Arguments about AI writing tend to rely on two assumptions: (1) that machine-generated prose can be identified by a conspicuous set of tells, and (2) that anything containing those tells is necessarily bad and embarrassing. Respectfully, these are terrible ways to think about writing.
We writers can pontificate about what AI will do to the craft all we want. The reality is that most writing isn’t some artful monomyth. It just needs to work. Does it explain things clearly? Can readers trust it? Is someone — anyone — behind the wheel, making decisions here? In short, is the writing doing its job? The extent to which a human was involved will matter in some cases and not others. The nice thing about these questions is that they apply equally well to writing produced by a person, a model, or some increasingly entangled combination of the two. They’re also useful in industries like ours, where scams (often written by AI) and jargon (too often written by humans) can easily distract from real goals and values.
What is actually useful about the AI detection discourse is that it’s made some longstanding writing problems much easier to see. So instead of using some percentage of “Human Written” as the relevant heuristic, we take a look at the hallmarks of machine-generated prose — its rhetoric, voice, structure, punctuation — as a taxonomy of problems that have always existed. Our goal as editors is to help founders, and anyone else in the business of communicating ideas, figure out when these habits get in the way of effective writing, regardless of provenance.
Rhetorical hallmarks: Insight-shaped writing
One of the most infuriating (then rewarding) experiences in editing is finding a sentence that sounds good but just kind of feels off. You rearrange it, move a phrase here and there, and slowly realize it never meant anything to begin with. Then you delete that slippery little pretender and move on with your life.
A quick example straight out of asking Fable for its most recent punctuation meta:
“Human punctuation has a body; it fidgets. My punctuation is uniformly deliberate, and uniform deliberateness is itself the rhythm.”
I would describe this tone of voice as “person at an otherwise nice party talking about jazz improvisation,” but I’m not going to pick these sentences apart here — you get the point. People say things about this sort of writing, like “it could just never have been written by a human.” And I take their point, but as an editor, ghostwriter, and lifelong reader of fantasy novels, I can assure you that people do indeed write these sorts of sentences without AI all the time.
Spotting this kind of language, however, isn’t always as straightforward as in the fidgeting-punctuation example. Unlike other AI hallmarks (like three consecutive sentences full of lists or the familiar Morse code of too many em dashes), these sentences don’t actually look like a problem. They just kind of look like sentences. Some of them even sound better than the other sentences around them: “Something real is happening.” “This matters for crucial reasons.” “The implications are significant.”
AI writing produces these sorts of phrases in bulk, because they’re plausible (if semantically vacant) connective tissue. Before it became associated with AI writing, we used to call this category of bland, meaningless language corporate, because it’s your typical, dispassionate business-blog speak. Can you think of a more spiritless way to say you’re excited about something than to start with, “We’re excited to announce…”? “We’re at an inflection point” or “this is the next chapter of our journey” are two more examples. They’re phrases that sound weighty but generally lack the specificity to survive paraphrase.
I’ve heard other editors miscategorize this as lazy or thoughtless writing. That’s uncharitable. It can be a matter of taste, but mostly, it’s not easy to swat away the generic, instantaneously retrievable phrases and idioms buzzing at the front of our minds and find something better. It is first-draft language. And honestly, it’s a perfectly fine place to start. We just don’t want to end there.
The approach I recommend for avoiding meaningless language is the same, whether you’ve used AI or not. If you spot a sentence that feels off — if it uses any of the phrases or phrasings below, if you don’t immediately understand what it means, or if it’s just got a bad vibe to it — start rearranging and try to restate it.
If there’s a paraphrase, keep the better version. If it boils down to “stuff exists” or “things are changing,” skip it, and see if your piece can live without it. If it can, it’s time to embrace the illicit thrill of deleting words from your page.
Some tells and when to edit them
Naturally essenced profundity (the La Croix of insight). “Something real is happening.” “The stakes couldn’t be higher.” Sounds important; is actually filler.
