AI has become remarkably good at synthesizing what we already know. Whether it can identify what we don't is a much harder question—one that Charan Kukunoor spends most of his time trying to answer.
Charan’s work on making cities more intelligent, focuses on three problems that look unrelated but aren't. City planners cannot see what lies beneath their own streets. City residents cannot see how City Hall decides to budget tax money. And language models can not see the edges of their own knowledge. Each, in its own way, has an opacity problem. Each becomes tractable the moment you can measure what you didn't know before.
Charan came to AI research the long way round: urban planning fieldwork on Amaravati, Andhra Pradesh's greenfield capital city, and then with UN-Habitat, before earning a PhD at NYU's Center for Urban Science and Progress, where he studied 3D Bayesian deep learning for underground space. He is now a postdoctoral fellow at the BCG X AI Science Institute.
The following conversation explores how AI can move from a retrieval tool to a genuine engine of innovation and civic accountability. The implications are broad, from how cities manage underground infrastructure and how self-driving vehicles integrate into mixed urban traffic to how AI might eventually make the machinery of government visible and trustworthy. The conversation has been edited for length and clarity.
Your research at NYU focused on something most people don't think about: underground urban infrastructure. What problem were you trying to solve, and how did AI change how you might have otherwise tried to solve it?
Most cities, including New York, don't have comprehensive maps of what's beneath their streets. Water mains, fiber cables, gas lines, subway tunnels... they've been built up over decades by different agencies using different systems, and no one has a single coherent picture of where they all are and how they might intersect each other. The consequence is constant disruption: a crew digs for one utility and damages another, a city tries to plan a new transit corridor without knowing what lies below the surface.
I tried to tackle this by developing a 3D Bayesian convolutional neural network, a type of deep learning model to map and plan underground spaces more holistically. The Bayesian element is important because it doesn't just produce an answer; it produces an answer with a confidence interval, which matters significantly when you are making infrastructure decisions under uncertainty. The goal was to move from fragmented, two-dimensional records toward something a city planner could actually use to understand the full three-dimensional picture of what lies below and where future investment could go without creating new conflicts. Underground space is a finite resource. Cities that don't plan it systematically are going to face increasingly costly surprises.
At the BCG X AI Science Institute, you're working on what you describe as a value function for AI, a way for LLMs to evaluate their own outputs and identify genuine knowledge gaps. Can you unpack what that means and why it matters for innovation?
The limitation of large language models today is that they are excellent at synthesizing what exists, but much weaker at identifying what doesn't. Ask one to generate research questions, and you tend to get a recombination of known concepts that are plausible but superficial, are sometimes impractical, and rarely point at a real frontier. That isn't unique to AI. It's structurally similar to how early-career researchers sometimes mistake familiarity for insight.
A value function is a mechanism that lets a model score its own output against a structured picture of what is established, what is contested, and what is genuinely open. The design draws on how innovation works as a cognitive process by drawing analogies across domains, noticing an absence rather than a presence, or asking what a field of study or nature itself has already solved but no one has applied it to the subject at hand. Encode some of that structure, and you give the system a way to tell the difference between recycling old ideas and developing new ones.
The aim is to create systems that can formulate hitherto unimagined questions, not just answer the ones we already know to ask.
You've spoken about using AI to turn governance from a "black box" into a "glass box." What would that actually look like for a city resident or a civic leader, and how close is it to being real?
Technically, public data such as permits, contracts, budget lines, and infrastructure spending exist. Practically, it's out of reach. If you want to know why a road project has stalled or whether a permit is moving at a normal pace, you generally need either to know how to properly file a records request or the technical skill to interpret the record you receive. Sometimes both. As such, attempts at civic oversight often end up gatekept by complexity rather than by intransigence.
What AI agents make possible is continuous analysis of public data. Active well beyond the human work-day, these agents can surface permit delays, unusual spending patterns and make correlations across contracts that no individual analyst has the hours to flag. Then they can explain what they found in plain language to anyone who asks. No GIS training, no records request, no waiting. That's the glass box: creating transparency not by publishing more data, but by making the machinery of government more legible.
With one caveat: these agentic systems produce hypotheses, not findings. A flagged anomaly is a reason to look, not a conclusion about willful conduct. Any system deployed in this space needs to be built and communicated with this distinction as a fundamental tenet.
India's Right to Information Act is the useful example of this kind of civic transparency. The act, enacted in 2005, created both a citizen’s right to access government data and an obligation on the part of government agencies to publish the data proactively. What it couldn't create was usable, structured, current, machine-readable, and fully analyzed data. But the gap between a right and the ability to exercise that right in the real world is a tooling problem, and it is the gap that AI can close.
The components needed to attain this level of transparency are closer to ready than most people assume. The binding constraint is institutional, not technical. When cities learn how to structure and expose their data in a form that makes it truly accessible, city planners can begin thinking in three dimensions before an engaged and informed citizenry.
This article was developed using AI-assisted tools.
