The Human Language Model

There is a remarkable sequence in Indiana Jones and the Last Crusade in which reaching the Holy Grail requires passing three tests before anyone is allowed to choose the cup. They are presented as ancient religious riddles, but watched from another angle they resemble something surprisingly modern: a sequence of tests designed to expose the assumptions of the human model.
The first concerns penitence, the second language, the third perception. Only after passing all three does Indiana arrive in the chamber where dozens of cups are waiting and confront the most difficult problem of all: not finding an object, but deciding which interpretation of that object is true. The sequence works because every test exploits the same human faculty that makes civilization possible. We receive incomplete information, combine it with what we already know, predict what it means and act upon the result. Most of the time this capacity is extraordinarily
useful. Occasionally, it is not.

Long before engineers built Large Language Models, human beings were language-learning models of a far more complicated kind. We do not arrive in the world with definitions of authority, beauty, success, holiness, danger, dignity, shame or belonging. We learn them. Like a child learns a word, but the word never arrives alone. It arrives attached to a voice, an expression, an object, a consequence, a story and eventually an entire network of associations. Mother says this when that happens. Teachers reward this answer and adults become uncomfortable around that question. This profession is spoken about with admiration; another with condescension. These people are called successful, those people are called foolish. This behaviour is respectable, that behaviour is embarrassing. By the time a child can speak fluently, language has done considerably more than make communication possible. It has begun constructing a model of reality.

This is the point at which the comparison with artificial intelligence becomes useful, but only if we resist the temptation to turn humans into machines. A Large Language Model encounters enormous quantities of language and learns patterns that allow it to estimate plausible continuations. A human being learns from far less linguistic material, but every sentence is grounded in something immensely richer: bodies, affection, hunger, hierarchy, embarrassment, imitation, danger, curiosity, reward, memory and consequence. We do not learn the word fire only from its statistical proximity to hot. We do not learn trust exclusively from sentences containing the word. We discover what happens when trust is honoured and what happens when it is betrayed. The human language model is therefore trained not simply on text but on language attached to life, and this distinction is precisely what makes it powerful enough to build civilizations.

Yet grounding does not make the human model infallible. It merely gives it another source of information. Humans still predict. We still fill gaps. We still construct the most plausible continuation from what we already know, and much of what we already know was inherited from people who learned it from other people. Culture is, among many other things, the accumulation of those inherited probabilities. It tells us what normally follows what: kings follow crowns, power follows wealth, wisdom follows age, authority follows titles, sacredness follows ceremony, success follows certain clothes and certain addresses. These associations are not necessarily false. The problem is a little subtler. Because they are often true, they acquire the authority to speak even when they are not.

This is where the three trials before the Grail become interesting. The first is the Breath of God:
“Only the penitent man will pass.” Indiana initially treats the phrase as information to be decoded,but survival requires something else. A penitent man kneels. The word must become action.

Language is not being tested as vocabulary; interpretation is. 

Knowing what penitence means abstractly is insufficient. Indiana must recover the embodied meaning hidden inside the concept. The distinction matters because civilizations are filled with words whose definitions everyone knows while their operational meaning has slowly disappeared. Justice, responsibility, freedom, dignity, service, faith, leadership, societies can become extraordinarily fluent in their vocabulary while becoming increasingly uncertain about what the words require in practice. Language survives remarkably well after meaning has begun to erode.

The second trial is even more explicitly linguistic. Indiana must cross a floor of letters by spelling the name of God. He begins with the familiar spelling of Jehovah and nearly falls because the alphabetic convention encoded in the puzzle is not his own: in Latin, the name begins with an I. The mistake is small enough to appear trivial, yet conceptually it is enormous. He knows the answer and can still be wrong because he has applied the wrong linguistic system to it. This is something humans do constantly across cultures. 

We assume translation transfers meaning when often it transfers only vocabulary.

