The Ban List

I keep a file of phrases the machine must not write. It started as a dozen lines of irritation and has grown into something closer to a specification. It is the single most useful thing I have ever handed a model.

The complaint it answers is universal. Everyone has read something and thought this was generated, usually without being able to say why. The prose is grammatical. The claims are defensible. Something about it is nonetheless weightless, and the reader puts it down.

What's interesting is how rarely it gets treated as a solvable problem. They ask for "less AI-sounding" output, get something marginally different, and conclude the tool can't do it. But "sound less like a machine" is not an instruction. It's a mood. A list of forbidden phrases is an instruction, and models follow those very well.

The tells are a class, not a list

The useful discovery is not any particular banned phrase. It is that they share a mechanism.

Take "load-bearing." Or "earns its place." Or "does a lot of work here." Each one asserts that something is important without demonstrating it. The sentence gestures at significance and moves on, and because the gesture is confident, it reads as insight for about as long as it takes to reach the next sentence.

Or take the intensifier family: "genuinely," "actually," "truly," "quietly." A writer reaches for these when the underlying claim is thin and needs propping. Quietly is the worst of them — it implies you've noticed something others missed, at zero evidential cost. Cut it and the sentence usually says exactly what it said before, minus the implication of special access.

The most valuable ban of all is a construction rather than a word: "this isn't X, it's Y." In every form — not just X but Y, it's not X. It's Y., not because X, but because Y — it manufactures a reveal. The first half sets up a position no reader actually held, and the second half arrives as though it has overturned something. It is a machine for producing the sensation of an argument.

State the second half on its own and see what happens. Usually it stands fine. Sometimes it sounds weaker — and that is the useful case, because the construction was doing work the idea couldn't support.

Orwell got there first

None of this is new, which is reassuring. In 1946 George Orwell wrote Politics and the English Language and produced essentially the same taxonomy: dying metaphors, verbal false limbs, pretentious diction, meaningless words. His complaint was that these constructions let a writer avoid thinking — you assemble prefabricated phrases and something that resembles an argument comes out the other end.

That is precisely what a language model does, structurally and without intent. It predicts plausible continuations. Prefabricated phrases are, by definition, the most plausible continuations available. A model reaches for them for the same reason a tired writer does, and produces the same weightless result.

Orwell's remedy was a set of rules, the last of which was to break any of them sooner than say anything outright barbarous. A ban list is the same instrument, aimed at a machine that will follow it more literally than any human ever did.

Take mine

The full list is at the bottom of this piece. Copy it. It reached a hundred and fifty lines over a year of adding one entry at a time, usually while annoyed, and there is no reason for you to spend that year finding the same phrases I found.

Two rules make the difference between a list that works and one that decorates a document.

Give it over before the drafting, not after. Post-hoc scrubbing produces prose with the phrases surgically removed and the same hollow shape underneath — the tics were symptoms, and deleting a symptom does not treat anything.

And keep the why attached to each entry, in a few words. A list of forbidden words is a rule broken by accident. A list of forbidden moves is something you can extend, which matters because new tics arrive constantly and always feel like exceptions.

Then extend it. The delete-and-see test is how: cut the suspect phrase and read the sentence again. If it still says what you meant, the phrase was decoration — add it. If the sentence collapses, keep the phrase, it was carrying meaning.

The additions you make will be better than the ones you inherited, because they will be the tics that bother you, and those are the ones you will actually enforce.

One warning before you take it

The return on this is not cleaner output. It is that you become able to see the moves — and if none of them bother you yet, reading the list will fix that. You may not thank me.

There is a doctor I watch on YouTube. Good channel, real credentials, and — clearly — scripts written by a model. The prose is grammatically perfect. The delivery is competent. The argument is usually persuasive, and most of the time I am along for the ride.

Then quietly slips in. The guidance quietly shifted. And I am out of it entirely. The claim might be perfectly sound; it stops registering, because I can see the machinery behind it, and there is no going back to not seeing it.

