Steph Zinn, writing for a16z Crypto, argues that conversations about AI writing often rest on two assumptions that do not hold: that machine-generated text can be identified through a stable set of tells, and that any text carrying those traits is automatically poor.
Her focus is not on catching AI. It is on helping founders and writers decide which habits get in the way of clear expression, and which ones should be cut or kept depending on the job the writing needs to do.
Rhetoric: prose that sounds smooth but says little
Zinn begins with rhetoric. One of the most frustrating, and oddly satisfying, parts of editing, she writes, is finding a sentence that reads well enough yet still feels off. After moving a few words around, the editor realizes the line had no meaning to begin with, and the cleanest fix is to delete it.
She uses a recent Fable self-assessment on punctuation style as an example: "Human punctuation has a body; it fidgets. Mine is uniformly deliberate, and the tidiness of that deliberateness is itself rhythm." Many readers would call that the kind of sentence that could not have come from a human. Zinn disagrees with the premise. As an editor, ghostwriter, and lifelong fantasy reader, she says humans have long written lines like that without any help from AI.
Some cases are less obvious. Unlike more visible tells, such as three list-shaped sentences in a row or a familiar burst of em dashes, these lines can look like perfectly normal prose. Some even sound smoother than the surrounding text: "What really matters is happening." "The stakes could not be higher." "Its implications are profound."
Her point is that AI produces these phrases at scale because they function as plausible but semantically thin connective material. Before they were tied to AI, this kind of bland writing already had a name: corporate prose. Phrases such as "we are excited to announce," "we are at an inflection point," or "this is the next chapter in our journey" sound weighty, but often fail to point to anything specific that can be restated plainly.
Zinn pushes back on the idea that this always reflects laziness. More often, she says, it is just draft language: the first stock phrase that floats to the top of the mind. That is acceptable as a starting point. It becomes a problem only when the writer stops there.
Her editing method does not depend on whether AI was used. If a sentence feels wrong, uses a familiar empty pattern, cannot be understood immediately, or simply has the wrong air about it, she advises rewriting it. If another version is clearer, keep that one. If the sentence boils down to little more than "something exists" or "things are changing," delete it and see whether the piece still stands.
Rhetorical tells she flags
- La Croix-style insight: lines such as "What really matters is happening" or "The stakes could not be higher" that sound important but work mainly as filler.
- Empty contrast: constructions like "This is not just about X — it is about Y," where Y is even vaguer than X. She suggests removing the first half and seeing whether the point arrives faster.
- Hedges: phrases like "in many ways," "to some extent," or "it could be said." Their function is often to make a sentence harder to falsify. Zinn notes that in areas such as finance and medicine, some hedging is necessary for compliance and should stay.
- Over-parallelism: bullet points and mirrored sentence structures that are too even in grammar and length, a pattern she says is both a common AI tell and a fast route to dull prose.
- Summary phrases: lines such as "when all is said and done" that can often be removed without changing the paragraph at all.
Her judgment here is clear. Meaningless, vague language usually needs to go, and almost never deserves protection. She says her own revision prompts tend to circle around two goals: make the writing more specific, and make it plainer. The target is higher information density with better readability. If an LLM helps with that, she is happy to use it.
She also suggests building a rough style guide with positive and negative examples, using AI to run a paraphrase test that converts a passage into its boring version, and asking an LLM to rewrite prose at a sixth-grade reading level as a blunt way to strip out filler.
Voice: the "Alexa voice" problem
At the level of voice, Zinn describes a default AI register as "Alexa voice." English, she notes, contains roughly 1 million words, yet stock AI prose can sound as if it relies on only about 400 of them. Instead of obsessing over whether a trendy word is an AI marker, she proposes a different test: fungibility. Can a sentence be lifted word for word out of your piece and dropped into somebody else’s article on an entirely different topic without anyone noticing?
She points to a period when readers could track low-friction vocabulary by feel. Researchers, she says, famously noticed a surge in the word "delve" in academic abstracts after ChatGPT launched in 2022. That was followed by waves of banned-word lists that included tapestry, testament, underscore, and more recently load-bearing, scaffolding, and broader.
Zinn argues that careful diction is not a luxury reserved for literary writers. It matters to anyone building or running a company because marketing increasingly depends on personality. She says readers can easily picture the writing styles of Brian Armstrong, Vitalik Buterin, and Chris Dixon: concise, earnest, and non-ironic in Armstrong’s case; technical, digressive, and dense in Buterin’s; restrained and philosophical in Dixon’s.
Her warning is straightforward: when using an LLM, do not hand over your personality along with the draft. A writer’s own word choices, including their flaws, add a patina that machines cannot replicate.