The same word can carry different histories, moral weights, hierarchies and expectations in different societies. Ambition can sound admirable in one environment and vaguely indecent in another. Individualism can signify liberty or selfishness. Tradition can evoke continuity, identity, authority, comfort or stagnation. A dictionary can translate the token. It cannot automatically translate the civilization surrounding it.
Then comes the Path of God. Indiana reaches an abyss where there is no bridge. Everything his eyes tell him agrees: one more step means death. And yet a bridge is there, camouflaged so perfectly against the rock that perception cannot distinguish it from emptiness. He must act against the prediction generated by his own visual system before new evidence becomes available. Once he crosses and looks back, the bridge becomes obvious. Reality did not change. His position did. 

There are few better illustrations of the difficulty of revising a human model. We routinely mistake the limits of our perspective for the limits of reality. What cannot be seen from where we stand becomes impossible; what has never occurred within our experience becomes absurd; what our cultural model has no category for becomes suspicious. Civilization advances partly because someone occasionally walks far enough to discover that the abyss contained a bridge.
Only after penitence, language and perception have been tested does Indiana enter the chamber of cups. And now culture itself becomes the test.
The Grail Knight gives no description. There are elaborate chalices, precious metals, ceremonial objects and ordinary vessels. Donovan chooses exactly as a well-trained cultural model might predict. He selects magnificence. The cup of the King of Kings should surely be worthy of a king: ornate, precious, exceptional. Gold completes the sentence perfectly. Indiana looks at the same collection but changes the question. What would the cup of a king look like? or What would the cup of a carpenter look like? The probability distribution changes. His attention moves away from splendour toward an unremarkable vessel. The information was available all along. Christ was king in one story and carpenter in another. The error was not ignorance. It was weighting.

This may be one of the most important things language teaches us without our noticing. Words do not merely describe objects; they assign probabilities to them, you know, or statistical relevance. Say king and the imagination begins decorating the room. Say carpenter and it removes the gold. Say expert and a voice, vocabulary and appearance begin forming before an actual person has entered. Say successful, dangerous, holy, educated, modern, backward, elite, ordinary, and an enormous amount of supposedly missing information suddenly appears. We experience the completed image as intuition because we rarely see the thousands of cultural observations from which it was assembled.

This is how language becomes culture and culture becomes civilization. Civilization is not created only by constitutions, roads, markets, armies, cathedrals or technologies. Before any of those structures can persist, a population must share enough predictions about the world to coordinate behaviour. Money works because enough people agree about value. Institutions work because enough people recognize authority. Contracts work because words can create expectations extending into the future. Education works because societies decide which knowledge deserves transmission. Reputation works because communities attach meaning to behaviour. Even rebellion requires a common language against which rebellion becomes intelligible.

Civilization is possible because millions of human models overlap sufficiently to create a shared reality.

But the same mechanism that creates civilization can also preserve its errors. Once a prediction is culturally stable, every generation can receive part of the previous generation’s output as new training material. A society can repeat that certain people belong in certain rooms until their presence elsewhere feels anomalous. It can describe wealth as evidence of competence, poverty as evidence of failure, suffering as evidence of virtue, complexity as evidence of intelligence, ceremonial language as evidence of legitimacy, or obedience as evidence of morality. 

None of these conclusions needs to be explicitly taught as doctrine. Repetition is often enough. 

Eventually statistical familiarity begins wearing the clothes of natural law.

Humans therefore hallucinate too. Not usually in the clinical sense, and not exactly as machines do, but in a sense increasingly useful for understanding culture: we complete incomplete information with the most plausible story available to our model. We meet someone and quietly manufacture the missing biography. We hear confidence and supply competence. We see expensive architecture and infer importance. We encounter ritual and infer meaning. We see hardship and sometimes invent a moral purpose for it. We encounter coincidence and supply intention. We hear the same claim often enough and repetition begins to feel strangely similar to evidence. The surface comes from reality, but the completion eventually comes from us.

