Dr. Timnit Gebru visited the Digital Democracies Institute and shared with us some context and direction for the newly established Distributed AI Research Institute, of which she is the founder and executive director. The name speaks to how her teams are so distributed: across Africa, Europe, Australia, North America, and despite the coordination challenges, this distribution allows for a clearer picture of how technology is harming people around the world.
After her firing from Google and based on her overall experiences from both academia and the corporate world, she wanted to create a different space to conduct AI research. She explained that the goal was to create a different incentive structure antithetical to the standard model that requires the most productivity for least pay, leading to extractive relationships. At DAIR, Timnit explained, “following this value, we would probably put in more money per resource, so we would have to accept ‘less work’ to make sure appropriate people are compensated.” For example, one project at DAIR involves coordinated social media activity done by people who are victims of online harassment. They are doing this work for free out of necessity, and DAIR will be finding ways to compensate this labour. Also, DAIR values healthy thriving researchers and promotes healthy workloads, “an alternative to the idea that great researchers should be working up to 96-hours per week. This amounts to discriminatory exploitation,” Timnit explained. In addition, DAIR is developing guidelines to encourage researchers to prioritize their health and resist putting out papers so quickly. For Timnit, these values are difficult to enact within current academic incentives that have researchers chasing deadlines, rushing to conferences, and following the H-index, with researchers preoccupied with how much we are maximizing research impact.
Timnit posed the question, how do we do research that benefits specific communities, such as in Nyeri, Kenya where Dr. Ciira wa Maina, a DAIR Advisory Committee Member, is doing data science for environmental conservation and food security. One challenge is their funding model, where they are supported by grants from a number of foundations. Timnit said that much of her time is being spent on developing a sustainable revenue model, as two-year grants lead to precarious and uncertain situations. Also, a root of the problem Timnit explained, is as journalist Karen Hao described: a big part of people who control the AI agenda and funding conclude that if large tech companies end up doing something great that’s still a good thing. The problem with this is that under this model, technologies are not redistributed equitably. “This hasn’t happened with other technologies in the past, so why would we expect this would happen in the future”, Timnit said.
To highlight some of the issues with incentives and big tech, Timnit explained the Wired Magazine story that featured Maori people who were having a local radio competition. An American company wanted to buy their recorded data from the competition and the community rejected their offer. The community published a video that outlined the reason for rejection, where they explained, “they suppressed our languages and beat it out of our grandparents. And now they want to sell our language back to us as a service.” In doing research in this area, the current incentive structure would drive the company to sell or license the data, and it would be difficult to release it in a way that doesn’t exploit people. A different incentive structure can help preserve a community’s own metrics of success.
Timnit then outlined some projects that she considers a good use of technology that would not be considered a priority under the current incentive model. Timnit’s former colleague at Google Ernest Mwebaze and his team, for example, is working with farmers to develop a cassava disease diagnosis tool. She explained that cassava is an important food security crop grown by smallholder farmers in sub-Saharan Africa, and disease is currently leading to poor yield. As the climate catastrophe intensifies, it will be even more important to increase the yield. The team worked with farmers to develop an app that classifies diseases, tells the disease severity and counts pests to better understand the problem. However, Timnit explained that “in the field of computer vision, current priorities do not push you to stuff like this. We can assert its importance, but it is difficult to publish a paper on this because publishers will say this is not novel, but just a data set paper.” She added, while conducting peer-reviewed research is very important, the priorities of the field push you away from this kind of work.
(Source: Ernest Mwebaze, 2021)
The second project is being led by DAIR research fellow Dr. Raesetje Safala, first conceived by Dr. Nyalleng Moorosi. Collaborators are working to understand the evolution of spatial apartheid in South Africa using aerial photos. This suburb and township segregation was enacted by law through the Group Areas Act, which forced non-white groups to live in townships where they allocated smaller budget than suburbs.
(Source: Sefala et al., 2021. Constructing a Visual Dataset to Study the Effects of Spatial Apartheid in South Africa)
The stark differences are very clear, despite the fact that apartheid is over legally. Timnit explained that the research team posed the question, “since this legal change, have lives changed? Have these spatial configurations changed?” They wanted to know, “can we use computer vision to do this kind of analysis?” Gathering a data set took years, as the South African government stopped annotating townships, something they must do if the government wants to understand the impact of spatial apartheid. In their paper, they compare changes over time to what amount of land has wealthy people living on it, compared to not wealthy people. They found that vacant areas were being mostly changed into wealthy neighbourhoods, and formerly non-residential areas were demolished, with some built into wealthy neighbourhoods. Timnit explained that they ran into similar publishing problems that she discussed earlier in her presentation, with conference reviewers questioning novelty and importance of this kind of work. Eventually they published it in NeurIPS, as there is a specific dataset and benchmarks track for data set gathering, ethical considerations, etc. They will be continuing their analysis of this work and look forward to seeing how it develops.
In closing, Timnit emphasized that these projects illustrate the kinds of things an independent research institute can do, and developing priorities and norms is a part of this. For example, at DAIR, it was important to them that the South Africa project was led by Raesetje who grew up in a township. “Her specific knowledge shaped the research questions, and here we have applied the principle developed by disabled activists, ‘nothing about us without us’.” This is a research principle that the researchers at DAIR can and have decided is important to them. As we at the Digital Democracies Institute continue shaping our own values and research principles, we thank Dr. Gebru for sharing the work happening at DAIR.