A Fact Is Shared Belief

A Fact Is Shared Belief

ai blog
July 8, 2026

Joel Kowalewski, PhD

The most obvious fact in the sky

For roughly fourteen centuries, the most obvious fact was that the Earth stood still and the heavens turned around it. It was tangible. It was perceived. The sun rose in the east and set in the west; the stars swirled overhead according to their fixed, mathematical schedule; the ground beneath us did not so much as tremble. Around 150 CE the astronomer Claudius Ptolemy, working in Alexandria, gave this fact its mathematical form in the Almagest, and the system worked — it predicted the positions of the planets well enough to guide navigation and set the calendar for more than a thousand years. When the planets misbehaved, drifting backward against the stars in what we now call retrograde motion, astronomers did not abandon the fact. They defended it, adding epicycles — small circles riding on the larger orbits — and then epicycles upon epicycles, an ever more intricate scaffolding whose only purpose was to keep the obvious fact standing. The fact was wrong. The scaffolding was ingenious. And almost no one could tell the difference, because the fact and the evidence for it had grown up together

I begin in the sky over Alexandria because the question I want to ask — what is a fact? — is usually treated as too simple to ask, and the history of astronomy is the cleanest proof that it is not. We are now living through a moment when a new kind of machine is being sold to us as an oracle of facts or not: ask it anything, and it returns a fluent, confident answer drawn from the accumulated text of trained experts. An alternative take is that, far from providing anything resembling fact or truth, AI finds the mathematically optimal set of symbols that matches or completes the pattern of symbols in the user’s prompt. If the prompt is out-of-distribution, containing words/symbols unlike its training data, there is no optimal pattern completion. It provides an authoritative guess.The user who asks in ignorance cannot distinguish between sensible and nonsense responses.

Before we can judge whether AI can tell us the truth or is problematic precisely because it cannot, we must carefully define truth and fact. My argument is that facts are fundamentally shared beliefs. They are not, as we often imagine them to be, immutable universals. The earliest astronomers did indeed record discrepancies in their measurements of the night sky. They just chose to disregard or fit them to the facts of the geocentric model. As a disinterested observer, we could claim that these discrepancies represent the closest thing to our current thinking about the immutability of truth and fact. Ironically, this would, however, imply that whatever is of negligible consequence to the group is truth in-disguise. The case of epicycles teaches us that what is universal and unchanging is the thing we cannot explain. What we actually call “fact” today is much closer to a shared belief that the discrepancies are irrelevant. I think this has significant implications for how we interact with conversational AI systems, as the measuring stick is not factualness. We should derive value from the AI system because it rejects factualness.

My overall take is that truth is a commitment to a method of inquiry (commitment to questions above anything else) and this, perhaps oddly, turns the AI system into a truth machine, excavating questions for us. Should we no longer know where our commits lie (e.g. our version of geocentrism and epicycles) and are unmoored in the 21st century we are inclined to follow a method of inquiry (to believe in it) to wherever it might lead. And this is more productive than percentages, averages, and charts, which are only as useful as our capacity to challenge what they lack. For instance, beyond their apparent factualness to the discrepancies.

What we mean when we say “fact”

Let us start with the word itself, because a great deal hides inside it. We use “fact” to mean a piece of the world reported without distortion — the thing as it really is, independent of anyone’s opinion. That picture is so intuitive it feels like bedrock. But in 1935 a Polish physician and philosopher named Ludwik Fleck wrote a small, neglected book, Genesis and Development of a Scientific Fact, that took the bedrock apart. Studying how the medical “fact” of syphilis and its blood test came to be, Fleck argued that a scientific fact is not discovered by a lone observer but produced by a thought collective — a community with a shared style of thinking that decides, in advance, what counts as a legitimate observation and what counts as noise. The fact is real; it is also manufactured, in the precise sense that it could not exist without the community and its method. Thomas Kuhn, who read Fleck, would later make the same point famous under the word paradigm in The Structure of Scientific Revolutions (1962): normal science spends its days, as the Ptolemaic astronomers did, refining the reigning picture, until the anomalies pile up and the picture breaks.

