Human-Computer Interaction Ph.D. Thesis Defense - Alicia DeVrio

AI systems often behave in ways that harm people, such as by perpetuating racial stereotypes, violating privacy, damaging health, and causing economic losses. There have been many efforts to address AI harms that are led top-down by researchers, industry practitioners, government officials, or other technologists. These top-down, more centrally-led approaches to AI harm remediation tend to view AI systems as inevitable and thus conceive of reformist solutions to harm.

At the same time, people affected by AI harm regularly take both small and large actions, from individual workarounds on algorithmic platforms to collectively organized protest and sabotage of algorithmic systems, to resist AI harm in the ways they can. These bottom-up forms of resistance to AI harm have broader imaginaries than top-down methods of addressing AI harm, envisioning beyond the bounds of AI system logics to see technology as part of the picture but not necessarily so.

Examining bottom-up resistance to AI harms is especially important because existing structures of power around AI systems often fail to empower those who directly interact with and are most affected by AI harm. But, there has been less work focusing on how to directly support those negatively impacted by AI systems in taking action in bottom-up ways. I argue that to make meaningful progress toward addressing the harmful impacts of AI systems, bottom-up resistance must be better understood, valued, engaged with, and broadly supported.

In the first part of this dissertation, I examine one form of bottom-up resistance that I call everyday algorithm auditing, a process by which people work from the bottom up to surface, interrogate, and work toward remediation of AI harm that may elude detection via more centrally organized forms of auditing. First, I conceptualize this process to understand its dynamics and characteristics. Then, I investigate how people go about searching for and making sense of potentially harmful AI behavior from the bottom up, highlighting ways they might be further supported in their auditing activities. Finally, I examine how industry practitioners engage with bottom-up auditing, drawing out a complex relationship with fraught power dynamics between practitioners and everyday auditors.

To that end, in the bridge section of this dissertation, I then step back to examine the function and impact of algorithm audits—including bottom-up audits—in larger structures of power. This highlights a need to attend to the relationship between acts of bottom-up resistance and larger power dynamics.

Thus, in the second part of this dissertation, I take a more expansive look at how bottom-up resistance interacts with power. First, I develop a taxonomy of bottom-up resistance to AI harm that connects different acts of resistance, including everyday algorithm auditing, to existing theorizations of power. Next, I study how people understand their own power in algorithmic systems, highlighting ways to support them in more fully realizing and taking action using their power. Finally, I examine disconnects and alignments in perceptions of bottom-up resistance held by AI deployers and AI resistors, using speculative cases of algorithmic system sabotage as probes, and explore challenges and opportunities for the powerful to more productively engage with, value, and support bottom-up resistance to AI harm.