There’s a famous story, definitely anecdotal, in which the philosopher and mathematician Bertrand Russell gives a public lecture on astronomy. He explains how the Earth orbits the Sun, and the Sun orbits the center of the galaxy. After the talk, an elderly woman approaches him and politely informs him that he has it all wrong. The world, she says, is a flat plate resting on the back of a giant turtle.
“And what does the turtle stand on?” Russell asks.
“Another turtle.”
“And that turtle?”
She smiles. “It’s turtles all the way down.”
Stephen Hawking tells a version of this in *A Brief History of Time*. William James told a version in 1882, but with rocks. Like most good stories, the attribution doesn’t quite hold up. Nobody can produce the original.
The turtle anecdote is a joke about infinite regress, an explanation that never finds bedrock. Philosophers have been wrestling with it for two thousand years. How do you justify a belief? With another belief. What does that belief rest on? Another belief. The pattern shows up in epistemology, in causation, in metaphysics, certain branches of match and computer scients…almost anywhere we ask the deepest version of “why.”
And there’s some rich irony that you can’t quite get to the bottom of the origins of infinite regress.
We don’t usually run into infinite regress in daily life, because we hit something concrete pretty quickly. The reason you believe the milk is in the fridge is that you put it there an hour ago and you can go grab it. The reason you believe the bridge will hold your car is that you’ve driven over it for ten years… and so have your neighbors. Beliefs ground out in experience, in physical evidence and in things that aren’t more beliefs.
A large language model doesn’t have that same grounding.
Inside an LLM, every word is defined by its statistical relationship to other words. “Hunger” lives near “stomach,” “food,” “craving,” “empty.” Those neighborhoods exist because those words appear near each other in billions of documents, web pages and comments. But what does “stomach” mean? More neighborhoods. “Food”? More neighborhoods. Pull on any thread and you never arrive at anything that isn’t more language.
It’s words all the way down.
The Dictionary Loop
You can verify this with a paperback dictionary in a few minutes (or a suitable digital version for those who recycled all your paper dictionaries).
Pick any word. Look it up. Take a key word from the definition and look that one up. Repeat.
Start with “love.” The first definition will probably lean on “affection.” Affection leans on “fondness.” Fondness leans on “tender.” Tender leans on “loving.” Loving becomes “Feeling or showing LOVE or great care.” Six steps in, we’re back where we started. We traveled in a tight loop without ever leaving the book.
This isn’t sloppy lexicography. It’s mathematically and structurally inevitable. A dictionary contains a finite set of words, and every word is defined using other words from the same set. The whole thing is a closed graph. You can make the graph bigger by adding more words. You can make it richer by adding cross-references. You cannot escape it. The system is closed by design.
A human child does not learn “hunger” from a dictionary. They learn it from the ache in their belly, from their own body’s need for energy, ( and eventually from green chili bacon cheeseburgers!) The word ‘hunger’ gets pinned to a bodily experience driven by a need for energy. That pin is the elusive turtle at the bottom, the thing that stops the regress. LLMs have no pins. They have an extraordinarily detailed map of how words relate to other words (or shapes relate to shapes etc. etc.), and no experience to anchor any of it.
If you’ve ever read a passage of AI-generated prose that was technically fluent and somehow felt hollow, you were sensing the absence of pins. The map is dense, but the territory is empty.
The Chinese Room
In 1980, the philosopher John Searle published a thought experiment called The Chinese Room. It’s become more relevant now as our AIs get better.
Imagine a person sitting in a sealed room. They speak no Chinese. They have a thick rulebook, written in English, that tells them what to do when various sequences of Chinese characters come in through a slot in the wall. The rules are mechanical. “If you see this sequence, write that sequence and push it back through the slot.” Nothing in the rulebook explains what any of the characters mean.
To a Chinese speaker standing outside the room, the responses look like fluent conversation. The room appears to understand Chinese. Inside, the person is just shuffling symbols according to instructions. They have no idea whether they’re discussing Tang dynasty poetry or what to order for dinner. There is no comprehension. There is only manipulation.
Searle’s point: syntax is not semantics. You can manipulate symbols correctly without understanding what they mean.
LLMs are the Chinese Room scaled to a planet. The rulebook is now trillions of parameters trained on most of the readable internet. The output is astonishing. The fluency would have been unthinkable a decade ago. And yet, when you trace what’s actually happening inside, it’s still symbol manipulation according to rules. The room got much bigger.
This is the part where AI enthusiasts often push back. Maybe the system as a whole understands, the argument goes, even if no individual component does. Maybe understanding is just what sufficiently complex symbol manipulation looks like from the inside. Searle anticipated this. His answer was that no amount of symbol shuffling, no matter how complex, ever crosses the line into meaning. Scale the Chinese Room up to fill a continent. You will not produce a Chinese speaker. You will produce a very large lookup table.