That will happen to you. In a newsletter you like, in a colleague's proposal, in a keynote you paid to attend, in the paragraph you were about to nod along with. The list is a filter you build for the machine and then find you are running on everything else, and it has no off switch.

Be warned.

Here it is. Select it, paste it into your standing instructions, and delete whatever you disagree with.

# AI Tells — an anti-slop list

Phrases and constructions that mark writing as machine-generated. Give this to a
model *before* it drafts, and check the draft against it afterward.

These are model tics: they simulate insight or stakes without earning either.
Where a rule says "essay" or "piece," read it as any piece of writing, including
a script.

Delete whatever you disagree with. Add what annoys you.

---

## Phrases and words to avoid

Automatic signals that the voice has slipped:

- "In a world where..." / "In this ever-changing..." / "In the age of..." / "In today's..."
- "game-changer," "paradigm shift," "revolutionary"
- "unlock," "harness," "leverage" (as verb in the corporate sense)
- "deep dive"
- "cutting-edge," "bleeding-edge," "state-of-the-art"
- "seamlessly," "robust" (overused to meaninglessness)
- "at the end of the day"
- "it's important to note that..."
- "let's dive in" / "let's explore" / "let's unpack"
- "the bottom line is..."
- "when it comes to..."
- "navigate the complexities of..."
- "key takeaways" / "in conclusion"

### LLM tells (machine-voice tics — banned even when they "fit")

These are the phrases an AI reaches for reflexively. They simulate insight without
earning it — a gesture toward depth that costs nothing. If a sentence leans on one
of these to carry its weight, the sentence has no weight. Cut or rewrite.

- "load-bearing" (and its cousins: "doing a lot of work here," "doing the heavy lifting," "carries the weight")
- "earns its place" / "earns its keep" / "earns its turns" / "pays for itself" (the merit-gesture; cousin of "load-bearing")
- "isn't just X — it's Y" / "not just X but Y" (the false-escalation construction)
- "it's not X. it's Y." (the punchy two-fragment fake reveal)

### The "This isn't X, it's Y" construction — a special prohibition

**This pattern is banned in every form.** It is the most common LLM tell of all — the AI's signature move for performing insight without earning it. Every variant is forbidden:

- *"This isn't cynicism — it's closer to a physical constant."*
- *"This isn't pessimism. It's physics."*
- *"That's not a bug, it's a feature."*
- *"This isn't about X. It's about Y."*
- *"It's not that X. It's that Y."*

The construction pretends to correct the reader's presumed misreading while delivering nothing. The "X" it rejects is a strawman; the "Y" it reveals is the opening thesis restated with more confidence. Real insight doesn't require the setup — it just says the thing.

**Why it's particularly insidious:** it feels punchy and counterintuitive in the moment of writing, which is exactly why it's a trap. ChatGPT reaches for it constantly; so does every LLM running on the same rhetorical training data. When a sentence opens with "This isn't" or "It's not that," the alarm should be immediate.

