a16z Deep Dive: Stop Chasing the 'AI Smell', Here's a Practical Guide to Writing with AI

Author: Steph Zinn, a16z Crypto

Compilation: TechFlow

TechFlow Introduction: When people talk about AI writing, they often fall into two assumptions: machine-generated text can be identified by a set of "tells," and anything exhibiting these tells is inherently poor. The author argues that both are false. She dissects the "characteristics" of AI-generated content into four dimensions: rhetoric, voice, structure, and punctuation, pointing out these flaws have always existed—AI just amplifies them at scale. The focus of this article is not "how to spot AI," but rather to help founders and writers judge: which habits hinder effective expression, and when to restrain them versus when to embrace them.

Rhetorical Features: Writing That Sees the Shape

One of the most frustrating yet satisfying (and then relieving) editing experiences is discovering a sentence that reads well but feels off. You rearrange it, shift a few words, and slowly realize: it never meant anything to begin with. So you delete that slippery impostor and move on with your life.

A ready-made example comes from a recent Fable self-analysis on punctuation style: "Human punctuation has a body, it fidgets; mine is uniformly deliberate, and that deliberate neatness itself is a rhythm."

I'd describe this tone as "that person at a perfectly good party lecturing about jazz improvisation," but I won't dissect it sentence by sentence here—you get the idea. People look at this type of writing and say "no human could ever have written this." I understand what they mean, but as an editor, ghostwriter, and lifelong fantasy novel reader, I can assure you: humans write like this without AI all the time.

However, spotting this language isn't always as blatant as "fidgeting punctuation." Unlike other AI features (like three consecutive list sentences, or that familiar dash Morse code), these sentences don't look like problems. They just look like sentences. Some even sound more polished than the ones around them: "Something truly important is happening." "The stakes could not be higher." "The implications are far-reaching."

AI writing mass-produces these phrases because they are "plausible (though semantically empty)" connective tissue. Before it was linked to AI, we called this bland language "corporate-speak," because it's the classic, impassioned tone of corporate blogs. Can you think of a more boring way to say "I'm excited" than "We are pleased to announce..."? "We are at an inflection point" and "This is the next chapter of our journey" are two more examples. These phrases sound weighty but often lack a specific referent that survives retelling.

I've heard other editors dismiss such writing as "lazy" or "careless." That's uncharitable. It might be a matter of taste, but more often, it's genuinely difficult to push past the ubiquitous, readily available phrases floating at the front of one's mind and find a better way to say something. This is "first-draft language." Frankly, it's a fine starting point. We just don't want to stop there.

The method I recommend for avoiding meaningless language is unrelated to whether you've used AI. If you find a sentence that feels off—if it uses the phrases or constructions below, if you can't immediately grasp its meaning, or if it just "feels wrong"—start rearranging it, trying to rephrase it.

If a different wording is clearer, keep the better version. If it ultimately just says "things exist" or "things are changing," delete it and see if the piece stands without it. If it does, it's time to enjoy the "illicit thrill" of deleting words from the page.

Some Tells, and When to Edit Them

• La Croix Insight (Flavored Profundity). "Something truly important is happening." "The stakes could not be higher." Sounds important, is filler.
• Hollow Contrast. "It's not just about X—it's about Y," where Y is vaguer than X. This structure is occasionally useful when the contrast is crystal clear. But before leaving it as is, try deleting the "It's not just about" part. You might get to the point faster.
• Hedges. "In many ways." "In a sense." "Arguably." Each exists to make the statement "never wrong." Note: In some industries (like finance or healthcare), certain hedges are necessary to soften claims for compliance. I imagine AI would call these "load-bearing" hedges.
• Excessive Parallelism. Bullet points and sentences mirror each other too neatly. Each list item has the same grammatical shape, the same length. These are AI tells, but they also make writing incredibly boring.
• Summative Phrases. "At the end of the day." "When the dust settles." Deleting these often leaves your paragraph unaffected.

Edit: Almost always. Meaningless, empty verbiage is the top priority on our "writing pitfalls" list. No matter how good it sounds, it doesn't advance your idea. It's a scourge. Exterminate accordingly.

Embrace: Almost never.

Advice: I often use prompts that revolve around two goals: (1) making the text more specific, and (2) making the text more direct.

My aim is always to remove unnecessary jargon and move towards "information-dense, more readable." If an LLM can help me do that, good.

