Default Voice: Why Your Studio’s Words Look Like System Font

You would never ship a brand identity set in the operating system’s default font. The whole point of the craft is that Helvetica-because-it-was-there communicates nothing except that nobody made a choice. Designers will argue for an hour about the counters on a lowercase g. Then the same studio publishes a case study, an about page, a project announcement, and the words are pure default: smooth, symmetrical, and identical in texture to every other studio’s words, because the same model drafted all of them.

Call it default voice. It is what happens when generative AI writes your copy and nobody kerns it afterward. And in a field that lives on distinctiveness, it is quietly becoming the most common branding mistake of the decade.

Type people already understand this

Here is the thing: nobody is better equipped to spot the problem than people who work with letterforms, because prose has typography. Not the visual kind, the structural kind.

A paragraph has rhythm the way a typeface has rhythm. Sentence lengths are strokes: some long and sweeping, some clipped. Word choice is weight and contrast. A writer’s quirks, the odd construction, the sudden short sentence, the phrase nobody else would pick, are the ink traps and irregular terminals that make a face feel alive at text sizes. Strip all that out and you get the prose equivalent of a geometric sans drawn by committee: technically clean, perfectly even, and dead on the page.

That evenness is precisely what language models produce by default. They are prediction machines, built to choose the statistically likely next word, and likelihood is the enemy of character. The result is text with no burstiness, no texture, no surprise. Every sentence lands in the same middle register, the way every letter in a bad revival has the same width. Readers may not articulate why it feels hollow. They feel it anyway, the same way a client who knows nothing about type can tell your identity work “feels off” when the spacing is wrong.

Design Twitter spent years mocking the sameness of startup branding, the identical geometric sans, the same five-color gradient blobs. Default voice is the same flattening, one layer down, and it is spreading faster because the tool that produces it is free and instant.

Where it actually hurts a studio

A studio’s writing carries more commercial weight than most designers admit. The case study is where you justify your rates. The about page is where a client decides whether they want to be in a room with you. The project announcement is how peers, and awards juries, and future hires, read your judgment. These are not neutral containers for information. They are specimens of your taste, exactly like everything else you ship.

Now run the current pipeline. A founder asks a model for “a case study about our rebrand for a sustainable coffee brand.” Out comes something fluent and hollow: “We embarked on a comprehensive brand journey, crafting a cohesive visual identity that resonates with modern consumers.” No specifics, no opinion, no friction, no evidence a human ever held the work in their hands. Multiply that across a portfolio and the studio has spent thousands of hours making the work distinctive and thirty seconds making it sound interchangeable.

The irony stings. We are in an era when craft is the pitch. Studios sell taste, discernment, the human hand. Publishing machine-average prose under that banner is like presenting a hand-lettering portfolio in a Word template. The contradiction is visible to exactly the audience you most need to convince.

Fixing it, fortunately, does not mean giving up the tools. It means treating the words like everything else the studio ships.

Fixing the typography of your prose

The good news: this is a craft problem, and studios are good at craft problems. The workflow that works looks a lot like any other design process. Generate roughs fast, then invest the human hours where they show.

Let the model do structure. Outlines, first drafts, alternative angles, the blank-page problem. This is genuinely what it is good at, and refusing the speed is romanticism, not rigor.

Then art-direct the words. Go through the draft the way you would go through a type proof. Vary the line lengths: if every sentence is the same size, cut one to four words and let another run. Replace the five most predictable word choices with the ones you would actually say. Add the specific detail only your team knows, the client’s weird constraint, the version everyone hated, the print vendor drama. Specificity is to prose what a distinctive terminal is to a glyph: the thing that proves a particular hand was here.

Some teams add a tooling pass between draft and polish. Text-humanization tools exist for exactly this flattening problem; they rewrite machine-drafted copy to restore the statistical variation that reads as human, before a person does the final pass for voice and accuracy. If you want to see the category, undetectedgpt.ai is one place to start, and the honest way to use anything in this space is the same as the honest way to use a grid template: as a foundation you then make yours, not as the finished work. A tool can restore texture. Only your team can restore taste.

Finally, read it aloud. The oldest proofing trick survives every technology shift for a reason. Default voice is unmistakable to the ear: it sounds like nobody. If a paragraph could appear on any competitor’s site without anyone noticing, it is system font. Reset it.

A specimen, before and after

Because this is a type crowd, here is the argument as a specimen sheet. Take a typical machine-drafted case-study opener:

“We partnered with a leading coffee brand to develop a comprehensive visual identity that seamlessly blends sustainability with modern design sensibilities, creating a cohesive brand experience that resonates across all touchpoints.”

