Grounding AI: Exploring the Role of Algorithms in Science

Unfolding AI, or Learning to Read the Sky

There is a version of AI that lives in headlines, in social media posts, in the quiet shared glance when someone mentions it at dinner. It feels enormous and yet hard to grasp, always there, almost inevitable by now. It presents itself as both promise and threat. You probably know this version, as most of us do.

But if you stop to think about it, what are we actually talking about when we talk about Artificial Intelligence? What do we have in mind, exactly? Is it a song our music platform recommended to us? The chat we ask for help when replying to emails? A hospital’s triage system? An autonomous flying drone?

When we talk about AI this way, we are building a figure — singular, capitalized, implicit — that doesn’t need to explain where it is or what it’s doing, because it’s everywhere and doing everything, or about to be. Even critique tends to accept this framing. Whether you think AI will save us or end us, whether you’re excited or exhausted by the conversation entirely, the grammar stays the same: “AI will, AI can, AI threatens.” A single actor with a single story.

This construction, this representation of AI, is real. We are all living through it, whether we work with algorithms daily or only encounter them as invisible infrastructures operating around us. The unease is real. The excitement is real. The feeling that something large is shifting, or has already shifted, is real. But all of this is a specific view on the matter. A view from somewhere. Maybe a view from nowhere. Definitely a view from a distance. And when we look at things from a distance, we tend to reduce, simplify. Our minds turn details into outlines, making the multiple singular. We see a shape, but only once we stop seeing everything that shape is made of.

You are holding something that was made in response to that distance. Not to dismiss what you can see from there, but to ask: what happens when you step closer? What is actually there?

Look at the map you are holding right now. A folded square, about the size of a book. Unfold it all the way. What you’re looking at, once it’s open, is what we called the Grounding AI Map: a visualization of around two million scientific papers about artificial intelligence, algorithms, and machine learning, retrieved from the Scopus database and published between 1985 and 2024.

In the map, each document becomes a dot, and these are arranged by semantic similarity: papers that share vocabulary, problems, and technical approaches sit close to one another, and the dense regions can be seen as clusters of research that grew up around common questions. The colors trace time, from red for older work to blue for newer. The labels you’ll find scattered across the map were generated through an AI-enabled pipeline, where we trained a large language model to work with us, reading each cluster from the inside and creating a title and a short summary for each of the 4062 clusters that we found. Just remember that ours is an invitation to explore, not an attempt to give answers.

From up close, you can see what scientists actually work on. There is research on smart wearables, on the chemical composition of soil, on how to identify trees from drones, on detecting fraud in financial transactions, on optimizing energy grids. Each cluster is a community of researchers working on something specific, with their own methods, their own debates, their own slow accumulation of findings.

People often say it looks like a galaxy — densities, gradients, something that resembles deep space, or weather systems seen from very far above. The shapes suggest nebulae, or the terrain of some other planet, something your eye keeps trying to resolve into a recognizable figure.

“Why is it called Grounding AI?” you might ask. This version of the map, the one you can hold, fold, and carry with you, is a pocket-size version of the same visualization that we printed as a floor mat large enough to walk on. We literally place AI research on the ground. The map becomes a stage where you move through research areas, questions, and ideas. Stand in the middle of a cluster about medical diagnostics, or flood prediction, or materials science, and you’ll be able to read only what’s immediately around you: the labels at your feet, the neighboring regions within a few steps. The rest of the map becomes blurry. You can see where it ends, but you can’t read all the labels at the same time. It is not a view from nowhere. It forces you to place yourself somewhere specific, reading from a position you had to choose — even if you chose it by accident. Your body becomes part of the experience, of the reading.

This foldable map does something different. Think of it as a wayfinder, an astronomer’s chart: you can see the whole sky at once, but you’re seeing it reduced, compressed, without the granular detail of the full installation. Of the 4062 clusters, only 72 are labelled here, not because they are more important than the others, they were just selected as a sample to be distributed evenly across the map. The trade-off is legibility. This version of the map helps you orient yourself, find the region you want to explore before you step onto it, or remember where you were standing after you’ve left. Folded, it contains everything the unfolded version does — the same two million papers, the same clusters, the same semantic distances — but in a form you can carry. The act of unfolding becomes more than a practical gesture. Unfolding is the gesture the whole project is asking you to make.

A galaxy has a name. The Milky Way. Andromeda. Names that make the unknown vastness just a bit more familiar, so that we can point up and say “that one”. We give names to things to make sense of them, to categorize the world. But once we give a name to something so immense like a galaxy, we turn its multiplicity into singularity. We make a figure out of something that was never trying to be one.

Constellations work the same way. We looked up at the sky and found hunters, bears, wagons — shapes that felt inevitable once someone pointed them out. But they’re not inevitable. They are stories we told from exactly this distance, because we needed the sky to make sense. The Greeks saw a hunter where the Japanese saw a drum. The stars were the same. They just got caught in different stories, held inside different sets of human needs and hopes and fears.

The stars themselves are real. The constellation doesn’t lie about them — it simplifies. It collapses something intricate and particular into a silhouette you can remember, a shape you can find again on a different night. We made up constellations to talk about the stars, and then, slowly, we started talking about the constellations instead. The hunter acquired a mythology. The stars disappeared behind it.

AI, as one thing, works like this.