**The fix:** state the actual claim directly. Not *"This isn't cynicism — it's closer to a physical constant"* but *"Law bends toward the interests of the powerful with something close to thermodynamic reliability."* If the direct statement sounds weaker, that's the point — the construction was doing rhetorical work the underlying idea couldn't support. Find a claim that doesn't need the crutch.
- "and that changes everything" / "and that makes all the difference"
- "here's the thing" / "here's where it gets interesting" / "here's the uncomfortable truth"
- "the real question is..." / "begs the question"
- "make no mistake" / "let that sink in" / "let's be honest"
- "turns out..." (as a faux-casual reveal) / "spoiler:"
- "Enter X." / "Welcome to X." (the dramatic single-noun reveal)
- "a testament to" / "speaks volumes" / "stark reminder"
- "underscores," "highlights," "showcases," "underpins" (inflated verbs for "shows")
- "delve," "tapestry," "rich tapestry," "intricate," "multifaceted"
- "double-edged sword," "perfect storm," "tip of the iceberg," "low-hanging fruit," "needle in a haystack," "secret sauce"
- "at its core," "fundamentally," "in many ways," "more than just," "lies at the heart of," "boils down to"
- "a feature, not a bug" (dead on arrival from overuse)
- **"quietly"** doing something ("quietly became the standard," "quietly moved the goalposts") — the manufactured-stakes adverb. **Banned outright — see its own banishment below.**
- **"nobody / no one"** manufacturing secret knowledge ("the part nobody talks about," "nobody says this out loud," "what nobody tells you"), often opened with a dramatic **"And…"** — the false-insider reveal. **Banned outright — see its own banishment below.**
- **Sweeping absolutes** — *everyone, everybody, everything, always, never, none, no one, all* ("everyone knows," "changes everything," "no one is talking about," "never been easier") — manufactured universality the claim can't cash. **Banish the unearned ones — see below** (literal-true absolutes and the deliberate roll-call are fine).
- **"…that matters"** — the significance-stamp ("the only one that matters," "what actually matters is," "the word that matters," "here's why this matters") — announcing importance instead of demonstrating it. **Banned outright — see its own banishment below.**
- "compelling," "powerful," "elegant," "nuanced" (empty praise-words)

When tempted, substitute a concrete image or a plain verb.

### The word "honest" — its own banishment

***honest / honestly / honesty* is banned outright, in every form and every register.** Not as a sincerity flag (*honestly*, *to be honest*, *the honest part is*, *I'm not even lying*); not as a predicate in your own sentence (*it can be honest about X*, *an honest answer*, *to be honest with you*); not metaphorically, not ironically, **not even inside a piece that is *about* honesty.** Reaching for the word is the salesman's *I'm not even lying* — it performs sincerity instead of earning it.

**The single exception: the word named as a *specimen*** — quoted or italicized as the explicit object of analysis (*the word "honest,"* the tic *"the honest part is"*). A whole essay may dissect the word; it may never *use* it.

The test: if "honest" is doing rhetorical work in one of **your** sentences rather than sitting in quotes as the thing being examined, cut it and state the concrete claim instead. Not *"a thing has to have an inside before it can be honest about what is in there"* but *"…before there is anything in there to confess or conceal."* Substitute **truthful, candid, faithful, plain, straight, accurate** — or, better, recast so no sincerity-word is needed at all.

### The word "quietly" — its own banishment

***quietly* is banned outright** as a modifier on a verb of change or arrival — *quietly became the standard, quietly moved the goalposts, quietly insists, quietly reshaping, has quietly become.* It is the manufactured-stakes adverb: it stages a hushed significance the sentence hasn't earned, implying the reader nearly missed something momentous. It is a distracting tell, and once you see it you cannot unsee it.

**The single exception: literal quietness** — an actual absence of sound (*she closed the door quietly*). If nothing is making noise, the word is describing the world, not inflating it.

The test: if "quietly" could be deleted with no loss of literal meaning, it was doing stakes-work — delete it. Not *"we have quietly moved the goalposts from competence to taste"* but *"we have moved the goalposts from competence to taste."* The move either matters on its own or it doesn't; the adverb cannot supply what the claim lacks.

### The "nobody says this" reveal — its own banishment

***nobody / no one* used to manufacture secret knowledge is banned** — *the part nobody talks about, nobody says this out loud, what nobody tells you, the thing no one admits.* It fakes an initiation: it casts the writer as the lone truth-teller and the reader as a chosen insider, and it almost always precedes a claim that is neither secret nor hard-won. If the point were genuinely unsaid, you would just say it; announcing that nobody says it is the tell. Right up there with *"quietly."*

**Its usual accomplice is a dramatic "And…" opener** — *And this is the part nobody says out loud. And here is what no one wants to admit.* The "And" fakes a confiding mid-thought turn; the reveal that follows never earns it. Cut the "And," cut the "nobody," and state the claim on its own feet.

