SJ Waller | Space Software Sound

Abstract

The term "AI" now spans a range so wide — from a trivial conditional statement to a frontier model — that it has largely ceased to convey specific information. This paper treats that familiar fatigue not as a complaint to be vented but as an instance of a documented linguistic process — semantic bleaching, the loss of specific descriptive content as a term broadens in use — accelerated by a market that rewards the broadest possible application.

Rather than beginning from "AI" as a stable category, the paper returns the word to the minimum condition under which a term carries information at all: the capacity to rule things out. It argues that "AI" has been bleached to the point where it rules almost nothing out, that the cost is epistemic rather than merely aesthetic — degrading public reasoning about capability, safety and regulation precisely where the trivial and the consequential are forced to share a name — and that the damage is partly repairable, not by reclaiming the word but by routing around it with more specific vocabulary in high-stakes contexts. The paper offers a discriminator for when imprecise use is harmless versus harmful, and applies its own demand for precision to its own key terms. The contribution is diagnostic: a precise account of what has gone wrong with a word, and a bounded proposal for what to do about it.

1. Introduction

The complaint is itself a cliché: the word "AI" is everywhere — on every product launch, every deck, every job posting. It is attached to so much that encountering it tells you almost nothing about what is inside the thing it describes.

Fatigue, though, is not an argument. This paper attempts to say precisely what has gone wrong with the word, why, what it costs, and whether anything can be done. The diagnosis is that "AI" is a textbook case of a process linguists have studied for decades, and that naming it as such is the first step toward using language more responsibly where it matters.

The framing concern is structural. A dominant term, like a dominant paradigm, can come to define not only the answers but the questions (Kuhn, 1962): once "AI" is the accepted label for a vast range of systems, debates inherit the label's vagueness without noticing. The aim here is to suspend the assumption that "AI" is a single informative category and return it to the minimum condition under which any word informs — then test whether restoring that condition changes anything we can observe in how debates actually go.

2. Terminology and Scope

Semantic bleaching is the gradual loss of specific descriptive content as a term is used across an ever-wider range of contexts, studied chiefly within work on grammaticalization (Sweetser, 1988; Bybee, 2003; Hopper and Traugott, 2003).
Informativeness is used in its plain information-theoretic sense: a term informs to the degree that hearing it lets the listener rule things out. A term applying to nearly everything in its domain carries nearly no information about its referent.
Category signal names the residual function a bleached term retains — marking something as belonging to a valued class (modern, advanced, investable) — after its descriptive content has thinned.

These are used as analytic tools, not as established claims about "AI" specifically; the paper's task is to show that the tools apply. The scope is the public and commercial use of "AI", not the technical use within the academic field, where "artificial intelligence" has long been an accepted broad umbrella and causes less harm because its users share context.

3. Where This Sits: Relation to Existing Work

The observation that "AI" is overused is common; the attempt to give it a rigorous diagnosis is less so.

Grammaticalization and semantic change (Sweetser, 1988; Bybee, 2003; Hopper and Traugott, 2003). This is the established science of how words lose descriptive content. The canonical cases are mundane — intensifiers, auxiliaries, discourse markers — and the mechanism is well understood. This paper's contribution is not to extend that science but to apply it to a contemporary term and to add the incentive analysis (§4) that explains the unusual speed.

Critiques of marketing and buzzword language (popular and applied). There is a long tradition of complaint about hollow corporate vocabulary ("synergy", "cloud", "smart"). This paper aims to be more than complaint by grounding the phenomenon in the linguistics and by specifying the epistemic (not merely aesthetic) cost.

Philosophy of science on conceptual hardening (Kuhn, 1962). The idea that a dominant term shapes which questions are askable is borrowed from this tradition and applied at the level of a single word rather than a whole paradigm.

Stated contribution. This is not new linguistics and not a new theory of meaning. The claim is that a recognised process (bleaching), an incentive structure (the market reward for the broadest claim), and a concrete cost (degraded high-stakes reasoning) together explain a familiar irritation and point to a bounded remedy. If the remedy reduces to "use words carefully", then the paper has at least specified which words, where, and why the careless use does measurable harm.

4. Inflation as a Warning Signal — and Its Limits

When a term spreads rapidly across products and contexts, one reading is that the underlying thing is genuinely proliferating; another is that the term is being inflated for advantage. Both happen, and conflating them is an error.

This creates a problem. If rapid spread is sometimes genuine diffusion of a real capability and sometimes mere inflation, then "the word is everywhere" is uninformative unless we can tell the two apart.

A discriminator is therefore required. Rapid spread of a term is more likely to be bleaching-driven inflation than genuine diffusion to the degree that:

Application is decoupled from capability. The term is applied to systems that differ by orders of magnitude in what they do, with no narrowing qualifier.
There is a one-directional incentive to apply it. Using the term confers value (perceived sophistication, investment, premium pricing) at no cost, and no actor faces an incentive to police it.
The term resists falsification in use. No realistic standard must be met to apply it; the claim "this is AI" is rarely one anyone can contest.
It increases rather than reduces the listener's uncertainty about the referent. After hearing the term, you know less about what the thing actually is, not more.