She acknowledges that practical advice on finding the right word is limited. Much of the work, in her view, is about emotional or atmospheric precision rather than technical precision. She notes that E.B. White himself admitted that no one can fully explain why some words ignite a reader and others do not.
Her recommendation is to start by identifying what you like in writing you admire. Learn a new word, or place an old one in a new setting. Match your diction to the mood you want to create. If you are explaining something obscure, short and simple words may serve best. If you want readers to sense trouble, she says, choose "scheme" over "plan."
Voice tells she lists
- Generic warmth: lines like "Great question!" that sound like customer support friendliness.
- Reusable sentence frames: phrases such as "one useful way to think about this is," "the key takeaway is," or "this can be understood as," which can be transplanted almost anywhere.
- Low-friction words: diction that feels too perfectly safe and frictionless.
- Abstract nouns: heavy reliance on words like efficiency, complexity, society, communication, and innovation.
- Bland motion verbs: verbs such as navigate, leverage, unlock, cultivate, shape, enhance, or streamline that imply movement without much force.
- Beacon language: formulas like "is a testament to," "is a beacon of," or "reminds us that."
- Fuzzy nouns: words like landscape, space, journey, ecosystem, or tapestry that gesture near the thing instead of naming it directly.
- Vague intensifiers: phrases such as "highly important," "significant impact," or "critical role" when unsupported by specifics.
Most of the time, she says, these should be edited. But not always. If the entire point is to sound like an NPC, the default voice can be useful. Documentation, error messages, terms of service, safety notices, and formal apologies to large groups of strangers often work better when personality is reduced rather than amplified.
In practice, she prefers using LLMs for detection rather than for drafting. Models are good at spotting jargon, business-speak, and other common markers. Writers can ask them to identify the most generic words and phrases in a draft, then revise from there. She also recommends a transplant test for individual lines: if a stranger could plausibly claim the sentence as their own, it may need a rewrite.
Her example is simple. "The Ethereum ecosystem is expanding" can be made more concrete by saying developers are building more wallets, exchanges, and lending markets around Ethereum. She also notes that more people are feeding voice notes into LLMs to get ideas onto the page. One advantage is that spoken quirks can reveal what makes a voice distinct. But the output still needs editing. There is no hole-in-one.
Structure: plenty of form, not always much function
Zinn says AI writing often becomes overstructured when left on its own: too many H2s, too many lists, and paragraphs broken into fragments. That does not make structure bad. Human writers have always borrowed from a stock of familiar forms because those forms are readable and dependable.
People break ideas into three points because three feels complete. They add signposts such as "first" or "in other words" to orient the reader quickly. They build neat categories because those are easy to scan. Many writers also learned some version of hamburger logic in school: say what you will say, say it in supported sections, then say it again at the end.
That model has value because it forces argument, evidence, and readable sequencing. The problem begins when the structure starts forcing the idea into a container that does not fit. For Zinn, structure is a set of decisions: what container to use, what matters most, which ideas belong together, and how they should be named. Whether those decisions are good depends on the job the piece is trying to do.
Narrative nonfiction needs discovery and tension to pull a reader forward. A product announcement needs efficiency because readers are already scrolling. Explanatory writing needs a sequence that builds one concept on top of another. Her advice is to ask, ideally at the start, what format best suits the idea, then borrow from writing that has already solved a similar problem.
If you are writing an argument, she says, study how other writers organize commentary. If you are writing technical explanation, pick one of the best explainers you have read and examine how the information is arranged. Once you have that template, adapt it to your own piece.
Structural tells she calls out
- Over-organization: excessive subheads, bullets, numbered sections, or mini-frameworks.
- Always arriving at three points: the triad is not the issue by itself; the problem comes when the idea is distorted to fit it, or when a weak third point is added just to complete the shape.
- Familiar article silhouette: broad intro, explanation, example, reminder, conclusion.
- Cliché openings: lines like "In today’s fast-changing world..."
- Too much signposting: first, next, finally, in summary, here is the breakdown, let us break this down.
- Rigid transitions: formulas such as "To understand why this matters, we first need to look at..."
- Section previews: lines like "There are three key reasons..."
- Bullet lists: she notes that AI detectors often flag them unfairly, though lists are useful. A more recognizable move is to introduce the list with a bolded lead-in.
- Short dramatic fragments: common in performative LinkedIn posts. Short. Punchy. Often in threes.
- Restatement endings: conclusions that merely say the same thing again.
- Didactic endings: vague closing lines about progress, the future, or what we can learn.
She says subheads that do not suit the format, especially in commentary, personal narrative, and pieces driven by voice and rhythm, usually deserve to be cut. Sections that are too weak or too similar should be reworked. And if a structure changes the meaning of the piece, it is the wrong structure even if the facts remain the same. She is especially skeptical of obvious scaffolding like "three reasons" signposts that add little beyond bulk.