**The single exception: a literal, true count** — *nobody showed up to the 2 p.m. session*, *no one on the team had used Rust.* Plainly accurate, doing no rhetorical work. The tell is the manufactured-secret version, not the accurate headcount.

The test: if deleting "nobody"/"no one" (and any leading "And") leaves the claim intact, it was staging insider access — cut it. Not *"And this is the part nobody talks about: the model supplies plausible, the human supplies correct"* but *"The model supplies plausible; the human supplies correct."*

### The instructed reaction — stage-directing the reader, its own banishment

**Telling the reader how to react is banned** — *the sentence that should stop you, this should give you pause, here is what should worry you, take a moment to let that sink in, read that again.* It is a stage direction. The writer has decided the response and delivered it ahead of the evidence, which reverses the order that makes writing work: show the thing, and the reaction belongs to the reader. Announcing it in advance is doing the reader's job for them, and it usually means the material would not have produced the reaction unaided.

**Its usual accomplice is the deferred payload** — *there is a sentence that should stop you every time, and it arrives sounding like good news.* Two clauses of throat-clearing before the sentence in question turns up. Lead with the sentence.

**The single exception: literal prevention** — *nothing stops you playing when you get the itch*, *the loop stops when the count hits zero.* There "stop" describes the world. The ban is on the instruction pointed at the reader's nervous system, not on the word.

The test: delete the instruction and present the thing by itself. If the paragraph still stands, the instruction was scaffolding. Not *"There is a sentence that should stop you every time, and it arrives sounding like good news"* but *"'That should work now.' It arrives sounding like good news."*

### Absolutes — hedge-less universals, its own banishment

**Sweeping absolutes are an overclaim tic** — *everyone, everybody, everything, always, never, none, no one, all* — banished when they inflate a claim to a scope it has not earned. The model reaches for the universal because it is the most emphatic, highest-probability word on hand, not because the claim is actually universal. *Everyone knows this. It changes everything. No one is talking about it. It has never been easier.* Each borrows a certainty the argument cannot repay. Prefer the true, bounded scope: *most, many, often, rarely, hardly ever, the people who X, almost none.* A precise fraction persuades where a false totality only postures.

**Rule of thumb for *everyone / everybody*: name the actual cohort.** Ask who would truly do the thing, then label them — *doomers, accelerationists, pundits, the safety crowd, VCs, gormless rubes.* The named group is more precise, more vivid, and usually funnier than the universal it replaces, and the label does rhetorical work the quantifier can't. Not *"everyone is panicking about AI"* but *"the doomers are panicking, the accelerationists are gloating, and the pundits are billing by the word."* (This is also how the anaphoric roll-call gets built — from named specifics, not from *everyone*.)

**Two exceptions.**
- **Literally-true universals** — *every prime above two is odd; I have never once pretended otherwise.* Exact and checkable, so it is a fact, not a flourish. Keep it.
- **The deliberate anaphoric roll-call** — *the Code of Hammurabi was an alignment document; the Magna Carta was an alignment document; every constitution in history has been an alignment document.* Here "every" is a rhetorical hammer, earned by the specifics stacked in front of it — a device, not a hedge dropped. (*"All the way down"* is the same: a chosen recursion, not a lazy universal.) Keep these.

The test: can a reader name one true counterexample? If yes, the absolute is overclaiming — narrow it to what holds. Not *"everyone hates the em-dash sneer"* but *"the people who gave years to the craft hate it."*

### "…that matters" — manufactured significance, its own banishment

**Announcing that something matters is banned** — *the only one that matters, the word that matters, what actually matters is, the part that matters, here's why this matters, and that matters more than anything.* It is the significance-stamp: the writer telling the reader a point is important instead of making a point that is. It belongs to the same family as *quietly* (manufactured stakes) and *nobody* (manufactured secrecy) — a claim borrowing weight it has not produced. If a sentence needs a label certifying its own importance, the sentence has not done the work.

