The myopia of model centrism
AI
models seek to intervene in increasingly higher stakes domains, such as
cancer detection and microloan allocation. What is the view of the
world that guides AI development in high risk areas, and how does this
view regard the complexity of the real world? In this talk, I will
present results from my multi-year inquiry into how fundamentals of AI
systems---data, expertise, and fairness---are viewed in AI development. I
pay particular attention to developer practices in AI systems intended
for low-resource communities, especially in the Global South, where
people are enrolled as labourers or untapped DAUs. Despite the
inordinate role played by these fundamentals on model outcomes, data
work is under-valued; domain experts are reduced to data-entry
operators; and fairness and accountability assumptions do not scale past
the West. Instead, model development is glamourised, and model
performance is viewed as the indicator of success. The overt emphasis on
models, at the cost of ignoring these fundamentals, leads to brittle
and reductive interventions that ultimately displace functional and
complex real-world systems in low-resource contexts. I put forth
practical implications for AI research and practice to shift away from
model centrism to enabling human ecosystems; in effect, building safer
and more robust systems for all.
Bio:
Nithya Sambasivan
is a Research Scientist at PAIR, Google Research and leads the
human-computer interaction (HCI) group at the India lab. Her current
research focuses on designing responsible AI systems by focusing on the
humans of the AI/ML pipeline, specifically in the non-West. Her research
is seminal to Google's products and strategy for emerging markets,
while also winning numerous best paper awards and nominations at
top-tier computing conferences. Nithya has a PhD. in Information and
Computer Sciences from UC Irvine.