Here is something that sounds paradoxical but is structurally true: the more meaningful a word is, the harder it is for AI to learn.

Not because AI is dumb. Because of statistics.

AI learns by counting. The word "the" appears billions of times in training data. The model learns it perfectly — where it goes, what it modifies, every syntactic groove it fits. But "thing-in-itself" — Ding an sich, Kant's term for the world as it exists independent of perception — appears, maybe, a few thousand times across all the text ever written. That is statistically close to zero. The model has almost nothing to learn from.

And here is the uncomfortable implication: the concepts that matter most — the dense, load-bearing ones that philosophers have spent centuries sharpening — are precisely the concepts that are rarest in the corpus. Meaning and statistical frequency move in opposite directions.

Two Roads That Don't Meet

To understand this, you need to know something about how AI reads text.

Before a language model sees a single word, a separate system called a tokenizer slices the text into units — "tokens" — that the model will actually process. This is not a trivial step. It decides what the model is capable of seeing at all.

The dominant approach, called BPE (Byte Pair Encoding), works by frequency: it finds the most common pairs of characters and merges them into single units, then repeats. The result is a vocabulary built from statistical patterns, not from meaning. High-frequency sequences become tokens. Rare sequences get chopped into fragments.

This approach has one large advantage and one structural limit. The advantage: it scales. Give it more data and more compute, and it keeps getting better. The limit: it is built entirely on the logic of compression. It asks, how do I fit more information into fewer units? — not which units actually carry meaning?

There is a second road. Call it the semantic path: instead of merging by frequency, you merge by concept. "Thing-in-itself" should be one unit, not three — because it is one indivisible concept, and splitting it loses what it means. But this kind of merging cannot be done by statistics alone. It requires someone who actually understands the concept to draw the boundary.

These two roads are not the same road at different speeds. You can drive the frequency road all the way to its logical end and you will not arrive at the semantic road. Compression and understanding are different problems.

The Inverse Law

This is where the structural limit becomes visible.

The more meaningful a token, the more distinct it is. The more distinct, the rarer it appears. The rarer it appears, the less statistical data exists to learn from. The less data, the weaker the model's grasp.

Flip it: the less meaningful a token — pure connective tissue, grammatical filler — the more it repeats. The more it repeats, the more statistical signal exists. The better the model learns it. But it has learned nothing that carries weight.

Meaning and learnability move in opposite directions. Call it the inverse law.

This is not an engineering problem. It is not "we just need more data." The scarcity of meaningful things is a feature of the world, not a flaw in the dataset. The word "the" genuinely appears billions of times. "Ding an sich" genuinely does not. No amount of internet scraping will change that ratio, because the ratio reflects how often humans actually use these concepts.

In the framework I work with — Self-as-an-End — this is an instance of what we call remainder conservation: every act of structuring leaves something out, and what gets left out cannot be made to disappear. The tokenizer structures text. What gets left out is meaning. The more you optimize for compression, the more meaning becomes the remainder.

The Remainder That Only Humans Can See

Here is the deeper problem: the model cannot see its own blind spots.

The model operates downstream from the tokenizer. It receives already-sliced fragments. It does not know what was cut, what was lost in the cutting, what concepts were split across token boundaries in ways that destroyed their coherence. The model is building on someone else's structure and does not know what the foundation looks like.

Who can see this? Only humans.

A human who genuinely understands a philosophical concept can recognize, immediately, when it has been mangled by mechanical slicing. Not because humans have more processing power — but because human understanding is not built on frequency. It is built on having lived with an idea long enough to feel it from the inside: the nights of confusion, the sudden moment of clarity, the argument with someone who didn't get it, the slow erosion of certainty and its rebuilding.

When statistical methods fail — when a concept appears so rarely that the model has no data to learn from — a human who understands that concept can work with it from a single example. Zero-shot. The model cannot do this, not because it lacks intelligence, but because its form of intelligence is statistical and statistics requires repetition.

This is why AI's ceiling is not compute. It is not data volume. The ceiling is the human judgment that AI cannot replicate by scaling up what it already does.

What This Means for Chinese

The inverse law has a specific geopolitical consequence that deserves naming.

