Why Does Your Writing Sound Like AI? a16z Editors Teach You How to Reclaim the 'Human Touch'

Author: Steph Zinn, a16z Crypto

Compiled by: Deep Tide TechFlow

Deep Tide Introduction: When people talk about AI writing, they often fall into two assumptions: that machine-generated text can be identified by a set of "flaws," and that any text with these flaws must be poor. The author argues that both of these assumptions are incorrect. She breaks down the "features" of AI-generated content into four dimensions: rhetoric, voice, structure, and punctuation, pointing out that these issues have always existed—AI has just amplified them in bulk. The focus of this article is not on "how to spot AI" but on helping founders and writers determine which habits hinder effective expression and when to restrain or retain them.

Rhetorical Features: Writing with Insight {#article-toc-33527-2}

One of the most frustrating yet satisfying experiences in editing is discovering a sentence that reads well but feels off in some way. You rearrange it, shift a few words, and slowly realize: it has no meaning from the start. So you delete that slippery imposter and move on with life.

A ready example comes from Fable's recent self-analysis on punctuation style: "Human punctuation has a body; it fidgets; my punctuation is all deliberate, and the very neatness of it is rhythm."

I would describe this tone as "someone talking about jazz improvisation at a party that was otherwise quite nice," but I don't intend to dissect it sentence by sentence—you get my point. People say about this kind of writing, "This could never have come from a human hand." I understand what they mean, but as an editor, ghostwriter, and lifelong fantasy reader, I can assure you: humans often write sentences like this without AI.

However, identifying this kind of language is not always as straightforward as "fidgety punctuation." Unlike other AI characteristics (like three consecutive sentences being lists or that familiar dash code), these sentences don't seem problematic at all. They just look like sentences. Some even sound smoother than the surrounding sentences: "Something important is happening." "This matters." "Its impact is profound."

AI writing churns out these phrases in bulk because they are "seemingly reasonable (though semantically hollow)" connective materials. Before associating them with AI writing, we called this bland language "corporate," as it embodies that typical, unremarkable business blog tone. Can you think of a more boring way to express "I’m excited" than "We are pleased to announce..."? "We are at a turning point" and "This is the next chapter of our journey" are two more examples. These phrases sound substantial but often lack specific references that can withstand paraphrasing.

I’ve heard other editors misjudge this kind of writing as "lazy" or "careless." That’s not fair. It may be a matter of taste, but more often than not, it’s not easy to shake off those clichéd, readily available phrases and idioms floating at the forefront of our minds to find better expressions. This is "draft language." Honestly, it’s perfectly fine as a starting point. We just don’t want to stop there.

The method I recommend for avoiding meaningless language has nothing to do with whether you’ve used AI. If you find a sentence that feels off—if it uses any of the phrases or structures below, if you can’t immediately grasp its meaning, or if it just has the "wrong vibe"—then start rearranging it, trying to rephrase it.

If a different phrasing is clearer, keep the better version. If it ultimately boils down to "things exist" or "things are changing," then delete it and see if the article holds up without it. If it does, it’s time to enjoy that "forbidden pleasure" of deleting words from the page.

Some Flaws and When to Edit Them {#article-toc-33527-3}

• Natural Flavor Insights (La Croix-style insights). "Something important is happening." "The risk is at its highest." Sounds important, but is filler.
• Hollow Comparisons. "This is not just about X—but about Y," where Y is more vague than X. This structure can be useful when the comparison is very clear. But before leaving it as is, try deleting the preceding phrase "not just about." You might get to the point faster.
• Hedge Language. "In many ways." "To some extent." "It can be said." Each of these exists to ensure the statement "can never be wrong." Note that in some industries (like finance or healthcare), some hedge language is necessary to soften claims and meet compliance needs. I think AI would call these "load-bearing" hedges.
• Overly Balanced Structures. Bullet points and sentence structures mirror each other too neatly. Each list item has the same grammatical shape and length. These are AI flaws, but they also make writing extremely dull.
• Summary Phrases. "At the end of the day." "When the dust settles." Deleting these phrases often has no impact on your paragraphs.

Edit: Almost always. Meaningless, nonsensical platitudes top our prioritized "writing traps" list. No matter how nice they sound, they do not advance your ideas. They are a plague. Clear them out accordingly.

Embrace: Almost never.

Advice: I often revolve around two types of prompts: (1) prompts that make writing more specific; (2) prompts that make writing more straightforward.