More ways to remove filler language:

• Create a rough style guide defining what "good" writing looks like for you, using positive and negative examples. This holds up better than a list of AI tells.
• Run the "paraphrase test" I mentioned above with AI—have it write a "boring version" of what you've written.
• Having an LLM "write at a 6th-grade level" is a decent, blunt tool for working on your prose.

Voice Features: The Alexa Voice

English has roughly a million words. Default AI writing sounds like it uses about 400 of them. But rather than struggling to pinpoint which "word of the day" is an AI marker, the test here should be "fungibility": could this sentence be lifted verbatim from your article and dropped into someone else's completely different piece without anyone noticing?

For a while, you could empirically track this "low-friction vocabulary." Researchers famously noted the spike of "delve" in academic paper abstracts after ChatGPT's launch in 2022, followed by waves of banned-word lists: tapestry, testament, underscore; more recently, load-bearing, scaffolding, broader.

Meanwhile, caring about word choice isn't just a literary author's obsession. It's important for anyone starting or running a company because people increasingly rely on personality for marketing.

The best founder writing is personal, expressive. Whether others think it's "good" or not, it makes the reader feel the person who believes it. That's why we can so easily imagine Brian Armstrong's style (concise, sincere, unironic), Vitalik Buterin's style (technical, digressive, dense), or Chris Dixon's style (restrained, philosophical).

So, when using an LLM, be careful not to hand over your personality to it. Your own word choices and their imperfections add a "patina" a machine can never replicate.

Despite the many volumes on "how to find the right word"... practical advice is scarce. Magic, alchemy. It's often a matter of "vibe curation" or "emotional" precision, not "technical" precision. Even E.B. White, the textbook rule-maker, admitted no one could say why some words "catch fire" and others don't.

Start by listing what you like about writing you admire. Learn a new word, or use an old one in a new context. Let your word choice match the mood you want to evoke—e.g., use short, simple words to explain something obscure; or choose "scheme" over "plan" when you want the reader to "smell trouble."

Some Tells, and When to Edit Them

• Generic Warmth. "Great question!" A customer-service-notes friendliness.
• Recyclable Sentence Structures. Those low-risk, transplantable phrases that could fit almost any other article. "A useful way to think about this is..." "The core idea is..." "This can be understood as..."
• Low-Friction Vocabulary. Every word is "just right," with an eerie appropriateness.
• Abstract Nouns. Text leans on words like "efficiency," "complexity," "society," "communication," "innovation." This is the ancient tactic of "saying nothing," long predating ChatGPT.
• Lifeless Action Verbs. "Navigate," "leverage," "unlock," "foster," including both "power" and "empower," "shape," "enhance," "streamline"—words that gesture at motion but land with a thud.
• Beacon of Something. "A testament to..." "A beacon for..." "A reminder that..."
• Vague Hand-Waving. "Landscape," "space," "journey," "ecosystem," "tapestry"—fuzzy, non-committal words pointing near the thing rather than naming it.
• Vague Intensifiers. Phrases like "very important," "significant impact," "critical role," especially when unsupported by specifics.

Edit: Most of the time. Sometimes "ecosystem" genuinely has no substitute, and we all have to make do.

Embrace: When being an NPC is the entire point. Support docs, error messages, terms of service, safety notices, mass apologies to vast numbers of strangers—in these places, personality is friction, and the Alexa voice is a mercy.

Advice: Most of these suggestions focus on using LLMs for "detection" rather than "writing." Models are actually quite good at identifying jargon, corporate-speak, and other vocabulary common in AI writing.

• Setting aside the "what percentage AI is this writing" debate, you can use an AI detector or any LLM to flag the most generic words and phrases in a draft, then edit accordingly. Use prompts targeting hedges, fungible sentence structures, etc.
• Afterward, try running any sentence through the "transplant test." If a stranger could claim it, consider rewriting.
• Go through your list of vague language and ask, "Is there a more specific word?" For example, "The Ethereum ecosystem is expanding" could become—slightly more specifically—"Developers are building more wallets, exchanges, and lending markets around Ethereum." If you can convey the meaning with a more precise word, choose it every time.
• Finally, people are starting to feed voice notes into LLMs to get ideas on paper. An underrated benefit: you can spot the personal quirks that make a voice distinctive. Note you must edit the output; there's no "hole-in-one" here.

Structural Features: Form Without Function

Left to its own devices, AI writing often seems "over-structured"—too many H2s and lists, paragraphs broken up like a William Carlos Williams poem. But this isn't all bad. People have always borrowed from a familiar, comfortable, and readable stock of structures.