Empty contrasts. “It’s not just about X — it’s about Y,” where Y is fuzzier than X. This format works sometimes when the contrast is very clear. But before leaving it as is, check whether you can delete that first “it’s not about” phrase. You may get to your point faster.
Hedges. “In many ways.” “At some level.” “Arguably.” Each of these exists to make the sentence impossible to be wrong about. Note that in some industries, like finance or healthcare, some hedges are necessary to soften claims and comply with the law. I think AI would call these “load-bearing” hedges.
Too much parallelism. Bullets and sentence structures that mirror each other too tidily. Every list item is the same grammatical shape and the same length. These are AI tells, but they also make for terrifically boring writing.
In-summation phrases. “At the end of the day.” “When the dust settles.” It’s often possible to delete these phrases with no impact on your paragraph.
Edit: Almost always. Meaningless, gobbledygook language is at the top of our stack-ranked list of writing pitfalls. It will never move your ideas forward, no matter how nice it sounds. It is a scourge. Eliminate accordingly.
Embrace: Almost never.
Tips: I often find myself circling two kinds of prompts: (1) those that make the writing more specific, and (2) those that make it more plainspoken.
My goal is always to drop unnecessary technical jargon, and work toward something that’s both information-dense and easier to read. If an LLM can help me do that, great.
More ideas for getting rid of filler language:
Creating a rough style guide that defines what “good” writing looks like to you with examples and counterexamples. This will hold up better than a laundry list of AI tells.
Using AI to run the paraphrase test I alluded to above. Just ask for the “boring” version of what you’ve written.
Prompting the LLM to “write it at a 6th grade level” works decently as a blunt instrument to take to your prose.
Voice hallmarks: Alexa voice
There are roughly one million words in the English language. Default AI writing sounds like it’s working with about four hundred of them. But rather than trying to pinpoint which word of the day is an AI hallmark, the test here should be fungibility: could this sentence be taken word-for-word out of your writing and dropped into someone else’s essay on a different topic without anyone noticing?
For a while you could track this category of low-friction vocabulary empirically. Researchers famously noticed the word “delve” spiking in academic abstracts after ChatGPT launched in 2022, and an onslaught of banned-word lists followed: tapestry, testament, underscore; more recently, load-bearing, scaffolding, broader, etc.
In the meantime, thoughtful diction isn’t just a preoccupation of “literary” writers. It matters for anyone who starts or runs a company because people increasingly market on the strength of their personalities.
The best founder writing is individually expressive. Whether or not others would deem it “good,” it gives readers a sense of the person who believes it. It’s why we can so easily conjure what it means to write like Brian Armstrong (pithy, earnest, unironic); like Vitalik Buterin (technical, digressive, dense); or like Chris Dixon (austere, philosophical).
So, when using an LLM, please make sure you don’t cede your personality to it. Your own word choices and their imperfections add patina a machine could never.
For all of the writing available on finding the right words (two of my favorites are Theodore Bernstein’s The Careful Writer and George Saunders’ A Swim in a Pond in the Rain), there’s very little actionable advice on how to actually do it. Magic, alchemy, etc. It’s often an exercise in vibe curation, or emotional precision, rather than technical precision. Even quintessential writer of rulebooks E.B. White caveats that nobody can say for certain why some words “ignite” and others don’t.
Start by listing what’s good about the writing you like. Learn a new word, or use an old word in a new context. Match your word choices to the mood you’re trying to evoke — for example, use short, simple words to explain something opaque; or choose a word like “scheme” over “plan” when you want the reader to smell trouble. There’s a know-it-when-you-hear-it quality to diction that you can’t compensate for with rules.
Some tells and when to edit them
Generic warmth. Friendly in a hold-message sort of way. “Great question!”
Recyclable phrasing. Low-stakes phrases and sentences that could be transplanted into any other piece of writing with few consequences. “A useful way to think about it is,” “The key idea is,” “This can be understood as.”
