When Machine Learning Meets Policy: Why AI Strategy Needs Interdisciplinary Thinking
Artificial intelligence is often discussed in extremes: either as a purely technical challenge to be solved by better models, or as a societal risk to be managed through regulation and ethics frameworks. In practice, however, AI lives in the space between these two worlds. Over the past several years, my work has focused on exactly that intersection, where machine learning meets policy, institutions, and governance, and on what we gain when we stop treating them as separate domains.
My research began with a simple but underexplored question: how do governments and institutions actually articulate their priorities, values, and assumptions about AI? National and sub-national AI strategies are not just technical roadmaps. They are narratives. They signal what a country believes AI is for, who it should benefit, what risks matter, and which trade-offs are acceptable. Yet much of the analysis of these strategies has relied on selective reading, predefined categories, or narrow ethical checklists.
To address this gap, I turned to machine learning, not as an end in itself, but as a tool for systematic interpretation. Using unsupervised learning methods such as topic modeling, I treated AI strategy documents as large-scale textual corpora and asked what themes emerge without being imposed in advance. This approach allows us to surface priorities that may not be explicitly labeled as ethics, equity, or innovation, but nonetheless shape how AI is governed in practice.
However, computational methods alone are not enough. Machine learning models can identify patterns, but they do not explain their significance. That is where interdisciplinary collaboration becomes essential. By working closely with policy scholars, social scientists, and domain experts, I focused on making these computational findings interpretable, stable, and meaningful in real governance contexts. In practice, this meant repeatedly reviewing model outputs with policy scholars, who flagged when themes shifted across runs in ways that undermined interpretability, prompting methodological refinements that prioritized stability and reliability. This process led to methodological innovations that emphasize reproducibility, robustness, and human-in-the-loop validation, which are critical requirements when AI is used to study high-stakes policy decisions.
More recently, my work has expanded to examine how AI strategies differ across national, sub-national, and international levels, and how values such as fairness, accountability, and health equity are unevenly emphasized across contexts. These analyses highlight an important lesson. AI governance is not monolithic. It is shaped by institutional capacity, political economy, and local priorities. Understanding this variation requires both technical rigor and contextual insight.
Why does this matter for AI strategy today? Because many of the challenges we face, including trustworthy AI, responsible deployment, and alignment with societal goals, cannot be solved by better models alone, nor by policy in isolation. They require tools that can translate between technical systems and institutional realities.
Interdisciplinary work is not always easy. It requires learning new languages, questioning assumptions, and resisting the temptation to oversimplify. But it is precisely this synthesis that allows us to move from abstract principles to evidence-based, actionable AI strategies. As AI continues to shape economies and governance worldwide, bridging machine learning and policy will be not just useful, but necessary.
Your approach of using topic modeling to uncover implicit values in AI policy documents is fascinating—so much insight lies in what’s not explicitly stated. It reminds me of how technical and practical knowledge often live in silos; for instance, understanding real-world measurements like actual lumber sizes or pipe dimensions requires both data and context. Resources like ChartsPrintables bridge that gap with clear, usable references—much like interdisciplinary research bridges AI’s technical and societal dimensions.
Quarterly data-use agreements and annual compute credits mean the county pilot can’t rerun topic models until themes settle; we freeze the corpus first, then log every stopword and topic-count choice like a custom name tracing worksheet generator locks letterforms. Policy scholars still flag drift, but the only fix is documenting cuts. We lean on a few best AI tools for policy work to version pipelines, yet human-in-the-loop validation still isn’t a budget line. That’s the gap.
In a county-level pilot, the constraint isn’t theory but procurement: compute credits reset annually and data-use agreements lapse every quarter, so rerunning topic models until themes stabilize is often impossible. The policy scholars flagging theme drift are right, but for us that means freezing the corpus and documenting preprocessing choices like the custom name tracing worksheet generator locks letterforms. We keep our own grow-a-garden value list at our own grow-a-garden value list to track which values survive those cuts. Does human-in-the-loop validation become a budget line?
Calling topic modeling a way to surface themes “without being imposed in advance” is too clean. Preprocessing, stopwords, and the number of topics are impositions, and the fact that policy scholars flagged themes shifting across runs shows the model isn’t neutral. Even a custom name tracing worksheet generator locks in letterforms and spacing; so does a corpus pipeline. Human-in-the-loop validation should be treated as part of the method, not a final check.
The point that AI strategy fails when it is framed as either a modelling problem or a governance problem is well made, and the space-between framing is more useful than the usual call to bridge the gap. The interesting part is where the two disciplines actually change each other's methods rather than just sitting in one room, and a concrete example of a policy question that changed a modelling choice would make it land. For turning an explainer like this into short video, AI Fruit can generate talking-object and meme-style clips from a prompt.