Set in prose terms: mono-width sentences, no contrast, every word the most probable choice, zero evidence of a hand. Now the reworked version a human might actually sign:

“Terra came to us with a problem we liked: their beans were exceptional and their bag looked like a pharmacy product. Eleven weeks, two rejected directions, and one heated argument about brown later, they had an identity that finally matched what was inside it.”

Same project, same facts. But the second version has typographic life: a long unspooling sentence against a short one, an opinion, a number, a joke that costs something. No classifier taught on statistical evenness reads that as machine output, and more to the point, no prospective client reads it as filler. The rewrite took four minutes. That ratio, minutes of voice work against the credibility of the whole page, is the cheapest trade in your studio.

Do this exercise once with your own homepage and the difference stops being theoretical. Most teams find their published copy sits a lot closer to specimen one than they would like to admit.

The new gate

There is one more reason to take the rework seriously, beyond taste. In the last two years, text started getting screened. Editors of design publications now scan contributed articles before accepting them. Content platforms screen submissions. Some clients quietly run agency deliverables through AI checkers before sign-off, especially where copywriting is part of the engagement. A flagged deliverable starts an awkward conversation about what, exactly, the client paid for.

It is worth understanding how AI detectors actually score text, because the mechanics are surprisingly close to how a type designer evaluates rhythm. Detectors measure predictability and uniformity: how likely each next word is, how similar sentences are in shape and length, how even the overall texture reads. Statistically flat text scores as machine-made. Text with genuine variation, the burst-and-lull of a human actually thinking, scores as human. A classifier is, in effect, squinting at your paragraph the way you squint at spacing, asking whether the color is suspiciously even.

The messy middle is where it gets interesting for working designers. Research from 2025, a paper titled “Almost AI, Almost Human,” looked at what happens with hybrid text, human writing lightly polished by AI or machine drafts lightly edited by people, and found that classifiers handle that gray zone far less cleanly than the pure cases. Which describes basically all real studio copy now: your words, smoothed by a grammar tool, restructured by a model, tweaked again by you. The practical consequence cuts both ways. Honest work sometimes gets flagged because polish flattened it, and lazy work sometimes slips through. The gate is real, but it is measuring texture, not virtue.

If your natural register is already minimal and controlled, and a lot of design writing is, machine polish can push it over the line into flagged territory without a single dishonest intent anywhere in the process. Knowing that before an editor or client runs the scan is just professional hygiene.

A voice spec is a design deliverable

Here is a practice worth stealing from the brand-guidelines world: write a voice specimen for your own studio, with the same seriousness you would bring to a type spec.

Not the vague adjectives every brand doc lists (“bold, human, approachable”) but operational rules a model can be prompted with and a human can edit against. We write short. We name real constraints. We never say “elevate,” “seamless,” or “journey.” We open case studies with the problem, not the deliverables. We admit what did not work. Two or three of your actual paragraphs, marked up with why they sound like you.

A spec like that changes the economics of AI drafting entirely. The model stops producing the internet’s average and starts producing raw material in roughly your register, which cuts the humanizing and editing pass from a rewrite to a tune. It is the difference between correcting a proof and redrawing the face.

The spec earns its keep beyond the marketing site, too. Microcopy, error states, onboarding emails, the empty-state one-liner: these are the text equivalents of punctuation glyphs, tiny surfaces where voice either persists or quietly dies. A studio that specs its voice down to the 404 page is doing for language what it already does for type, carrying the system through the smallest sizes, where craft is hardest to fake and easiest to notice.

It also future-proofs something more important than any single page. Distinctiveness is compounding. Every piece of writing that sounds unmistakably like your studio makes the next one more recognizable, the same way consistent identity work compounds. Every piece of default voice spends that equity down.

Make a choice

The tools are not going anywhere, and they should not. A studio that refuses AI drafting on principle is just paying a speed tax its competitors are not paying. But the studios that will look smart in five years are the ones treating generated text the way they already treat every other default: as a starting point that no self-respecting designer would ever ship untouched.

You have a typeface policy. You have a color system. You proof everything twice before it goes to print. Extend the same standard to the words. Because the audience that matters, clients, editors, peers, and increasingly the classifiers they run, can all tell the difference between a voice somebody designed and a default nobody bothered to change.

System font is fine for a terminal window. It has no business on your specimen page.