Low-friction vocabulary. Every word is precisely acceptable in an uncanny sort of way.
Abstract nouns: The writing leans on words like “efficiency,” “complexity,” “society,” “communication,” and “innovation.” A timeless tactic for saying nothing in 3+ syllables that long predates ChatGPT.
Insipid dynamism. “Navigate,” “leverage,” “unlock,” “foster,” both “power” and “empower,” “shape,” “elevate,” “streamline” — words that signal motion but fall flat.
A beacon of something. “A testament to...” “Stands as a beacon of...” “Serves as a reminder that...”
Gestures vaguely. “Landscape,” “space,” “journey,” “ecosystem,” “tapestry” — ambiguous, noncommittal words that point to the vicinity of things instead of identifying them directly.
Vague intensifiers. Phrases like “very important,” “significant impact,” or “major role,” especially when there’s nothing concrete backing them up.
Edit: Most of the time. Sometimes there is simply no other word for “ecosystem,” and we’re all just going to have to live with that.
Embrace: When being an NPC is kind of the whole thing. Support docs, error messages, terms of service, safety instructions, apologies at massive scale, and anything else read by a million strangers in a million contexts. Here, personality is friction and Alexa voice is a mercy.
Tips: Many of these tips focus on using LLMs for detection, not for writing. Models are actually very good at identifying jargon, business-speak, and other words frequently used in AI writing.
The Pangram “percent of AI in writing” debate aside, you can use an AI detector or any other LLM to flag the most generic words and phrases in your draft, and then edit accordingly. Prompt for hedging, fungible phrasing, etc.
From there, try running the transplant test on any given sentence. If a stranger could claim it, consider reworking.
Go through your laundry list of vague language and ask, “Is there a more specific word for this?” For example, “The Ethereum ecosystem is expanding” can become, with a little more specificity, “Developers are building more wallets, exchanges, and lending markets around Ethereum.” If you can clarify your meaning with more precise words, choose them every time.
Finally, people are starting to feed voice notes into LLMs as a way to get words on paper. An underappreciated benefit is finding personal idiosyncrasies that set your voice apart. Note that you do have to edit the results; there is no such thing as a hole-in-one here.
Structural hallmarks: All form, no function
AI writing tends to feel a bit overstructured when left to its defaults — too many H2s and lists, enough paragraph breaks to resemble a William Carlos Williams poem. But it’s not all bad. People have always drawn on a bank of familiar structures that work well and feel good on the brain.
We group ideas into three because three feels complete. We add signposts (e.g., “First,...” or “In other words, …”) to instantly situate readers within our writing. We create neat little taxonomies because they’re easier to skim.
As students, most of us learn some version of “hamburger” logic, where we tell our audience what we’re going to say, say it in discrete and supported chunks, and then restate it to conclude. It’s useful while forming an argument, because it forces us to articulate a thesis, collect evidence, and arrange our thoughts into something legible — a good thing!
A good, clear structure helps readers immediately understand their surroundings, whether that’s an op-ed, an explainer, or even a piece of fiction. The trouble is when pre-set structures pressure writers to draw and quarter their ideas into formats that don’t necessarily make sense.
Structure is a set of decisions — which container to use, which information matters most, which ideas go together, and what to call them. The right decisions depend on the job to be done. A narrative essay needs discovery and tension to keep readers moving. A product announcement needs to be brutally efficient to catch readers mid-scroll. An explainer needs an order that builds on concepts sequentially learned.
Ask yourself — anywhere in the writing process, but hopefully at the beginning — “what is the best format for my idea?” Then, borrow from something that already works. If you’re working on an argument piece, take a moment to understand how other writers structure op-eds and similar articles. If you’re writing a technical explainer, pick out the best explainer you’ve ever read, and look at how its author organizes information. Once you have a template, you can apply it to your own work.
