How do moderators of online communities see their roles? As a firefighter, as police, as a gardener, as a mediator? This paper explores the metaphors that volunteer moderators use to make sense of their roles.

Data work is foundational to any AI application, but often undervalued. In practice, this results in “data cascades,” which occur when data issues compound into downstream technical debt. This work studies how data cascades happen, what they cause, and how to avoid them.

How do the designs of mental health helplines influence people’s experiences with them? This paper weaves together the lived experience of 18 helpline stakeholders to understand barriers along the pathway to care.

I’ve decided to move on from both Nielsen and data science and start a new role in engineering. This is a short post with some reflection.

BERT is one of many large language models taking NLP by storm. But little is known about how or why it works, leading to papers studying it specifically: “BERTology.” This survey paper synthesizes current research on how BERT works and what remains unknown.

What does it mean for a human to trust an AI system? What are the requirements for it to be present, and how do we know if it is? This paper, which just appeared at FAccT, explores these questions.

What does collective action look like online? The authors design and study Dynamo, a platform for organizing Amazon Mechanical Turk workers into action, and discuss the unique affordances and challenges of the web.

Datasets, which are a central part of all machine learning, are often alarmingly undocumented. How can we do this better, prioritizing transparency and accountability? The authors propose datasheets that describe the characteristics, collection practices, and recommended uses of datasets.