None of this makes prediction defective. Without it, human existence would be almost impossible. We cannot verify the entire universe every morning before deciding whether to leave the house. Intelligence requires compression and culture is an astonishing compression system: centuries of observation, mistakes, adaptation and knowledge can be transmitted through stories, customs and language to someone who did not personally experience any of them. A child can learn not to touch fire without rediscovering combustion. A civilization can preserve engineering knowledge beyond the lifespan of any engineer. 

Language allows the dead to contribute training data to minds that do not yet exist. This is one of humanity’s greatest achievements.

The danger begins elsewhere: when we forget that compression loses information, when the model becomes invisible to the person using it, and when plausibility acquires the status of truth.

Artificial intelligence has unexpectedly given us a vocabulary for examining this problem. When a machine generates an elegant falsehood, we ask what it was trained on, why this continuation had such a high probability, whether its sources were grounded, whether the context was sufficient and why the model expressed confidence beyond its evidence. These are excellent questions, and perhaps the great cultural gift of AI will be that we eventually learn to ask them about ourselves.

What trained this conclusion? Which association made it feel obvious? Is this observation, or completion? Am I looking at the object in front of me, or at everything my language taught me an object of this kind should be?
That question becomes especially important when societies attempt to change themselves.
Changing vocabulary is relatively easy; changing the model beneath it is not. Institutions can adopt new terminology while preserving old assumptions. Political systems can rename categories without changing their distribution of power. Corporations can replace entire dictionaries of management language while employees continue predicting precisely the same rewards and punishments. Moral vocabulary can modernize while ancient status mechanisms quietly migrate into the new words. Humans are remarkably capable of learning new tokens while running an old model. 

Civilizational change therefore cannot be measured simply by the language a society adopts.

We have to ask which predictions survived the translation. And perhaps this is where enlightenment belongs in the story. Enlightenment is often described as accumulation: more knowledge, more education, more information, more sophisticated language. But Indiana does not survive because he continuously accumulates facts. In each trial, he must correct something. Penitence requires him to move from word to embodied meaning. The Word requires him to recognize that his familiar linguistic convention is not universal. The Path requires him to distrust the completeness of his own perception. The Grail requires him to revise the cultural probability attached to greatness. Each step removes certainty rather than adding ornament to it.
Enlightenment, in this sense, is not the absence of a model. No human being can live without one. It is the ability to see the model while using it. That ability may be one of the highest forms of cultural intelligence. To understand that one’s language contains history without assuming history is destiny. To inherit civilization without becoming imprisoned by its categories. To recognize the extraordinary value of cultural knowledge while retaining the ability to inspect its predictions. To know that intuition can contain generations of accumulated wisdom and generations of accumulated error at exactly the same time.

We must understand that another civilization may not be irrational simply because its sentence completes differently from ours.

This also changes what enlightenment means collectively. An enlightened civilization would not be one that finally discovered the perfect vocabulary and imposed it forever. That would merely create another golden cup. It would be a civilization capable of examining its own model: preserving what remains grounded, correcting what reality has falsified, distinguishing inherited wisdom from inherited probability and allowing evidence to update identity without treating every correction as destruction. Civilization needs continuity, because no society can rebuild its knowledge from zero every generation. But it also needs revision, because a model that cannot update eventually stops describing the world and begins demanding that the world resemble the model.
Perhaps that is why the Grail sequence remains so compelling. Beneath the adventure story lies an elegant progression through the architecture of human understanding. First, understand what a word requires. Then understand the system in which the word operates. Then confront the limits of perception. Finally, confront the seduction of cultural plausibility. Only after all of that are you ready to choose.

People spend much of our lives reaching for golden cups. Just because somebody taught them what gold means. The lesson is not to distrust language, abandon culture or reject civilization. They are among the most extraordinary technologies humanity has ever created.
The lesson is to understand what they do. Language gives us categories. Culture gives those categories probability. Civilization stabilizes them across generations. 

Enlightenment gives us the possibility of looking at the entire construction and asking, perhaps for the first time, whether the thing that feels most plausible is also the thing that is actually there. The human language model will always predict. Wisdom may begin when we know when to let reality answer instead.