A name given from a distance, one that turns something multiple, specific, and deeply interconnected into a single figure with a single story. Sometimes the figure is a threat, sometimes a promise, sometimes both at once, depending on who’s speaking and what they need it to mean. But it remains one thing, with one trajectory, moving through the world as if it had intentions and moods and a coherent will.

When you only have the constellation, you lose the specific distances between the stars, the difference in their temperatures, the fact that some of them burned out centuries ago and you’re seeing old light. You lose the granularity. You’re left with a story about them — a story that can be useful in its way, navigable, orienting, but that cannot substitute for seeing what’s actually there: the specific, burning, aging, dying particularity of each one.

What the Grounding AI Map does is let you step close enough that the figure starts to dissolve. Not into nothing — into its actual parts. Into the two million conversations researchers have been having about what algorithms can do, where they fail, what problems they were built to solve, and what problems they created instead. The stars don’t disappear. They just stop telling you a story about a hunter.

The view from far away, the one that lets you see AI as one. One solution, one controversy, one story, remains real and sometimes necessary. But it remains a specific view. And there are other views to look from. This map is one of them.

Should we call it deflation? Not because AI gets smaller — it doesn’t — but because the inflated story finally meets the ground.

Consider the constellation as the inflated version. When we talk about Orion, we’re talking about a hunter with a belt and a bow, a figure with mythological weight and narrative built all around him. We forget that there’s no hunter up there. There are stars doing what stars do: burning, dying, drifting slowly through space on their own trajectories, indifferent to the stories we’ve organized around them. The hunter exists because we looked up from exactly here and told a story that made sense from this angle, at this distance, in this moment of human history — to make the sky more readable.

Inflation serves a purpose. It’s a way of organizing complexity so we can talk about it, navigate by it, form an opinion about it. But the story develops its own gravity, its own pull. Soon we’re asking questions about the hunter — what is he doing, what does he want, should we be afraid of him — and we’ve stopped asking about the stars.

Deflation is what happens when you step close enough that the story can’t hold. The hunter dissolves — not into nothing, but into the actual, interconnected parts that were always there. This map deflates AI: not by claiming it isn’t real or doesn’t matter, but by refusing to let it stay singular. Every cluster you see — Smart Wearables, Underwater Sensor Networks, Student Performance Prediction, Secure e-voting — contains papers that mention algorithms, describe specific techniques, and belong, fully, to what we mean when we say AI. Deflation recontextualizes the matter rather than removing it. AI doesn’t vanish. It lands somewhere specific: inside a problem, a field, a community of researchers who were mostly trying to solve something that had nothing to do with AI in the first place, and who found that certain algorithmic approaches turned out to be useful.

The political and social weight doesn’t disappear either. Some of these clusters carry real stakes, real consequences for real people: who gets diagnosed, whose neighborhood floods, what counts as quality, who gets to decide. Deflation doesn’t dismiss urgency or critique. It doesn’t decrease controversiality. If anything, it refines it. It makes critique possible in a different register, because now we can ask: which AI, where, doing what, for whom, solving what problem, and at whose expense?

When AI is no longer one thing with one story, the questions get better. Not what will AI do to us, as if we’re standing still and it’s approaching from outside, but: what is being done here, in this context, through these decisions, with which consequences, and for whom? Paradoxically, deflation makes the thing bigger. The singular story compresses everything into one trajectory. The deflated view opens into multitudes. Two million papers. Hundreds of fields. Thousands of specific, local, deeply situated problems where algorithms turned out to matter.

Walk onto the floor map, and you might find yourself standing on Bean Authentication. There is something funny in realizing that the vast intelligence you’ve been reading about is also, for some reason, concerned with coffee beans. Sit with that reaction for a moment. What is it, exactly? Perhaps a kind of relief. Or recognition. The algorithms are still there — the papers in that cluster describe machine learning techniques, optimization methods, classification models — but they describe them in relation to something concrete, something grown and harvested and tasted: authenticating bean quality, predicting yields, modeling roasting temperatures. You suddenly face the fact that AI, this almost magical thing, is also grounded in something as simple as a cup of coffee.

Seen from inside the clusters, AI means something different than it does from the satellite view. More specific. More entangled with things that smell and taste and grow and flood and break and need to be authenticated or predicted or optimized or understood. More ours, in some sense, because it turns out to live inside the same material world we do.

You’ll fold this back up soon. Crease it along the lines it already knows, tuck it into a bag or a pocket. A condensed galaxy to carry around for a while.

This map won’t answer the big questions people keep asking at dinner tables and on social media, the ones that make AI feel enormous and inevitable. Perhaps that’s not what it’s for. What it offers instead is a different kind of question — the kind that can only be asked from somewhere specific. Not what AI will do to us, but what is happening here, in this cluster, in this field, with these people, working on this problem?

The next time someone says AI is going to change everything, you might remember what it felt like to look at it from close up. To stand on a cluster about flood prediction or bean authentication and see the situated, entangled, deeply material ways that algorithms are already changing things — some of them worth celebrating, some of them worth resisting, most of them worth understanding better than we currently do. The map doesn’t tell you which is which. It helps you find a way to look, and reminds you that the one story dissolves into many once you step close enough.

What becomes possible from here is yours to find.

Matilde Ficozzi, April 2026