A term spreading because a specific, capability-bearing technology is genuinely diffusing scores low on these. "AI" in current public use scores high on all four — which is what marks it as bleaching, not diffusion, and which is precisely the configuration in which the cost analysis of §5 becomes worth taking seriously.

5. The Minimum-Information Principle

Begin not from "what does 'AI' mean?" but from the minimum condition under which any term means anything: it must let the listener rule something out. "Spaniel" informs because most animals are not spaniels. A term applying to nearly everything in its domain has lost the capacity to distinguish, which is to say it has stopped describing.

"AI" has slid toward this condition. It is now routinely applied across a span including a hand-written conditional (if age > 65: flag()), a linear regression, a recommendation heuristic, a classical search algorithm, a trained classifier, and a frontier model — systems differing by many orders of magnitude in capability, cost, and risk. A word covering all of them with equal comfort no longer distinguishes among them.

The minimum-condition reading clarifies what has happened. "AI" has not become meaningless; it has become a category signal — a marker of membership in a valued class — rather than a description of capability. That residual function is real and is why the word persists. But a category signal cannot do the work of a description, and the harm arises precisely when it is asked to.

6. The Method: Term Collapse

The procedure for any high-stakes claim involving "AI" is the deliberate stripping of the umbrella term until the specific referent remains. Call it term collapse.

State the claim as written: "AI can/cannot/should/will do X."
Identify the referent actually intended. Which specific class of system — rule-based, statistical, a particular model architecture, an autonomous agent?
Substitute the specific term and re-read the claim. Does it still say what was meant?
Test the claim against an adjacent referent the umbrella also covers. Is it still true of a trivial system? Of a frontier one?
If the claim's truth-value flips between referents the same word covers, the word was carrying an illegitimate generalisation.
Critically: does the claim survive with the specific term substituted? If it only sounds true under the umbrella and fails once the referent is named, the word was doing the arguing.

Step 6 is load-bearing. Without it, "AI" debates generate confident claims whose plausibility depends entirely on the referent silently shifting. With it, the claim is forced to name what it is about — and many such claims do not survive the naming.

7. Case Study: Regulation, and the Limits of the Umbrella

The clearest cost appears in debates about regulating "AI".

7.1 The reframing

A proposal to "regulate AI" is nearly incoherent at the level of the word, because the referent silently swings between a spreadsheet macro and a frontier system with novel risks. Applying term collapse (§6) is genuinely useful: it forces the proposal to name what it governs, and most serious regulatory proposals, on inspection, do target specific capabilities or risk classes — they are not really about "AI" at all, but about, say, automated decisions affecting legal rights, or models above a capability threshold.

7.2 Why the reframing does not dissolve the difficulty

The central caution: naming the referent clarifies the debate but does not settle the regulation. Precision exposes that the disputants meant different things; it does not tell you what the right rule is. Two people who agree they are discussing "frontier models with autonomous tool use" can still disagree profoundly about how to govern them. The bleaching analysis removes a spurious disagreement (they were talking past each other) without touching the real one. The cost of the confusion relocates from "we cannot tell we disagree" to "we can now see exactly how much we disagree" — progress, but not resolution, and the genuine difficulty survives the clarification intact.

7.3 What honesty requires us to say

Term collapse earns a narrow but real conclusion in the regulatory case. It dissolves the pseudo-debates that exist only because the word's vagueness lets incompatible referents share a sentence, and it forces proposals to state their actual scope. It does not make hard regulatory questions easy; the substantive disagreements about frontier-model governance are real and survive every clarification. The honest one-line summary: precision about the word ends the arguments that were only ever about the word, and reveals the harder arguments that were hiding behind it.

8. Case Study: Safety Claims and Illegitimate Migration

A second cost: claims established for one referent migrate, unnoticed, to another under cover of the shared word.

A reassurance accurate about a narrow classifier ("the AI just flags transactions; it cannot act on its own") becomes a false reassurance when the word carries it to an autonomous agent. An alarm appropriate to a frontier system ("AI could pursue goals we did not intend") becomes scaremongering when applied to a regression. In both directions, the bleached word is what permits the migration: because the term spans both referents, a true claim about one slides into a false claim about the other without anyone appearing to change the subject.

The relevance is methodological, mirroring §7.2's caution. Term collapse exposes the migration — substitute the specific term and the slide becomes visible — but it does not adjudicate the underlying safety questions, which remain genuinely hard once correctly scoped. Naming the referent stops the illegitimate inheritance of truth-values; it does not hand you the answers for each referent. The difficulty is relocated to where it belongs, not removed.

9. Inflation, Incentive and Directionality

Most bleaching is slow. A word like "smart" or "natural" loses its descriptive edge over generations, as countless ordinary uses each wear it down a little. "AI" has hollowed in years. That difference in rate is the part worth explaining, and the explanation is not linguistic but economic.

