Caring for Datasets: Eshta Bhardwaj on Applying Information Lenses in the Age of AI

Before a machine learning model produces an answer, makes a prediction, or classifies an image, there are datasets. And before there are datasets, there is data work: the often-invisible labour of collecting, selecting, cleaning, labelling, documenting, interpreting, reusing, and maintaining data.

For Eshta Bhardwaj, a PhD Candidate at the University of Toronto’s Faculty of Information and a member of the Just Sustainability Design Lab, that work deserves far more attention than it usually receives. Her research asks what changes when machine learning treats dataset development not as a technical precondition, but as a form of careful, situated, and accountable practice. Drawing from data curation, archival theory, library science, critical data studies, and prior machine learning research, Eshta studies how data practices can become more rigorous, reflexive, and responsible.

This question sits at the centre of INF1005 and INF1006: Caring for Datasets, a workshop Eshta is teaching at the Faculty of Information in Summer 2026. The course invites students to consider dataset development as craft work requiring reflexivity, criticality, and rigour, and explores how norms and practice-based processes from information fields can elevate responsible machine learning data work.

For students encountering machine learning through the lens of information studies, the workshop offers a different starting point. Rather than asking only how models perform, it asks how data becomes AI, who participates in that process, what assumptions are embedded along the way, and what responsibilities follow from those decisions.

Rethinking what counts as progress in machine learning

Eshta defended her thesis proposal in July 2025. Her publication-based thesis, “Adopting a Data Curation Lens for Rigorous, Reflexive, and Responsible Machine Learning Data Practices,” examines how machine learning can learn from information fields that have long dealt with questions of provenance, integrity, preservation, representation, access, future use, and accountability.

The thesis begins from a clear diagnosis: many concerns in machine learning, including bias, data power, hidden labour, weak accountability, lack of transparency, and environmental cost, can be traced back to how data becomes AI. Eshta frames data practices as what happens across the lifecycle of data: how data are collected, managed, used, interpreted, reused, deposited, curated, and more. Her central argument is that concepts and methods from data curation can improve machine learning data practices by making model development more rigorous, reflexive, and responsible. Just as importantly, she argues that a curation perspective can help redefine what “progress” in machine learning means.

Progress does not have to mean only larger models, larger datasets, and higher benchmark scores. It can also mean better-documented datasets, more accountable design decisions, greater care in reuse, and more honest attention to social and ecological consequences. This is one of the central contributions of her work: she shifts attention from model work to data work, and asks how the two should relate.

Evaluating machine learning data practices through a curation lens

One of Eshta’s major contributions appears in FAccT 2024, in a paper titled “Machine Learning Data Practices through a Data Curation Lens: An Evaluation Framework,” with Harshit Gujral, Siyi Wu, Ciara Zogheib, Professor Tegan Maharaj, and Professor Christoph Becker. The paper develops a framework for evaluating machine learning datasets using data curation concepts and principles, then applies it to 25 machine learning datasets. In doing so, it translates a central insight from information studies into a concrete research method: datasets can be assessed not only as inputs for models, but as curated objects shaped by decisions, documentation practices, institutional norms, and future reuse.

While data curation offers a promising lens for improving fairness, accountability, and transparency in machine learning, applying curation principles in ML is not straightforward. Machine learning researchers often work in a model-focused environment, while data curation brings a lifecycle view that asks how data are maintained, contextualized, documented, preserved, and made usable over time.

These findings show why the work matters. If ML researchers struggle to apply standard data curation principles, then responsible AI requires more than adding a checklist at the end of a project. It requires training, collaboration, shared vocabulary, and new expectations for what rigorous data work looks like.

Evaluating structured documentation as a tool for reflexivity in dataset development

In her most recent publication, co-authored with Ciara Zogheib and Professor Christoph Becker, Eshta examines documentation frameworks such as datasheets, data statements, and dataset nutrition labels. These frameworks are designed to make dataset development more transparent by asking creators to describe where data came from, how it was collected, what it contains, how it should be used, and what risks it might pose.

But Eshta and her co-authors ask a more difficult question: does documentation actually prompt reflexivity? Their answer is nuanced. Documentation can make dataset development more visible. It can help researchers explain what they did. But transparency is not the same thing as reflexivity. Reflexivity requires deeper engagement with the assumptions, values, power relations, histories, institutional conditions, and social positions that shape research decisions. Put simply: transparency asks what happened. Reflexivity asks why it happened that way, who shaped the process, what assumptions guided it, and what consequences might follow.

AI scaling, sustainability, and the limits of growth

Eshta’s research also moves beyond individual datasets to examine the broader systems that shape AI development. In “Limits to AI Growth: The Ecological and Social Consequences of Scaling,” with Professor Christoph Becker, Eshta studies the accelerating development and deployment of AI through the lens of scaling. The paper reviews AI growth across four interrelated dimensions: technical, economic, environmental, and social. AI depends on increasing levels of capital investment, infrastructure, data, and natural resources, producing financial, environmental, and social costs. It asks a pointed question: How can we characterize the dynamics of AI growth to identify and analyze its limits? To answer it, Eshta and Becker use system dynamics concepts, including the “limits to growth” archetype, to understand how technical scaling, investment cycles, competitive pressure, AI hype, infrastructure development, and ecological consequences interact.

Bringing research into the classroom

In Caring for Datasets, Eshta Bhardwaj brings these research questions into the classroom. The workshop frames care for datasets not as a softer alternative to technical rigour, but as a way of practicing it. It invites students, researchers, and practitioners to slow down and ask better questions about provenance, context, documentation, accountability, power, reuse, sustainability, and responsibility. In doing so, it opens a space to think about responsible machine learning long before a model is trained.