This is the first thing I want to establish, and it is not a mystical claim. What we call a fact is inseparable from the procedure that certified it. Change the procedure — the instrument, the community, the question — and the inventory of facts changes with it. The Earth did not start moving in 1543 when Copernicus published his famous rejection of the geocentric model; what changed was the method by which we decided where to stand and what to measure. When Copernicus first proposed a sun-centered system, it did not predict the planets any better than Ptolemy’s. It won because it was simpler and more fruitful, not because it was, on day one, irrefutably more “factual.” It was aesthetically appealing though contrary to the facts of Christian theology, and therefore it was never actually settled as fact during Copernicus’ lifetime, as he was a scientist but also a devout believer. He endorsed a view whereby his work was a method for elegantly calculating the movement of stars and planets. His is an interesting case as he is often lauded as a scientific hero for the wrong reasons. His commitment to methodological rigor and confidence in the accuracy of his measurements was heroic. But he importantly remained a man of his time, struggling with the implications of the heliocentric model as much as anyone else. Science just made the tension real and unavoidable.

Science argues; it does not prove

In the past, the predominant method of truth construction was theological. It began with creation and a divine, perfect creator; this “theological method” was certified by the creator. If facts are certified by a method and there is no longer the assumption of perfection or divinity, then everything turns on what that method can and cannot do — and here we have to be exact, because the popular image of science as a machine that stamps claims TRUE is simply wrong. Karl Popper saw this most clearly in The Logic of Scientific Discovery (1959): no quantity of confirming observations can ever prove a general claim, because the next observation might refute it, but a single solid counterexample can knock one down. Science, on this account, never proves the positive claim. What a working scientist actually does is more modest and more adversarial: she gathers a result and argues that it is unlikely to be an accident, and her colleagues argue back. What we grandly call the body of scientific knowledge is, seen up close, a collection of battling interpretations that have so far survived the fight. That is not a weakness to be embarrassed about. It is the entire source of science’s strength, and it is exactly what a confident one-paragraph answer erases.

Notice what follows for the machine. When a large language model returns an answer, it is not conducting this argument; it is reporting the settled-sounding output of arguments that happened elsewhere, stripped of the dissent, the error bars, and the live disagreement that made the knowledge trustworthy in the first place. It gives you the verdict without the trial. The danger is not primarily that it is sometimes wrong — people are wrong too. The danger is that it launders interpretation into fact by presenting the residue of a contested process as if it were a reading taken directly off the world.

The method is objective; the world is not

It is easy to mishear this as relativism, and it is the opposite of relativism. What is objective is the method, not the world. Truth is what we construct through a disciplined procedure of inquiry, and the discipline — the willingness to be refuted, the shared rules of evidence, the argument held open — is the thing we can defend as objective. The “bare facts” the procedure delivers are always interpretations, and we rarely have access to anything more direct. Consider Lamarck, long treated as the textbook example of a scientist who was simply, factually wrong about heredity; the rise of epigenetics has shown that the inheritance of acquired traits never in fact had a probability of zero. The lesson is not that anything goes. It is that our confidence should live in the quality of the method, the diversity of questions we ask, and our commitment to the method over answers.

To deny the facts objectivity is not to say that all views are equal or that truth is whatever you can get away with. Alternative views do not always collapse to a probability of zero, but neither do they all stand equally; a good method sorts them, slowly and provisionally. William James, in Pragmatism (1907), located truth in what works — the “cash-value” of an idea in the flow of experience — and was accused of surrendering to relativism, yet he insisted we must still pursue universal truth as a regulative goal. That is an unending and challenging path that we must walk: truth is constructed, and some constructions are far more useful descriptions about the world than others. Method is the compass that points to new opportunities to meaningfully construct and evaluate our descriptions of the world. The formalization of this is called “scientific training.”

