Abstract
Artificial intelligence (AI) and machine learning (ML) systems are often presented as a more objective and rational form of decision-making than humans.
Because AI/ML techniques are seen as applying fewer assumptions than people, they are a kind of tabula rasa reflecting just the data.
This perception is flawed; a variety of thinkers have pointed out the ways in which AI/ML systems necessarily reflect subjective, non-falsifiable choices and assumptions by the people designing them.
These choices, which this dissertation refers to as methodological choices, are frequently implicated in problems around AI/ML fairness and accountability.
Recognizing that methodological choices are a fundamental aspect of AI/ML systems means that the way in which models are used and built must change.
The first contribution this dissertation makes towards this goal is through a case study in the use of synthetic data.
Synthetic data, realistic approximations of real data, can be used to analyze settings for which we lack a real world ground truth.
In our case study, we use synthetic data to analyze the possible consequences of different infectious disease surveillance approaches.
We use simulations of disease spread to show that aggregate structural statistics of infection cascades can robustly distinguish between different disease scenarios, even when only observing a portion of the cascade.
These results are not, and could not, be verified using actual ground truth data, but indicate that policymakers should consider building the capacity to privately aggregate disease transmission statistics.
While using a synthetic ground truth may seem to weaken these results, under the framework of methodological choices, these results have equivalent epistemic status to AI/ML problems with a "real" ground truth.
Furthermore, using synthetic ground truth - nearly entirely a product of methodological choices and assumptions - enabled us to investigate questions that would have been impossible to answer with real data.
A related question this dissertation takes on is under what circumstances methodological choices related to data creation have substantive effects, and when are they irrelevant?
As a first step towards this goal, this dissertation studies different methods for measuring data distance through a methodology for generating datasets with bounded statistical differences.
Through empirical benchmarking experiments, we find that most commonly used methods for measuring data difference are not sensitive to substantive differences between datasets.
However, for some tasks, these differences may be irrelevant.
These results suggest that trying to filter methodological choices based on their impact on the data may be difficult to do without additional contextual knowledge.
Lastly, this dissertation looks at, in practice, whether prompting model designers to more carefully consider their methodological choices leads to more fair models and practices.
We describe a computational notebook interface and associated user study meant to nudge model designers to more carefully consider methodological choices, in particular the impact some methodological choices may have on fairness.
In a user study, we show that nudging throughout the process leads to more thoughtful engagement with the fairness and bias of their methodological choices.