That said, she makes room for structure when the format calls for it. Commentary often benefits from a recognizable argumentative path. Lists, explainers, how-to guides, and other genres where readers expect headings and subheads also benefit from explicit organization. If an idea naturally comes in three parts, there is no problem with using three parts. She also notes that well-structured information is rewarded in LLM and search contexts, and that reference material such as documentation, guides, and FAQs is often meant to be scanned rather than read straight through.
When using a model for structure, she suggests telling it what the structure needs to do for the reader, whether that means building suspense or making the piece easier to scan. Another tactic is to give it a published article in the same genre that you like and ask it to study how the argument works before restructuring your own. Or ask for a reverse outline that describes what each paragraph is doing, then use that rough skeleton to draft.
Punctuation: em dashes do not need to be sacrificed
Zinn closes the practical section with punctuation and the current panic around em dashes. They have become one of the classic AI tells, she writes, but that does not mean writers should abandon them. Strunk and White generally advised restraint, reserving dashes for cases where more common punctuation was not enough. Yet the em dash has a relaxed, conversational feel that other marks do not quite reproduce.
She acknowledges the complaint against dashes. Inserting a whole side thought into the middle of a sentence can disrupt momentum. But once the dash becomes shorthand for bad AI prose, writers risk avoiding it for the wrong reason.
Her position is that punctuation is governed by specific rules and style preferences anyway. Whether AI favors the dash is secondary because some uses are simply correct. It remains well suited to setting off longer insertions or marking abrupt shifts in thought and feeling. She notes that horror writer R.L. Stine has praised it for that reason.
Zinn says plainly that she is not giving up the em dash because of current AI anxieties. Fashions change. Writers should not discard punctuation that fits their prose simply to avoid looking assisted. She jokes that while people race to avoid the em dash, AI has moved toward the colon, raising the absurd question of whether the interrobang will become the next proof of humanity.
For her, the short rule is this: choose the mark that is most correct and least distracting. Detailed usage questions, including the Oxford comma debate, can vary with style and taste. Perfection is less important than making sure punctuation supports meaning and does not irritate the reader.
She also warns against uniformity. If every sentence looks and sounds the same, the result feels mechanical. LLMs often produce repeated colon-led list constructions that grow tiring. The same is true of multiple sentences repeatedly interrupted by dashes. Her recommended test is simple: read the piece aloud and treat punctuation like stage direction. If it sounds unnatural, the reader will likely feel it too.
Punctuation tells she highlights
- Colon-heavy constructions: formulas such as "The problem is:" or "The key point is:" followed by a grocery list.
- Clustered em dashes: not the dash itself, but the density, especially when mixed with lists and asides. She says there is no hard numerical rule, but do not go past two in a sentence.
- Unserious parentheses: bracketed comments that import self-aware or jokey tone into prose that does not support it.
- Performative semicolons: she cites Kurt Vonnegut’s famous swipe at semicolons while still acknowledging that they do have legitimate uses, just not very often.
Her advice is not to overthink whether punctuation makes you look AI-generated. Edit when marks are repetitive, distracting, or fail the read-aloud test. If a model must help, do not simply ask it to remove every dash. It will often swap in another symbol while preserving the same underlying sentence pattern. A better prompt is to default to periods and commas, using special punctuation only when grammar requires it. She again recommends feeding the model your own writing, or writing you admire, to help it estimate a more personal punctuation fingerprint.
The better question is whether the writing does its job
Zinn ends by arguing that as more people use these tools, asking whether a piece was machine-generated is becoming less meaningful. The answer, she writes, will almost always be "to some extent."
She makes room for cases where bad AI writing still needs to be exposed, whether by human judgment or by Proof of Person technology: when disclosure is required, when the byline itself is the whole point, or when one person pretends to be a thousand.
For nearly everything else, she returns to a simpler test. Does the writing do the work it is supposed to do?
In her view, AI helps people express and publish ideas that otherwise might never be written down. It can save time on research and organization. It may even help people become better writers. Reducing expanded expressive capacity to something merely embarrassing or low-quality is, she says, ungenerous. If a piece works as intended, how much does it really matter which step in the process gave away the presence of a machine?
Her final question is aimed at the fear itself: why grant LLMs so much reverence that writers would give up an entire piece of punctuation, one that has existed since the dawn of printing, simply to avoid appearing assisted. She calls that a ridiculous concession, especially when these tools are already used so often and will continue to be used that way.
The article closes with thanks to the a16z Crypto editorial team — Tim Sullivan, Robert Hackett, and Sonal Chokshi — for feedback, as well as to the many years of editing and discussion that shaped the piece.