We divide ideas into three points because "three" feels complete. We add signposts (like "First..." or "In other words...") to help readers instantly orient themselves. We create neat little categories because they're easier to scan.

As students, most of us learned some form of the "hamburger" logic: tell readers what you'll say, say it in discrete, supported sections, then tell them what you said. This is useful for constructing arguments because it forces us to state a thesis, gather evidence, and arrange thoughts into a readable shape—a good thing!

Clear, good structure helps readers immediately understand "where they are," whether it's an op-ed, an explainer, or a short story. The trouble is that default structures can pressure the writer, forcing ideas into formats that may not fit.

Structure is a set of decisions—what container to use, what information is most important, which ideas belong together, how to label them. The right decisions depend on "the job to be done." Narrative essays need "discovery" and "tension" to pull readers forward; product launch announcements need extreme efficiency to grab readers mid-scroll; explainer articles need a sequence that "builds step-by-step on already-learned concepts."

Ask yourself—at any stage of writing, but ideally early on—"What's the best format for this idea?" Then, borrow from what already works. If you're writing an op-ed, spend some time understanding how other authors organize arguments and similar pieces. If you're writing a technical explainer, pick the best explainer you've ever read and see how the author structured the information. With a template, you can apply it to your own writing.

Some Tells, and When to Edit Them

• Over-organization. An abundance of subheadings, bullet points, numbered paragraphs, or mini-frameworks.
• Always Three Points. Though the "rule of three" remains best practice; more on this below.
• Familiar Article Shape. Broad intro, explanation, examples, reminder, conclusion.
• Formulaic Opening. "In today's rapidly changing world..."
• Excessive Signposting. "First," "next," "finally," "in summary," "here's the breakdown," "let's break this down."
• Clunky Transitions. "To understand why this matters, we must first look at..."
• Chapter Previews. "There are three key reasons..."
• Bulleted Lists. AI detectors often unfairly flag them; but lists are useful! However, a telltale move is using bold lead-ins for bullet points. For example, see these exact points (but it does make for much easier reading).
• Short, Punchy Fragments. Common in many dramatic LinkedIn posts. Short. Punchy. Often in threes.
• Conclusion that Restates. I.e., a conclusion that merely rephrases the whole piece rather than extending it or pointing to new directions.
• Preachy Ending. A vague wrap-up about progress, the future, or "what we can learn."

Edit:

• Subheadings that don't suit your format (op-eds, personal narratives, and most pieces driven by voice and rhythm).
• When your sections aren't solid or distinct enough (e.g., two very similar takeaways in a "5 Key Takeaways" list).
• When the structure dilutes or changes your original meaning (e.g., a numbered list of "why a startup pivoted" reads very differently from the story of "how it happened," even if the facts are the same).
• Signposts that introduce "self-evident structure" (i.e., "three reasons..."). Delete signposts that are just filling space.

Embrace:

• When the article's structure echoes what it's saying. For example, op-eds often have an easily recognizable, easy-to-follow argument.
• Similarly, when writing listicles, explainers, how-to guides, and genres where we expect headings and subheadings.
• When ideas genuinely come in threes. This principle is only a problem when it twists ideas into unnatural shapes, or when the author is obviously straining for a third thing.
• When optimizing for LLMs and search engines, which reward well-structured information.
• When creating reference content meant to be "consulted repeatedly" rather than "read start-to-finish"—docs, guides, FAQs, and anything readers "browse and scan" rather than "read sequentially."
• When creating content for "scanners." Headings can act as a table of contents, letting readers decide in a tenth of a second whether to stay or leave. Headings can even tell the whole argument when read together, with paragraphs below filling in the details.

Advice: Models are quite good at prompts like "organize this better," but only if you actually want more organization. If you don't:

• Try telling the LLM what your structure needs to do for the reader—whether that's creating suspense or making the piece more scannable.
• Give the LLM an article in the same genre that you like, have it study how that argument works, then restructure yours.
• Paste a published article in the same genre and have the LLM do a "reverse outline," describing what each paragraph does. Then use that rough structure as a skeleton for drafting.

Even if the model doesn't give you something, the extra thought invested in "organizing and presenting an idea" often pays off.

Punctuation Features: Dash Panic

The em dash—now a textbook AI tell—has always been controversial (😏). Strunk and White generally advise restraint: "Use a dash only when a more common mark of punctuation seems inadequate." But the dash, especially, has a very casual, conversational feel that other punctuation lacks.