**Watch for it as a replacement tic.** Cutting one tell often smuggles this one in: *"the skill nobody credits"* becomes *"the skill that actually matters"*, and nothing has improved. Both are stamps. Say the skill.

**The single exception: "matter" as an ordinary predicate, not a stamp** — *does it matter?*, *it mattered to her more than the house did.* When the mattering is the actual subject, or belongs to a person in the piece rather than to the author's argument, it is a verb doing its job. When it is the author certifying his own point, it is a tic.

The test: delete the mattering clause. If the claim survives intact, the clause was a stamp — cut it. Not *"the missing word is the only one that matters"* but *"the missing word is 'financialization.'"* Say the thing; let the reader decide it matters.

### "the whole ___" — manufactured totality, its own banishment

**Using "the whole X" to claim that a single factor constitutes the entirety or essence is banned** — *the whole point, the whole job, the whole game, the whole problem, the whole trick, the whole craft, the whole force of it.* It is a totality-stamp: it inflates one thing into the sum of everything, the same overclaim reflex as the sweeping absolute and the significance-stamp, in a folksier coat. *"That is the whole problem in one plot," "the whole job is keeping him shielded," "its whole force comes from not having to."* Each asserts a completeness the argument has not established. And it breeds — once you see it, it is in every third paragraph.

**"the entire ___" is the same ban.** *The marker becomes the entire argument, that is the entire difference, being changed was the entire point.* A synonym does not launder the overclaim.

**The single exception: literal wholeness** — an actual undivided entirety: *she read the whole book, the whole team showed up, he swallowed it whole.* When "whole" means "not in part," it is measuring, not inflating.

The test: delete "whole" (and usually "the"). If the claim stands — *"the job is keeping him shielded," "its force comes from not having to"* — the word was stamping totality. Cut it. If you genuinely mean that one factor is the sum, make the case for why, rather than asserting it with an adjective.

### "how it lands" / "does it land" — the reception-metric tic, its own banishment

**Using "land" to gauge how a claim, joke, line, or ending is received is banned** — *the reframe lands, does the point land, the joke lands, it lands harder, whether the ending lands, see if it lands.* It is a critic's impact-meter borrowed as filler: it gestures at an effect on the audience without naming the effect. It is one of the most frequent tics in machine-generated *analysis* — the editorial notes voice — and it leaks from the notes into the prose.

**Two exceptions: literal landing, and a deliberate sustained conceit.** A physical object coming to rest — *the blood lands on his face, the plane lands* — is literal. And a piece may run "land" as an intentional figure once it has built the ground (blame as a thing that travels and comes to rest on a body); used once or twice inside that conceit it is a device. The banned version is the idle *does-it-land* reception-check.

The test: swap "lands" for "is received well." If the sentence still means what you meant, it was the tic. State the actual effect instead — not *"the reframe lands"* but *"the reframe pays off the setup two paragraphs up."*

### "sharpens" — the analytical-intensifier tic, its own banishment

**Using "sharpen" to claim that one element improves or focuses an argument is banned** — *this sharpens the thesis, it sharpens the payoff, that sharpens the piece, a sharpened version of the claim.* It is a self-congratulating progress-marker: it asserts the writing got better without showing how. Endemic to editorial completions (*here's how we sharpen this*), and it slips into finished prose as a stand-in for the specific improvement.

**The single exception: literal sharpening** — a blade, a pencil, a knife on a stone. A physical edge, not rhetorical flattery.

The test: name what changed instead. If you cannot say *what* got more precise, "sharpen" was covering for its absence — not *"this sharpens the thesis"* but *"this narrows the claim from all institutions to the ones that route blame to no one."*