The accelerant is a one-directional incentive (§4). Calling a product "AI-powered" confers value at no cost, because there is no penalty for overclaiming and no agreed standard the claim must meet. Every actor faces the incentive to broaden; none faces an incentive to narrow. Bleaching is therefore driven not only by ordinary generalisation but by a market that rewards it — which is why "AI" has hollowed in years what "smart" or "natural" took decades to do. This is offered as a disciplined account of the rate, not a completed economics of language; it is explicitly subject to the objection that genuine capability gains also drive some of the spread. The paper does not claim all the spread is inflation. It claims the discriminator of §4 can, in principle, separate the inflation from the diffusion — and that the public use scores as mostly inflation.

10. Guardrails Against Undisciplined Purism

The rejection of careless use is not a licence for pedantry. The argument has limits.

10.1 Casual use remains locally valid

Where stakes are low — chat, headlines, ordinary conversation — the vague word is harmless, because nothing turns on the referent. Insisting on "transformer-based language model" at a dinner party is not rigour; it is noise. The argument targets load-bearing claims, not all speech.

10.2 Imprecision must be located

When the word causes trouble, ask: is the trouble in the word (a genuine referent-shift, per §6) or in a substantive disagreement the word merely dressed? The §4 discriminator and §6 procedure are the tools; they are fallible.

10.3 Precision must eventually touch a real claim

Substituting a specific term is worthwhile only if it changes whether a claim is true or false or actionable. Precision for its own sake, where no claim turns on it, is the purist mirror of the inflater's vagueness. This is why §6 step 6 is the test: does the claim's status actually change under substitution?

10.4 Precision is not correctness

Naming the exact referent does not make a claim true. A perfectly scoped claim about "frontier models" can still be wrong. The method removes a source of confusion; it does not confer accuracy, and the two should not be confused.

10.5 The paper's own terms must be collapsed

"Semantic bleaching", "informativeness", "category signal" are this paper's primitives. If the argument replaces the vague "AI" only to lean on equally loose terms of its own, it has failed by its own standard. "Bleaching" is borrowed from an established literature and is reasonably precise; "category signal" is the weakest, closest to a placeholder, and earns its place only insofar as it can be made to predict which uses of "AI" cause harm — which §4 attempts and which remains to be tested at scale.

11. Beyond "AI"

The discipline transfers, with the same demand that something turn on it. Other valued terms are bleaching under the same incentive — "quantum", "sustainable", "neural", "agentic" — and the §4 discriminator applies unchanged: decoupled from capability, one-directional incentive, falsification-resistant, uncertainty-increasing. In policy and journalism, term collapse is a general tool for telling pseudo-debates (resolved by naming referents) from real ones (which survive). In any field with a hot label, the question is the same: is this word describing, or signalling category membership?

In each case the caution of §7.2 transfers: precision exposes the confusion but does not resolve the substance beneath it. The method finds words doing illegitimate work and forces them to name their referents. It does not make the named questions easy.

12. Discussion: Routing Around a Word Without Banning It

"Collapsing the term" does not mean abolishing "AI". The broad sense will not surrender ground; bleached words rarely un-bleach. The realistic goal is to route around the word where stakes are high, reserving it for the casual register where its vagueness is harmless. In practice one cannot speak only in precise architecture names; to communicate at all, one reintroduces umbrella terms. The value of the collapse is that it stops the umbrella from being mistaken for a description in the contexts — regulation, safety, honest specification — where the difference between trivial and consequential systems is the whole point.

The market will not hold this line, because the incentive that drives the bleaching is unchanged. Particular communities can, for their own purposes, the way scientific disciplines maintain technical vocabularies the wider culture erodes. The proposal is a norm for those communities, not a reform of the language.

13. Conclusion

The exhaustion with "AI" points at something real: the term has been bleached, by an ordinary linguistic process accelerated by a market that rewards the broadest possible use, to the point where it signals category membership rather than describing capability. The cost is not that it grates but that it degrades reasoning wherever the trivial and the consequential are forced to share a name — manufacturing pseudo-debates and licensing the illegitimate migration of truth-values.

What remains after these subtractions is smaller but more defensible than "the word is meaningless": a discriminator for telling inflation from diffusion, a procedure for forcing high-stakes claims to name their referents, a demonstration that doing so dissolves pseudo-disagreements while leaving the real ones intact, and an admission that the paper's own weakest term is itself a near-placeholder. The deepest question is not:

What does "AI" mean?

but:

When this word is used to reason about something that matters, is it describing the system — or merely marking it as the kind of thing we are excited about?

References

Bybee, J. (2003) 'Mechanisms of change in grammaticization: the role of frequency', in Joseph, B.D. and Janda, R.D. (eds.) The Handbook of Historical Linguistics. Oxford: Blackwell, pp. 602–623.

Hopper, P.J. and Traugott, E.C. (2003) Grammaticalization. 2nd edn. Cambridge: Cambridge University Press.

Kuhn, T.S. (1962) The Structure of Scientific Revolutions. Chicago: University of Chicago Press.

Sweetser, E. (1988) 'Grammaticalization and semantic bleaching', Proceedings of the Fourteenth Annual Meeting of the Berkeley Linguistics Society, pp. 389–405.

The Overclaim