Semantic satiation and drift of meaning: why continuous ingestion follows

The word that stops meaning anything

Say a word aloud once and it means what it always meant. Say it thirty times in twenty seconds and something odd happens: the sound comes loose from the sense. You hear "spoon" enough times and it turns into noise wearing a spoon-shaped costume. The meaning does not vanish permanently. It goes inert, briefly, and returns when you stop. Psychologists call this semantic satiation: a temporary fatigue in the link between a token and what it denotes.

That is the fast version, measured in seconds. There is a slow version, measured in decades, and it is more consequential because it never announces itself. A word can keep its spelling, its pronunciation, its grammatical behaviour, while the thing it points at moves. "Awful" once meant awe-inspiring. "Meat" once meant food in general. Historical linguists call this semantic change, or drift: narrowing, widening, shift in register, the quiet substitution of a technical definition for an ordinary one. The mechanism differs from satiation — this is not fatigue but migration — but the surface symptom is the same. The token gives no warning. You cannot tell from looking at the word whether it still means what it meant.

That absence of warning is the entire problem this page is about. A system, or a person, that learned a word's meaning at one moment and stopped updating carries that meaning forward with full confidence and no internal flag that anything has changed. The error is not a gap. It is a false positive dressed as fluency.

Origin: two lines of work, one discovery

Edward Titchener described the satiation effect in 1915, calling it a "lapse of meaning" under repetition, part of his broader programme of introspective psychology at Cornell. The effect sat unnamed and half-studied for decades until Léon Jakobovits James and Wallace Lambert gave it its modern name and its laboratory measurement in 1960, running repetition tasks and quantifying how fast and how far the sense-fatigue went before recovery. That is the fast timescale: seconds to minutes, reversible, a property of a single listener's nervous system.

The slow timescale has an older and separate pedigree. Michel Bréal's Essai de sémantique (1897) laid out the mechanisms by which words change reference over generations — narrowing, widening, metaphorical transfer — founding semantics as a historical discipline in the process. Stephen Ullmann systematised the taxonomy in the mid-twentieth century; Elizabeth Traugott extended it into grammaticalisation, showing how content words hollow out into grammatical markers over centuries. The field stayed largely qualitative until 2016, when William Hamilton, Jure Leskovec and Dan Jurafsky trained word embeddings separately on each decade of large historical corpora and measured drift directly, as movement in vector space, deriving statistical laws relating a word's frequency and polysemy to how fast it changes. Two traditions, a psychology lab and a century of historical linguistics, converge on the same fact: the token is stable, the meaning is not, and nothing on the surface tells you which state you are in.

The turn

Here is where the two facts about words stop being a curiosity of human cognition and become a design constraint.

A Large Language Model learns meaning as geometry over a corpus fixed at a cutoff date. The embedding for a term is, roughly, an average over that term's uses up to that date. This is not a metaphor for how the model behaves; it is close to a description of the training objective. If the referent of a term has moved since the cutoff, the vector still points confidently at the old sense, because nothing in the objective ever told it otherwise. The output is fluent. The grammar is correct. The definition is wrong, and wrong in a way indistinguishable, from the outside, from being right. This is semantic drift transplanted wholesale into an architecture, at industrial scale, silent by construction.

Sepsis is the clean case. In 2016 the Sepsis-3 consensus discarded the older systemic inflammatory response criteria and redefined sepsis as life-threatening organ dysfunction, scored by the SOFA index. The word did not change. The referent did. Clinical text written before 2016 and after 2016 uses one token for two different clinical realities, and a model trained across that boundary averages them into a single, confidently wrong composite. Pluto is the public case: on 24 August 2006 the International Astronomical Union redefined "planet" to require clearing the orbital neighbourhood, and pre-2006 text — which still dominates the written record — is internally coherent and simply outdated. LIBOR is the financial case: the last US dollar tenors ceased on 30 June 2023, yet the phrase "the benchmark rate" survives in decades of contracts and analysis referring to an instrument with different economics under a different name, SOFR. In every case, the surface offers no warning.

A Large World Model improves on this in one specific way: it can re-anchor a word to an object present in a current scene. Point the system at the thing, and the word attaches to what is actually there now, not to an average of past mentions. That is a real repair, and it should not be understated. But it only works for vocabulary with a visible referent. Most of the vocabulary that carries institutional weight — regulatory capital, clinical diagnosis, policy wording, benchmark rate — is defined by committee, not by sight. You cannot point a camera at Tier 1 capital and see what Basel III made it mean.

A Large Universe Model, on this argument, treats meaning itself as a dated, sourced, revisable belief: this term meant that, under this authority, between these dates. Redefinition is not noise to be averaged away. It is an event, ingested when it happens, filed with provenance, superseding the prior reading without erasing it. Because intake never closes, the gap between a redefinition occurring and a system knowing about it shrinks to the latency of the stream, not the interval between retrainings.

What continuous intake buys, and what it does not

This is where the misreading needs to be named and refused. The weak, common version of this argument says: language models fail on new words, and the fix is fresher training data. Both halves are wrong. An unfamiliar token is the easy case — it announces itself, and a well-built system can simply say it does not know. The hard case is the opposite: an old, familiar word whose referent has moved, producing fluent, well-formed, obsolete meaning with no signal of staleness anywhere in the output. Fresher data does not fix this either, unless the new usage is stored with a date and a source. Otherwise it just blends into the running average of the old sense, and the drift continues, only slightly delayed.

Three objections deserve to be taken on directly, because they narrow the claim in ways worth keeping.

Retrieval augmentation already solves this. Put current documents in context and the model reads the up-to-date definition.

This is correct for queries you know to ask. It fails exactly where drift is silent: the model has no reason to retrieve a redefinition of "sepsis" because it does not know the word has been redefined. Retrieval repairs missing facts. Drift is a fault in the interpretation of the query itself, and no retrieval step fires when nothing looks wrong.

Semantic change is slow. Quarterly retraining outruns it easily.

This is true on average, and the 2016 drift-rate measurements support it: frequent, general vocabulary moves slowly per decade. But the terms that carry decisions are not general vocabulary. Regulatory and clinical definitions change by fiat, overnight, with a stated effective date. The distribution of drift is heavy-tailed, and the tail is precisely where consequences concentrate. This genuinely narrows the claim: for most of the lexicon, the problem is minor. For the vocabulary that governs contracts, diagnoses and capital, it is not.

Unbounded intake destabilises meaning. An audit needs a fixed reading, not a moving target.

Against a naive design that overwrites old readings with new ones, this objection is simply right. The claim being made here is narrower than that design: meaning stored as dated, superseded claims, with old and new readings both retained and provenance attached to each, is more auditable than a frozen model, not less — because a frozen model asserts one undated reading and calls it truth.

What this establishes, and what it does not

This concept establishes that fixing intake at any cutoff embeds a silent, undetectable class of error into every term whose institutional definition later changes, and that grounding in a present scene repairs only the subset of vocabulary with a visible referent. It does not establish that continuous intake is easy, cheap, or currently built at scale — the Large Universe Model here is an argued category, not a working system. It does not establish that all drift matters equally; most of it does not. It establishes only that after continuous, provenance-carrying observation, there is no further axis of intake left to climb — what remains is coverage, latency and trust, not a new kind of evidence.