Some tells and when to edit them
Over-organization. Lots of subheads, bullets, numbered sections, or mini-frameworks.
Ideas and lists that come in threes. Though this remains a best practice; more below.
Familiar essay shapes. Broad introduction, explanation, examples, caveat, conclusion.
Formulaic openings. “In today’s rapidly changing world…”
A little too much signposting. “First,” “Next,” “Finally,” “In conclusion,” “Here’s a breakdown,” “Let’s unpack this.”
Hard-pivot transitions. “To understand why this matters, we first need to look at…”
Section previews. “There are three key reasons…”
Bulleted lists. AI detectors often flag these unfairly; bullets are useful! A signature move, however, is the bold lead-in to a bullet point. See, for example, these exact bullets (but they’re so much easier to read this way).
Punchy fragments. Seen in many a dramatic LinkedIn post. Short. Punchy. Often in threes.
Ending on a restatement. In other words, the conclusion paraphrases the piece instead of expanding and hinting at further directions.
Moral-of-the-story endings. A vague final sentence about progress, the future, or “what we can learn.”
Edit:
Subheadlines that don’t naturally fit your format (op-eds, personal narratives, most pieces argued through voice and momentum).
When your sections aren’t solid or discrete (e.g., a “five takeaways” post with two very similar takeaways).
When structure dilutes or changes your meaning (e.g., a numbered list of reasons a startup pivoted reads very differently than the story of how it happened, even when facts are the same).
Signposts for self-evident structures (i.e., “three reasons why…”). Remove signposts that are just adding to your word count.
Embrace:
When the structure of a piece complements what it’s saying. Op-eds, for example, all tend to have a built-in argument that’s easy to pick out and follow.
Likewise, when writing listicles, explainers, how-tos, and other genres where we expect headers and subheadlines.
When an idea actually comes in threes. This principle is only a problem when it contorts ideas into unnatural forms. Or when it’s obvious the writer is arbitrarily forcing a third thing.
When optimizing for LLMs and search engines, which reward well-structured information.
When creating reference content people will return to rather than read once — documentation, guides, FAQs, and anything else readers navigate and scan vs. follow from beginning to end.
When creating content for skimmers. Headers can serve as a table of contents for readers deciding whether to stay or bounce within a tenth of a second. Headers taken together can even tell the whole argument, with the paragraphs underneath fleshing out the details.
Tips: Models are great at prompts like “organize this better,” but only if you actually want more organization. If not:
Try telling the LLM what your structure needs to do for the reader — whether that’s creating suspense or making a piece easier to scan.
Give the LLM a piece you like in the same genre and ask it to study how the argument works before it restructures yours.
Paste a published piece in the same genre, and prompt it to create a “reverse-outline,” describing the function each paragraph serves. Then draft using this rough structure as a skeleton.
Even if the model fails to provide, the extra thought put into organizing and presenting an idea tends to pay off.
Punctuation hallmarks: The em dash panic
The em dash — now a canonical AI tell — has always been divisive (😏). Strunk and White advise restraint more generally: “Use a dash only when a more common mark of punctuation seems inadequate.” But there’s a very easy, conversational feel to em dashes, in particular, that no other punctuation mark really has.
Em dashes are often seen as the enemy of efficient writing — it’s admittedly kind of distracting to jam this whole other phrase into the middle of a sentence — but as soon as they started signaling slop writing, they were an easy target for a ban.
The thing about punctuation, though, is that it has extremely specific rules, based on whatever 1,000+ page style guide you’re following. Whether AI favors em dashes is beside the point, because there are certain situations where you should probably use them. At the very least, they’re the most appropriate choice for setting apart longer parenthetical information, or for illustrating some abrupt changes in thought and emotion (as rhapsodized by horror writer R.L. Stine).
