A recent paper from the Center for Applied Data Ethics (CADE) at the University of San Francisco urges AI practitioners to adopt terms from anthropology when reviewing the performance of large machine learning models. The research suggests using this terminology to interrogate and analyze bureaucracy, states, and power structures in order to critically assess the performance of large machine learning models with the potential to harm people.
“This paper centers power as one of the factors designers need to identify and struggle with, alongside the ongoing conversations about biases in data and code, to understand why algorithmic systems tend to become inaccurate, absurd, harmful, and oppressive. This paper frames the massive algorithmic systems that harm marginalized groups as functionally similar to massive, sprawling administrative states that James Scott describes in Seeing Like a State,” the author wrote.
The paper was authored by CADE fellow Ali Alkhatib, with guidance from director Rachel Thomas and CADE fellows Nana Young and Razvan Amironesei.
The researchers particularly look to the work of James Scott, who has examined hubris in administrative planning and sociotechnical systems. In Europe in the 1800s, for example, timber industry companies began using abridged maps and a field called “scientific forestry” to carry out monoculture planting in grids. While the practice resulted in higher initial yields in some cases, productivity dropped sharply in the second generation, underlining the validity of scientific principles favoring diversity. Like those abridged maps, Alkhatib argues, algorithms can both summarize and transform the world and are an expression of the difference between people’s lived experiences and what bureaucracies see or fail to see.
The paper, titled “To Live in Their Utopia: Why Algorithmic Systems Create Absurd Outcomes,” was recently published and accepted by the ACM Conference on Human Factors in Computing Systems (CHI), which will be held in May.
Recalling Scott’s analysis of states, Alkhatib warns against harms that can result from unhampered AI, including the administrative and computational reordering of society, a weakened civil society, and the rise of an authoritarian state. Alkhatib notes that such algorithms can misread and punish marginalized groups whose experiences do not fit within the confines of data considered to train a model.
People privileged enough to be considered the default by data scientists and who are not directly impacted by algorithmic bias and other harms may see the underrepresentation of race or gender as inconsequential. Data Feminism authors Catherine D’Ignazio and Lauren Klein describe this as “privilege hazard.” As Alkhatib put it, “other people have to recognize that race, gender, their experience of disability, or other dimensions of their lives inextricably affect how they experience the world.”
He also cautions against uncritically accepting AI’s promise of a better world.
“AIs cause so much harm because they exhort us to live in their utopia,” the paper reads. “Framing AI as creating and imposing its own utopia against which people are judged is deliberately suggestive. The intention is to square us as designers and participants in systems against the reality that the world that computer scientists have captured in data is one that surveils, scrutinizes, and excludes the very groups that it most badly misreads. It squares us against the fact that the people we subject these systems to repeatedly endure abuse, harassment, and real violence precisely because they fall outside the paradigmatic model that the state — and now the algorithm — has constructed to describe the world.”
At the same time, Alkhatib warns people not to see AI-driven power shifts as inevitable.
“We can and must more carefully reckon with the parts we play in empowering algorithmic systems to create their own models of the world, in allowing those systems to run roughshod over the people they harm, and in excluding and limiting interrogation of the systems that we participate in building.”
Potential solutions the paper offers include undermining oppressive technologies and following the guidance of Stanford AI Lab researcher Pratyusha Kalluri, who advises asking whether AI shifts power, rather than whether it meets a chosen numeric definition of fair or good. Alkhatib also stresses the importance of individual resistance and refusal to participate in unjust systems to deny them power.
Other recent solutions include a culture change in computer vision and NLP, reduction in scale, and investments to reduce dependence on large datasets that make it virtually impossible to know what data is being used to train deep learning models. Failure to do so, researchers argue, will leave a small group of elite companies to create massive AI models such as OpenAI’s GPT-3 and the trillion-parameter language model Google introduced earlier this month.
The paper’s cross-disciplinary approach is also in line with a diverse body of work AI researchers have produced within the past year. Last month, researchers released the first details of OcéanIA, which treats a scientific project for identifying phytoplankton species as a challenge for machine learning, oceanography, and science. Other researchers have advised a multidisciplinary approach to advancing the fields of deep reinforcement learning and NLP bias assessment.
We’ve also seen analysis of AI that teams sociology and critical race theory, as well as anticolonial AI, which calls for recognizing the historical context associated with colonialism in order to understand which practices to avoid when building AI systems. And VentureBeat has written extensively about the fact that AI ethics is all about power.
Last year, a cohort of well-known members of the algorithmic bias research community created an internal algorithm-auditing framework to close AI accountability gaps within organizations. That work asks organizations to draw lessons from the aerospace, finance, and medical device industries. Coauthors of the paper include Margaret Mitchell and Timnit Gebru, who used to lead the Google AI ethics team together. Since then, Google has fired Gebru and, according to a Google spokesperson, opened an investigation into Mitchell.
With control of the presidency and both houses of Congress in the U.S., Democrats could address a range of tech policy issues in the coming years, from laws regulating the use of facial recognition by businesses, governments, and law enforcement to antitrust actions to rein in Big Tech. However, a 50-50 Senate means Democrats may be forced to consider bipartisan or moderate positions in order to pass legislation.
The Biden administration emphasized support for diversity and distaste for algorithmic bias in a televised ceremony introducing the science and technology team on January 16. Vice President Kamala Harris has also spoken passionately against algorithmic bias and automated discrimination. In the first hours of his administration, President Biden signed an executive order to advance racial equality that instructs the White House Office of Science and Technology Policy (OSTP) to participate in a newly formed working group tasked with disaggregating government data. This initiative is based in part on concerns that an inability to analyze such data impedes efforts to advance equity.