Dashes are often seen as the enemy of efficient writing—admittedly, shoving a whole other sentence into the middle of a sentence is distracting—but once they become "markers" of low-quality writing, they become easy targets for prohibition.

But punctuation has extremely specific rules, depending on which 1000+ page style guide you follow. Whether AI favors dashes or not is irrelevant because there are scenarios where you should use them. At the very least, dashes are the most appropriate choice for setting off long parentheticals or portraying a sudden shift in thought or mood (as horror writer R.L. Stine has passionately extolled).

My take: Shift + Option + Dash (macOS) is humanity's own troublesome, distracted, often-tangential child, and I, for one, will never abandon it. Sometimes, nothing fits quite like a dash, so I urge you not to let the current "AI writing zeitgeist" influence your punctuation choices, especially since all zeitgeists, by definition, change.

Now people are desperately avoiding dashes, and AI itself bypasses them with colons. Should we start using interrobangs to seem human enough?! Since no punctuation is safe, choose the one that works best for your writing.

So, how to judge what's best for your writing? The short answer: the one that's most correct and least distracting to the reader. The specifics of the rules can be subtle and vary by taste and style (Oxford comma vs. no Oxford comma is a religious debate popular among those who turn grammar and usage into "part of their personality"), so aiming for perfection is less important than ensuring they support your meaning and don't irritate the reader.

Another punctuation-related feature to watch for is "sameness": Do all your sentences look and sound the same? LLMs' frequent use of colon-led list sentences reads tiresome. Similarly, multiple sentences interrupted by parenthetical dashes, regardless of who writes them, scatter your focus.

A good test: Read your writing aloud, using punctuation as stage directions. Anywhere it sounds unnatural, a reader will likely notice too.

Some Tells, and When to Edit Them

• Colon-Heavy Constructions. "The problem is:" "The result is:" "The key point is:" ... especially when followed by a grocery list of items.
• Dash Clustering. Not dashes themselves, but their density, especially when paired with lists and interjections. There's no hard limit on dashes per sentence, but don't exceed two.
• Flirty Parentheses (like this one). These typically carry a "self-aware" or joking register, while dashes carry a restrictive one.
• Performative Semicolons. Writer Kurt Vonnegut once said the only reason to use a semicolon is "to show you've been to college." That's actually the nicest thing he said about semicolons, but it's still unnecessarily grumpy. Semicolons have many legitimate uses; though, seriously, those uses don't come up often.

Edit: When punctuation is repetitive, distracting, or otherwise fails the "read aloud" test.

Embrace: When it works. Punctuation should support your argument and suit your taste. Its goal is grammatical correctness and near-invisibility to the reader.

Advice: Don't overthink whether punctuation makes you seem like AI. Just use dashes or colons or whatever is currently on the "banned list." If you must:

• Don't rush to have a model "delete all dashes." It often just replaces them with something else while keeping the same underlying sentence structure.
• Instead, ask it to default to periods and commas, using "special" punctuation only where grammar demands it.
• Again, giving an LLM your own writing (or writing you admire) as a sample helps establish a "punctuation fingerprint" to gauge your frequency of use for a given symbol.

As more people use these tools, asking "was this machine-generated?" is becoming somewhat pointless. The answer is almost always "in some way." Of course, egregious cases of AI writing still need to be exposed, whether by our own standards or via "Proof of Person" technologies: e.g., when disclosure is required, when the person signing it is the entire point, or when one person pretends to be a thousand.

For almost everything else, we can ask the question we've always asked: Is this writing "doing its job"?

AI helps people express and publish ideas they otherwise never would have written down. It can save time on organization and research. It might even help us become better writers. Dismissing "the ability to express oneself" as "awkward" or "low-quality" is unkind. And if a piece works as intended, does it really matter which parts "gave it away"?

Finally, amidst all the anxiety about "whether our writing will be outed as AI-assisted," there's a question worth asking: Why are we showing this "awe" towards LLMs? We're willing to surrender an entire punctuation mark—an invention dating back to the dawn of printing!—just to avoid seeming like we had machine help. This is an absurd concession, especially considering how often people already have, and will continue to, use this machine.

Acknowledgments: Thanks to the a16z crypto editorial team—Tim Sullivan, Robert Hackett, Sonal Chokshi—for feedback on this piece, and for the countless edits and discussions that shaped it over the years.