Objectivity has a history

It helps to remember that even our craving for a view from nowhere is a dated invention rather than a timeless standard. In their history Objectivity (2007), Lorraine Daston and Peter Galison trace the very ideal of “mechanical objectivity” — the scientist as a self-effacing instrument who adds nothing of himself — to the middle of the nineteenth century, arriving alongside the photograph and the standardized atlas. Before it, scientific images openly relied on the trained judgment of the expert who knew which specimens were “typical” examples; objectivity, in our modern sense, was not the eternal essence of science but a scientist providing a plausible description, given the data and beliefs at that time. If the beliefs about science made the statistical average truth, “atypical” becomes irrelevant. We only reclaim objectivity in our awareness of these various descriptions of objectivity throughout history.

Objectivity has a history, and it is full of surprises.

Lorraine Daston and Peter Galison, Objectivity (2007)

When knowledge runs thin, truth becomes a logo

Where a question is genuinely settled by an accessible method, we can point to the procedure and check the work. But most of what we argue about lies past that edge, where the experimentally verifiable claims are quickly exhausted and something has to fill the vacuum. What fills it, reliably, is the authority of the source. We saw this during the pandemic, when public truth came to rest less on any argument the audience could follow than on the credentials of the expert being quoted; the spectator assigns validity to a claim because it came from MIT, or from a name they trust, not because they have walked the reasoning. Truth collapses into its letterhead.

An artificial intelligence is the purest instance of this failure yet built, and I say that as someone who thinks these systems are genuinely remarkable, not as someone waving them away. When an AI is marketed as trustworthy because it was trained on the writing of experts, we are being asked to accept its output as fact on the strength of its source — argument from authority, wearing a new and frictionless costume. It hands us the verdict with no trial and a very good brand. And the brand may be wrong in the deepest way: today’s expert consensus, confidently retrieved and fluently phrased, may turn out to be our version of geocentrism, complete with its own epicycles that we mistake for rigor because they are so elaborate. The problem is not that the machine lies. It is that it lets us stop doing the one thing that ever produced truth — the argument, the dialogue, the commitment to questions.

Truth becomes a commodity

There is a further turn, and it is where the epistemology becomes political, because facts have never floated free of the institutions that pay for them. The moment inquiry is fully commercialized, its findings can no longer be received as truth in the old sense; they arrive as product. Jacques Ellul warned in The Technological Society (1954) that the ruling logic of modern life is technique — the relentless optimization of means for efficiency, spreading from the factory into science, governance, and conversation itself. Under that logic a scientific result becomes a kind of documentary: produced, packaged, and consumed, persuasive precisely because it is well-made, and no longer straightforwardly true. And “facts” become a rhetorical device in politics, statements engineered to look like neutral readings of the world while doing the work of persuasion. Once truth is a commodity, the question stops being “is it true?” and quietly becomes “does it sell?” — which is the same engagement logic now tuning the machines.

Hold two things at once

What, then, is the healthy stance? Not the search for bedrock facts, and not the shrug of relativism, but something harder that I have come to think of as the whole discipline of thinking well: the capacity to hold two contradictory positions at once without collapsing them into one. There is a class of optical illusion — the drawing that is both an old woman and a young woman — that never resolves; you can flip between the readings but you cannot make one of them the fact. The natural world, I’d argue, is more like that non-collapsing illusion than like a puzzle with a single answer at the back of the book. If you can entertain and even accept two rival interpretations at the same time, what you gain in critical power is worth more than any reassurance about your sources. This is also why the relational view I keep returning to matters here: “traffic” is not a fact about cars but a pattern in the relationships between them, and it is no less real for that. Truth, at its most solid, is relational — a pattern we maintain, not an object we hold.

So: what is a fact? It is the durable residue of a good argument — provisional, constructed, and answerable to a method we can defend, rather than a bare piece of the world we happened to pick up. The danger of the new machines is not that they will lie to us, though they will sometimes. It is that their fluency invites us to outsource the method itself — to accept the verdict and skip the trial, to take the epicycle for the orbit because it is rendered so beautifully. We should keep telling our stories and using our models, and we should never forget that this is what we are doing. Ptolemy’s sky was not a lie. It was an honest, brilliant, and wrong account of what everyone could plainly see — and the only thing that ever moved us past it was the argument we were willing to keep having. A fact was never a thing you find. It is a relationship you keep in good repair.