Because BPE training data is overwhelmingly English, English high-frequency words survive the tokenizer as single units. Chinese, by contrast, gets fragmented more aggressively — a common two-character Chinese compound that functions as one concept often becomes two or three tokens, while its English semantic equivalent stays whole.

The result: Chinese reasoning costs more tokens, runs slower, uses more of the context window for the same work. This is not because Chinese is more complex — it is because the tokenizer was built with English as the default and Chinese as the remainder.

This is not malicious. No one designed it to be colonial. But the structure is colonial: the method itself encodes a standpoint, and whatever falls outside that standpoint becomes the residue. The people who built the tokenizer were not excluding Chinese — they were optimizing for what they had the most data on. But optimizing for one thing is always, simultaneously, marginalizing another.

The Human in the Loop Is Not a Consolation Prize

A common response to AI anxiety is: "Don't worry, humans will always be needed to supervise the machines." This is usually meant as comfort — a promise that we won't be fully replaced.

What the inverse law suggests is something stronger: the human in the loop is not a stopgap. It is a structural necessity that cannot be engineered away.

The semantic path — the one that requires understanding, not just frequency — has no automated solution. The boundary between "thing-in-itself" as one concept and its three-character fragments is not a boundary that statistics can discover, because statistics does not know what a concept is. Only someone who has genuinely understood the concept can draw the boundary correctly. And if they draw it wrong — if they merge what should stay separate, or split what should stay whole — the model internalizes the error as correct structure. The mistake happens below the level where the model can detect it.

The quality of the human judgment sets the ceiling on the model's understanding. Not the amount of compute. Not the size of the training corpus. The quality of the human who looks at the fragments and says: this belongs together, this doesn't.

Self-as-an-End — humans as ends in themselves, not means to AI's performance — is not, in the end, a moral argument. It is a technical description of how the system actually works.

A Practical Consequence

If meaning and learnability are inversely related, then the most dangerous failure mode in human-AI collaboration is not the dramatic one — AI going rogue, developing its own agenda. It is the quiet one: humans gradually outsourcing their judgment.

The loop runs like this. The model gets better at producing plausible-sounding text. The better it sounds, the less humans feel the need to push back. The less humans push back, the less the model receives genuine human judgment as input. The less genuine human judgment it receives, the more it optimizes for what is already there — statistical patterns, majority signal, the well-represented — and the more the rare, meaningful, irreducible things become invisible.

The cure is not clever. It is: go first.

Before asking the model, write your own judgment. Write it badly, incompletely, hesitantly. The moment your fingers pause over the keyboard and you are not sure — that pause is the remainder knocking on the door. Do not skip it. That is exactly what the model cannot do for you.

Then show the model what you wrote. Ask it to reflect it back, to find the gaps, to push against your framing. Not to replace your judgment — to sharpen it.

Then write again, yourself.

The loop only works in that order. Reverse it — start with the model's output and edit from there — and you are no longer bringing your remainder into the system. You are becoming the model's editor. Which is useful. But it is not the same thing.

The Remainder Cannot Be Dissolved

Every technology that structures human knowledge leaves a remainder. Writing left out the gesture, the silence, the timing that spoken language carries. Print left out the handwriting, the marginalia, the individual voice. Digital text left out — we are still discovering what.

AI tokenization leaves out meaning. Not all meaning. Not irreparably. But structurally: the most meaningful things are the most statistical rare, and the most statistically rare are the hardest to learn.

This is not a reason to mistrust AI. It is a reason to understand what AI can and cannot do — and to understand that what it cannot do is not a temporary limitation waiting to be overcome by the next model release. It is a consequence of what statistical learning fundamentally is.

The remainder cannot be dissolved. It can only be carried — by humans who are willing to stay in the loop, go first, and bring what only they can bring.

We cannot help not knowing — just for now. But the not-knowing is not the model's. It is ours to work with.

This essay develops themes from the Self-as-an-End theory series. The full academic treatment appears in Why Humans Matter in the Age of AI: An SAE Architecture Application (working paper, 2026). The SAE framework is documented at self-as-an-end.net.