My goal is always to remove unnecessary jargon and strive for "high information density and readability." If LLM can help me achieve that, great.

More methods for clearing filler language:

• Create a rough style guide that defines what "good" writing looks like for you with examples of both good and bad. This is more robust than listing a bunch of AI flaws.
• Use AI to run the "paraphrase test" I mentioned above—directly have it help you write a "boring version" of what you wrote.
• Having LLM "write at a sixth-grade level" is a somewhat useful blunt tool to tackle your prose.

Voice Features: Alexa Voice {#article-toc-33527-4}

There are about a million words in English. Default AI writing sounds like it’s using about 400 of them. But rather than struggling to pinpoint which "hot word of the day" is an AI hallmark, the test here should be "fungibility": can this sentence be pulled verbatim from your article and dropped into someone else’s piece on a completely different topic without anyone noticing?

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

Meanwhile, careful word choice is not just a concern for "literary" authors. It’s important for anyone starting or running a company, as people increasingly rely on personality to market.

The best founder writing is personal and expressive. Whether others consider it "good" or not, it makes readers feel the person who believes in it. That’s why we can so easily imagine the writing styles of Brian Armstrong (concise, sincere, un-ironic), Vitalik Buterin (technical, divergent, dense), or Chris Dixon (restrained, philosophical).

So when using LLM, be sure not to hand over your personality to it. Your word choices and their flaws will add a patina that machines can never replicate.

Despite the abundance of writings on "how to find the right words," truly actionable advice is scarce. It’s often a matter of "atmospheric curation" or emotional precision rather than technical precision. E.B. White, a textbook rule-maker, also admitted that no one can explain why some words "ignite" readers while others do not.

Start by listing "what you like about good writing." Learn a new word or use an old word in a new context. Let your word choices match the atmosphere you want to evoke—e.g., use short, simple words to explain obscure things; or choose "scheme" instead of "plan" when you want readers to "smell trouble."

Some Flaws and When to Edit Them {#article-toc-33527-5}

• Generic Warmth. "Great question!" A friendly, customer service-style remark.
• Clichéd Phrasing. Those low-risk phrases that can be transplanted into almost any other article without issue. "A useful way to think about this is" "The core point is" "This can be understood as."
• Low-Friction Vocabulary. Every word is "just right," carrying an eerie appropriateness.
• Abstract Nouns. Writing relies on words like "efficiency," "complexity," "society," "communication," "innovation." This is an old tactic of "saying nothing" that predates ChatGPT.
• Bland Action Verbs. "Navigate," "leverage," "unlock," "cultivate," including both "power" and "empower," "shape," "enhance," "streamline"—those words that suggest "movement" but fall flat.
• Lighthouse of Something. "Is a testament to..." "Is a beacon of..." "Reminds us of..."
• Vague Language. "Framework," "landscape," "journey," "ecosystem," "tapestry"—words that point to things "nearby" rather than directly naming them, vague and non-committal.
• Vague Intensifiers. Phrases like "very important," "significant impact," "key role," especially when lacking specific content to support them.

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

Embrace: When "being an NPC" is the entire point. Support documents, error messages, terms of service, safety notices, large-scale apologies to strangers—these are places where personality is friction, and Alexa voice is a kindness.

Advice: These suggestions mostly focus on using LLM for "detection" rather than "writing." The model is actually quite good at identifying jargon, corporate speak, and other vocabulary commonly found in AI writing.

• Regardless of the debate over the percentage of writing done by AI, you can use AI detectors or any LLM to highlight the most common words and phrases in your drafts, and then modify accordingly. Focus on avoiding vague language and suggesting alternative phrasing.
• Next, try running any sentence through a "transplant test." If a stranger can claim it, consider rewriting it.
• Go through your list of vague language and ask, "Are there more specific words?" For example, "The Ethereum ecosystem is expanding" can be made slightly more specific by saying, "Developers are building more wallets, exchanges, and lending markets around Ethereum." Choose the more precise words whenever possible.
• Finally, people are starting to feed voice notes into LLMs to get their ideas down on paper. An underrated benefit is that you can discover personal quirks that make your voice unique. Remember, you must edit the results; there’s no such thing as a "hole-in-one" here.

Structural Features: Form Without Function {#article-toc-33527-6}

When AI takes over, its writing often appears "over-structured"—with too many H2s and lists, and paragraphs broken up enough to rival a poem by William Carlos Williams. However, this isn’t without its merits. People have always borrowed from a familiar, useful, and comfortable structure.

