieces that show the history and where these things came from. Mimi Onuoha’s work in this space does a really interesting job exploring agency and power as it relates to data collection.

HS: I agree. The best AI art won’t come through prompt engineering, but through hacking. To close, can you share anything you’re excited about or interested in exploring, and what’s coming up next for you?

DB: Yeah, I’m really excited to share that DAIR is about to launch an essay series on possible futures of technology, where we’re trying to think more generatively about the future of tech and AI. I think there will be a lot of interesting fine-tuning work, and people playing with their own data in these models. I hope that at its best, it will be an interesting avenue for people to reflect on their own relationship to data and digital creation. I hope that this catalyzes organizing in a lot of creative professions that are currently threatened. I don’t want to end up in a boring dystopia, so I’m hopeful about this being a potential time for the mundanity of generative AI to come to the surface and have that spark a renewed appreciation for the reasons that we create things in the first place.