My take: shift + option + dash (on macOS) is humanity’s own troubled, distracted, often off-topic child, and I, for one, will never abandon it. Sometimes nothing hits quite like an em dash, so I urge you to avoid letting the current AI writing meta influence your choice of punctuation marks, not least because all metas must, by definition, change.
Now that people are bending over backward to avoid em dashes, AI itself is routing around them with colons. Shall we start using interrobangs to look appropriately human?! Since no punctuation mark is safe, choose the one that’s best for your writing.
So how do you tell what’s best for your writing? The short answer is: whatever is most correct and least distracting. The particulars of the rules can be nuanced and vary across taste and styles (Oxford vs. no Oxford comma is a popular religious debate among people who’ve made grammar and usage part of their personalities), so getting them perfect isn’t as important as ensuring they support your meaning and don’t infuriate your readers.
Another punctuation-related hallmark to look out for is sameness; that is, are all of your sentences looking and sounding the same? LLMs often use consecutive lists starting with a colon, which makes for an exhausting read. Likewise, several sentences interrupted by em dash asides will distract from your point, no matter who or what wrote it.
A good test is reading your writing aloud, using punctuation like stage directions. Anything that sounds unnatural will probably stand out to readers too.
Some tells and when to edit them
Colon-heavy phrasing. “The issue is:”, “The result is:”, “The key point is:”... but especially colons followed by grocery lists of things.
Em dash clustering. Not the dash itself, but the density, especially paired with lists and asides. There’s not really a numerical limit on the number of em dashes you can use in a sentence, but don’t use more than two.
Unserious parentheticals (like this one). These usually carry the self-aware or joking register, while dashes carry the qualifying one.
Performative semicolons. Writer Kurt Vonnegut once said that the only reason to use a semicolon is to show that you’ve been to college. It’s actually the least mean thing he said about semicolons, but it’s still unnecessarily crotchety. They have many legitimate use cases; however, these don’t come up all that often, honestly.
Edit: When punctuation is repetitive, distracting, or otherwise fails the read-aloud test.
Embrace: When it works. Punctuation should support your point and suit your taste. It should aim to be grammatically correct and mostly invisible to readers.
Tips: Don’t overthink whether punctuation is making you look like AI. Just use the em dash or the colon, or whatever is presently on the ban list. If you must:
Avoid the urge to tell a model to “remove all em dashes.” It will often just replace them with something else, while preserving the same underlying sentence structure.
Instead, ask to default to periods and commas, and only use “special” punctuation when grammatically required.
Again, giving an LLM examples and samples of your own writing (or writing you admire) helps build a “punctuation fingerprint” that estimates your use of a given mark.
As more people use these tools, interrogating whether something is machine-generated is becoming kind of pointless. The answer is almost always going to be “to some degree.” Of course, the egregious cases of AI writing still need outing, whether by our own standards or Proof of Person technologies: for example, when disclosure is required, when the human on the byline is the whole point, or when one person is pretending to be a thousand.
For almost everything else, we can ask the same question we’ve always asked: is the writing doing its job?
AI helps people articulate and publish ideas they’d otherwise never write down. It can save time on organization and research. It may even help us become better writers. It’s mean-spirited to dismiss the ability to express oneself as “cringe” or “slop.” And, if a piece works as intended, does it really matter which parts of the process show through?
Finally, with all of the hand-wringing over whether our writing will be exposed as AI-assisted, one question worth asking is: why are we showing LLMs this kind of deference at all? We would relinquish an entire punctuation mark — an innovation conceived at the dawn of printing! — all to avoid looking like a machine helped us. It’s a ridiculous concession, especially given how often people already use, and will continue to use, the machine.
Acknowledgments: Thanks to the a16z crypto editorial team — Tim Sullivan, Robert Hackett, and Sonal Chokshi — for feedback on this post, and for the many years of edits and debates that informed it.
Stop asking "does this look like AI," ask "is the writing doing its job." Fantastic.
Best article I’ve ever read on the topic. I’ll ask Claude to plagiarise it into a skill!