We break ideas into three points because "three" feels complete. We add signposts (like "First..." or "In other words...") to help readers locate themselves immediately. We create neat little categories because they are easier to browse.

As students, we often learned some form of "hamburger" logic: first tell the reader what you’re going to talk about, then present it in discrete, supported sections, and finally restate it to conclude. This is very useful for constructing arguments because it forces us to clarify our points, gather evidence, and arrange our thoughts in a readable manner—this is a good thing!

Clear, good structure helps readers immediately understand "where they are," whether it’s a commentary, a popular science piece, or a novel. The trouble is that preset structures can pressure authors to force their ideas into formats that may not be reasonable.

Structure is a set of decisions—what container to use, which information is most important, which ideas go together, and how to name them. The right decisions depend on "the work to be done." Narrative essays need "discovery" and "tension" to propel readers forward; product release announcements need to be extremely efficient to catch readers as they scroll; popular science articles need a sequence of "building on learned concepts."

Ask yourself—at any stage of the writing process, but best at the beginning—"What is the best format for this idea?" Then borrow from what has already worked. If you’re writing an argumentative essay, take some time to understand how other authors organize commentaries and similar articles. If you’re writing technical popular science, pick the best piece you’ve read and see how the author organizes the information. With a template, you can apply it to your own writing.

Some Pitfalls and When to Edit Them {#article-toc-33527-7}

• Over-organization. Excessive subheadings, bullet points, numbered paragraphs, or mini-frameworks.
• Always forcing three points. Although the "rule of three" remains best practice; this will be elaborated on later.
• Familiar article shapes. Broad introductions, explanations, examples, reminders, conclusions.
• Formulaic openings. "In today’s fast-changing world..."
• Excessive signposting. "First," "Next," "Finally," "In conclusion," "Here’s the breakdown," "Let’s break this down."
• Stilted transitions. "To understand why this is important, we first need to look at..."
• Chapter previews. "There are three key reasons..."
• Bullet point lists. AI detectors often unfairly flag them; but lists are useful! A hallmark move is to lead into bullet points with bolded prompts. For example, check out these exact points (but this does indeed make reading much easier).
• Short, choppy fragments. Common in many dramatic LinkedIn posts. Short. Punchy. Often in threes.
• Restating to conclude. In other words, the conclusion merely rephrases the entire text rather than extending or hinting at new directions.
• Didactic endings. A vague wrap-up about progress, the future, or "what we can learn."

Edit:

• Subheadings that don’t fit your format (commentaries, personal narratives, and most articles driven by voice and rhythm).
• When your sections aren’t solid or independent enough (for example, if a "five takeaways" piece has two very similar takeaways).
• When the structure dilutes or alters your original intent (for instance, a numbered list of "why a startup pivoted" reads very differently from a story of "how it happened," even if the facts are the same).
• Signposts added to "self-evident structures" (i.e., "three reasons..."). Remove those that are just filler.

Embrace:

• When the structure of the article itself echoes the content it’s trying to convey. For example, commentaries often come with an easily recognizable, easy-to-follow argument.
• Similarly, when writing lists, popular science, how-to guides, and genres where we expect titles and subtitles.
• When ideas genuinely appear in threes. This principle only becomes problematic when ideas are distorted into unnatural forms; or when the author is clearly forcing a third item.
• When optimizing for LLMs and search engines, they reward well-structured information.
• When creating content that people will "reference repeatedly" rather than "read from start to finish"—documents, guides, FAQs, and any content where readers "browse and scan" rather than "read sequentially."
• When writing for "scanning" readers. Titles can serve as a table of contents, allowing readers to decide in a tenth of a second whether to stay or leave. Titles can even encapsulate the entire argument, with the following paragraphs fleshing out the details.

Advice: Models excel at prompts like "make this more organized," but only if you genuinely want more organization. If you don’t want that:

• Try telling the LLM what your structure needs to do for the reader—whether to create suspense or make the article easier to scan.
• Give the LLM a piece of the same genre that you like, let it first analyze how that argument works, and then reconstruct yours.
• Paste a published article of the same genre and have the LLM create a "reverse outline," describing the role of each paragraph. Then use this rough structure as a skeleton to draft.

Even if the model doesn’t provide you with something, the extra thought put into "organizing and presenting an idea" often pays off.