The Grounding Problem
A decade after Searle, the cognitive scientist Stevan Harnad gave the underlying issue its formal name: the Symbol Grounding Problem. His question was simple and uncomfortable. How do the symbols inside a cognitive system get connected to the things they refer to?
Harnad’s answer, broadly: the connection has to come from outside the symbol system. Through senses. Through bodies. Through interaction with a physical world. A robot with cameras and grippers can learn that “cheeseburger” means the thing it sees right there, and “heavy” means the thing it struggles to lift. The symbols touch the ground because the robot has a ground to stand on. And ultimately the words relate to some form of energy.
So...
To be clear, this isn’t an argument against AI. As stated in An Imperfect Circle, generative AI is a once-in-a-generation force multiplier, especially for small organizations that can’t afford to hire away their problems.
AI systems are providing meaningful value – operating fusion reactor magnets, predicting protein structures, controlling self-driving cars, helping draft contracts and grant proposals. Those systems work because they’re grounded. The AI is operating in a closed loop with sensors, physics, and consequences. It can be wrong, and being wrong has costs, and the system gets corrected. When a fusion reactor magnet drifts out of spec, the plasma destabilizes and the physics tells you so immediately. By contrast, when AI ‘generates’ a city council vote, nothing tells you. That gap is the whole problem in one sentence.
The claim in this article is narrower, and more important. Language-only AI has a ceiling, and the ceiling is the grounding problem. No matter how much text you feed in, no matter how many parameters you stack, the model is still operating inside the dictionary. Better training data doesn’t fix this. Bigger context windows don’t fix it. You can’t escape a closed system by making the system larger.
The near term fix is the inside-out approach we sketched in “An Imperfect Circle”. AI that operates tools grounded in math, in physics, in real data about real things… running tools and generating commands rather than generating content. The SVG circle works because the math grounds out in geometry. A transcription works because the audio grounds out in actual sound waves. A CRM query works because the records ground out in actual customers who actually exist. Every inside-out use of AI breaks the regress by connecting language to something that isn’t language.
The LLM by itself is the woman with the turtle argument. It has an answer for every question, and every answer is more turtles. The inside-out approach is the moment you walk outside and step on the actual ground.
Local News - Supporting all the other turtles
This brings us back to Local News.
Local journalism is grounded by definition. The reporter was at the council meeting. The photographer was at the scene. The weather data came from real sensors at a real location. The high school football score was witnessed and recorded by someone in the bleachers. The new business that opened on Main Street has an address you can drive to.
That’s what makes local news the anti-turtle. It doesn’t rest on another abstraction. It rests on someone who was there.
In an age of AI-generated content that is words all the way down, verified local reporting is the ground floor. Journalists do something language models can’t. They show up. They observe. They get the name right because they asked. They check the address because they drove past it. They quote the source because they were in the room when the source spoke. And most importantly, the fact is locally verifiable.
Set aside any romance about the dignity of the profession for a minute. The point is structural. Information has to come from somewhere. The reason you can trust the council vote count in your local paper is the same reason you can trust the SVG circle. There’s an unbroken chain back to something real. The reporter saw the vote. The vote was a physical event in a physical room. The number is grounded.
The reason you can’t trust an AI summary of “what’s happening in city government this week” is the same reason you can’t trust the Franken-circle. There’s no unbroken chain. The summary was assembled from fragments of stories about other councils, in other towns, governing other communities. It is shaped like a story about your city. It is a story about a statistical average of cities, draped over your city like a sheet over furniture you can’t quite see.
Local information and local signal works the same way. Every reliable claim, traced far enough, lands on something physical that someone (probably someone you *know*) actually encountered.
A society that loses the habit of pinning its words to its world doesn’t notice the loss right away. The conversation continues. The fluency improves, in fact. But the stack starts to drift, and at some point you look down and there’s no Earth at the bottom anymore. Just turtles.
Local journalism is one of the last institutions whose business model depends on standing at the bottom of the stack. On standing on dirt and reporting what’s actually there. In a world where most of what fills our feeds is symbols generated from symbols generated from symbols, that work has never been more valuable.
The bottom of the stack is still there. But like water or oxygen, sometimes we take it for granted.
* Full Disclosure
I used AI to help draft this article, but the ideas are all human. My brilliant daughter and I built the math late one night, providing the foundation of the core argument. I picked the sources, ran the dictionary experiment, and edited every paragraph. The AI helped with the writing, but I take responsibility for the final product. Like the circles in An Imperfect Circle, the goal isn’t to avoid AI, it’s to use it from the inside out.
This is the second article in a series exploring the intersection of technology, humanity, and community. Previous: “An Imperfect Circle,” Next: “The Map and the Territory,” looking for grounding in SEO/AEO & “Does It Get Hungry?”, on intelligence, life, and the missing internal driver that language models don’t have.