Here is a paradox worth treating as a warning rather than a universal law: the more singular and context-bound a meaning is, the less direct statistical support a model may have for learning it.

Not because AI is dumb, and not because frequency is the opposite of meaning. The tension arises because distributional learning is strongest where many examples constrain a pattern, while some of the meanings that matter most to a person, discipline, or culture appear only rarely and in changing contexts.

Language models do not merely count words. They learn parameters by predicting patterns across enormous corpora and can compose familiar pieces to handle expressions they have rarely or never seen. Still, contrast "the" with "thing-in-itself" — Ding an sich, Kant's term for what exists independently of our mode of appearance. The first supplies abundant evidence about ordinary usage; the second demands a sparse and contested philosophical context. A model can say a great deal about both, but equal fluency does not imply equally grounded understanding.

The uncomfortable implication is narrower and more defensible: dense, load-bearing concepts often receive less direct supervision than ordinary connective language. Statistical frequency does not measure meaning, yet it shapes which distinctions are easiest for a model to stabilize.

Two Roads That Don't Meet

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

Before a language model processes text, a tokenizer maps it into units — "tokens" — handled by the network. This is not a trivial step. Tokenization affects efficiency, sequence length, and the representation from which learning begins, although a model can represent concepts across several tokens.

One widely used family of approaches, derived from BPE (Byte Pair Encoding), builds a vocabulary by repeatedly merging frequent byte or character sequences. Related tokenizers use different algorithms, but the trade-off is similar: common sequences tend to receive compact representations, while rare sequences may be split into several units.

This approach has one large advantage: it scales without requiring a human to define every concept in advance. Its limit is not that splitting a word destroys meaning automatically; transformers routinely compose meaning across multiple tokens. The limit is that the vocabulary boundary itself is optimized for statistical utility, not offered as a theory of which concepts matter.

There is another task we can call the semantic path: deciding that "thing-in-itself" should be treated as one concept even if its written form occupies several tokens, locating its disputes, and judging when a use is faithful or superficial. Statistical learning can contribute to this task through composition and context. It does not make the task identical to token compression.

The two roads interact, but they do not collapse into one. Better compression can support learning; it cannot by itself decide which distinctions deserve conceptual weight.

The Inverse Law

This is where a structural tension becomes visible.

Some highly significant meanings are also rare: a private vow, a new scientific distinction, a minority idiom, a philosophical term used precisely rather than decoratively. Rarity gives a learner fewer direct examples. But rare does not always mean meaningful, common does not mean empty, and models can generalize compositionally. Frequency is therefore not a law of meaning; it is one pressure on learnability.

Call the warning the inverse law: as direct statistical support decreases, human judgment about context and significance becomes more important. It is a heuristic, not a theorem assigning a fixed meaning score to every token.

More and better data can help, as can retrieval, tools, curated examples, multilingual training, and improved architectures. Yet some scarcity is not merely a dataset accident. New, local, contested, and deeply personal meanings may remain rare because people actually use them rarely. Scaling changes the boundary; it does not guarantee that every remainder disappears.

In the framework I work with — Self-as-an-End — this is an instance of what we call remainder conservation: every act of structuring makes some distinctions cheap and leaves others costly or obscure. Tokenization does not erase meaning wholesale. It is one of several structures through which meaning must pass, and its optimization target is not identical to meaning.

The Remainder a Human Position Can Expose

Here is the deeper problem: no learner sees all of its own blind spots from inside the same representation.

A model operates downstream from a tokenizer and training pipeline it did not choose. It can often infer a multi-token concept from context, and with tools it can even inspect tokenization. What it cannot obtain from fluent continuation alone is an independent guarantee that the distinctions favored by its pipeline are the distinctions that matter in this case.

Humans do not possess that guarantee either. What they can contribute is another position: lived stakes, disciplinary practice, minority usage, and the ability to say that a statistically plausible reading has missed the point.

A human who genuinely understands a philosophical concept can sometimes recognize when a mechanical representation has flattened it. Not because humans have more processing power, or because human learning is free of frequency, but because it is not built on frequency alone. It can include years of disciplinary practice, lived stakes, confusion, argument, and revision.

Humans can sometimes learn a new distinction from one carefully situated example because they bring bodies, projects, shared practices, and prior concepts to it. Models can also perform striking few-shot and zero-shot generalization, but they bring a different history. The practical question is not which learner uses "statistics" and which does not; it is which background makes the relevant distinction available and which hides it.

Compute and data therefore matter greatly, but they are not the only variables. At the frontier of sparse or contested meaning, the quality and plurality of human judgment can set limits that scale alone does not automatically remove.

What This Means for Chinese

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

Tokenizer efficiency differs by model and training mixture. In some widely used vocabularies — especially earlier or English-heavy ones — Chinese text is represented with more tokens than comparable English text. Newer multilingual tokenizers have narrowed the gap, and no single ratio describes every system.

Where the gap persists, Chinese can consume more of a context window and incur greater cost for comparable work. This is not because Chinese is intrinsically more complex. It reflects design choices, vocabulary allocation, and the distribution of training data.

This need not be malicious. A method optimized around one distribution can reproduce center–periphery effects: what is efficient near the statistical center becomes costlier at the edge. Tokenizer design can therefore disadvantage Chinese without anyone deliberately deciding to exclude it. Optimization creates trade-offs; their distribution should be measured rather than naturalized.

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 but more limited: in work where stakes, novelty, or local meaning matter, a responsible human role should not be treated as a temporary inconvenience to be engineered away.

The semantic path — judging what a concept means here and why it matters — is not solved merely by assigning it one token. Automated systems can support that judgment, retrieve its history, and expose competing uses. They can also reproduce a dominant reading while missing the case that gave the question its stakes. Human responsibility begins where plausible continuation is not enough.

The quality of human judgment can therefore set a ceiling on the quality of the whole human–AI process, just as poor data, weak architecture, or inadequate compute can. The point is not that only humans can ever draw a conceptual boundary. It is that someone must remain answerable for why this boundary, in this context, deserves to govern.

Self-as-an-End — humans as ends in themselves, not merely means to AI's performance — remains a moral and institutional claim. The technical architecture does not prove it. What the architecture shows is why abandoning human answerability would also be epistemically costly.

A Practical Consequence

If sparse and situated meanings receive less direct support, then one dangerous failure mode in human-AI collaboration is quiet rather than dramatic: humans gradually outsourcing their judgment.

The quieter risk lies in the workflow rather than in a model secretly retraining itself during the conversation. The less a workflow is exposed to independent human judgment, the more its outputs remain anchored to already represented patterns, majority signals, and whatever the system can make most plausible. Rare or situated meanings can then disappear without anyone noticing the loss.

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 and statistical training leave remainders — not all meaning, not always, and not irreparably. The narrower claim is enough: rare and situated meanings can be harder to stabilize, while fluent language can conceal that weakness.

This is not a reason to mistrust AI or to freeze its present limits into metaphysical law. It is a reason to distinguish gains that scale can plausibly deliver from questions that still require situated judgment, better evidence, and answerability.

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.