Punctuation Features: Dash Panic {#article-toc-33527-8}

The em dash has become a textbook AI pitfall—historically controversial (😏). Strunk and White generally advise restraint: "Use dashes only when more common punctuation proves inadequate." However, dashes, especially, have a very relaxed, conversational feel that other punctuation lacks.

Dashes are often seen as the enemy of efficient writing—it must be admitted that cramming an entire other thought into the middle of a sentence can indeed be distracting—but once they begin to "mark" low-quality writing, they become a target for prohibition.

However, punctuation has very specific rules, depending on which 1000-page style guide you follow. Whether AI favors dashes is irrelevant, as there are some scenarios where you should use them. At least, when separating longer inserted information or depicting sudden shifts in thought and emotion, dashes are the most appropriate choice (as horror writer R.L. Stine passionately extols).

My take: Shift + Option + dash (macOS) is the troublesome, distracted, often off-topic child of humanity, and I, at least, will never abandon it. Sometimes, nothing fits quite as perfectly as a dash, so I advise you not to let the current "AI writing trend" influence your punctuation choices, especially since all trends, by definition, will change.

Nowadays, people are desperately avoiding dashes, while AI itself sidesteps them with colons. Should we start using "interrobangs" to seem human enough?! Since no punctuation is safe, just choose what fits your writing best.

So, how do you determine what’s best for your writing? The short answer is: the most correct and least disruptive to the reader. The specific details of the rules can be subtle and vary by taste and style (the Oxford comma vs. no Oxford comma is a popular religious debate among those who make grammar and usage a "part of their character"), so achieving perfection is less important than ensuring they support your meaning and don’t irritate the reader.

Another punctuation-related feature to watch out for is "monotony": do all your sentences look and sound the same? LLMs often use lists starting with colons, which can become tiresome to read. Similarly, multiple sentences interrupted by dashes, regardless of who wrote them, will dilute your focus.

A good test is to read your writing aloud, using punctuation as stage directions. Any place that sounds unnatural is likely to be noticed by readers as well.

Some Pitfalls and When to Edit Them {#article-toc-33527-9}

• Colon-heavy sentences. "The problem is: " "The result is: " "The key point is: "... especially when followed by a string of grocery-list-like content.
• Dash overuse. Not the dashes themselves, but their density, especially when paired with lists and interruptions. There’s no hard limit on how many dashes can be used in a single sentence, but do not exceed two.
• Casual parentheses. (Like this.) These often carry a tone of "self-awareness" or humor, while dashes convey a restrictive tone.
• Performative semicolons. Writer Kurt Vonnegut once said the only reason to use a semicolon is "to show you went to college." This is actually the politest thing he said about semicolons, but it still carries a bit of unnecessary hostility. Semicolons have many legitimate uses; however, to be honest, these uses don’t come up often.

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

Embrace: When it works. Punctuation should support your argument and align with your taste. Its goal is to be grammatically correct and nearly invisible to the reader.

Advice: Don’t get too hung up on whether punctuation makes you seem like AI. Just use dashes or colons, or any currently "prohibited" symbols. If you must:

• Don’t rush to have the model "remove 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 when grammatically required.
• Again, emphasize giving the LLM your own writing (or writing you admire) as examples, which helps establish a "punctuation fingerprint" to estimate your frequency of use for a particular symbol.

As more people use these tools, questioning whether something is machine-generated is becoming somewhat meaningless. The answer is almost always "to some extent." Of course, those egregious cases of AI writing still need to be exposed, whether by our own standards or through "Proof of Person" technology: for instance, when disclosure is required, when the person signing is the entire point, or when one person pretends to be a thousand.

As for almost everything else, we can ask the question we've always been asking: does this text "do its job"?

AI helps people express and publish thoughts that would otherwise never be written down. It can save time in organizing and researching. It might even help us become better writers. To dismiss "the ability to self-express" as "awkward" or "low quality" is a cruel act. Moreover, if an article works as intended, how important is it really which part "gave it away"?

Finally, amid the anxieties of "will our writing be exposed as AI-assisted," there is a question worth asking: why do we feel this "awe" towards LLMs? Why are we willing to give up an entire punctuation mark—an invention that was born at the dawn of printing—just to avoid appearing as if we had machine assistance? This is an absurd concession, especially considering how frequently people have already used and will continue to use this machine.

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

This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.