Verbal Identity Design in the Age of LLMs
June 19, 2026
What Does It Mean to Have a Brand Voice When Any Language Model Can Imitate It?
A brand manager at a mid-sized consumer goods company sits down on a Tuesday afternoon and does something that would have seemed impossible three years ago.
She opens a new chat window, pastes her company's fifteen-page voice guide, adds a brief about an upcoming product launch, and hits send.
The response comes back in twelve seconds. Clean. Polished. The adjectives are right. The rhythm is right. The personality markers are exactly where the verbal identity design guide said to put them. Not close. Correct.
She reads it once. Twice. Then she does something interesting. She does not celebrate. She does not forward it to her creative director. She stares at it for a long moment and feels something she cannot quite name.
The feeling is not about quality. The output is good.
The feeling is about what quality proves now.
When any language model can generate passable copy in any brand's tone, verbal identity design stops being a product and starts being a proof. The question of what a brand voice is has been permanently altered by what can now convincingly imitate it.
That is what this piece is about.
In this piece:
1. Verbal Identity Design Used to Prove Craft. Now It Must Prove Position.
2. What LLMs Can Generate vs. What They Cannot Originate
3. The Three Components of a Voice AI Cannot Fake
4. Three Brand Voices That Prove the Point
5. What Verbal Identity Design Looks Like Now
6. Frequently Asked Questions
7. The Limit of Language Is the Limit of the World
Verbal Identity Design Used to Prove Craft. Now It Must Prove Position.
For roughly two decades, the brief for verbal identity design was a brief about capability.
Brands needed consistency across channels. They needed a tone that could be trained across teams. They needed copy that scaled without losing its personality when it moved from a homepage to a push notification to a customer service script. This work required specialist agencies because competent brand language was genuinely scarce. Most marketing teams could not write well at volume, and most could not write in a unified voice across formats.
The value proposition was essentially: we can write, consistently, in a voice that sounds like you.
That brief is over.
Language models can now produce vast amounts of consistent, on-brand copy across every format imaginable, given nothing more than a style guide and a prompt. The range of linguistic adaptation. The ability to shift register while staying in-voice. The generation of twenty variants of a headline in four seconds. All of it automated.
What this collapsed was the industry's implicit argument that verbal identity was scarce because making it was hard.
Here's the thing: it was never the right argument to make. It just happened to work, for a while, because the alternative was too expensive or too slow. Now the alternative is free. Which means the discipline has to argue from a different foundation entirely.
The value is no longer "we can write in your voice."
It is: "we can determine what your voice is actually saying. And whether that matters".
That is a much harder brief. And it is also the only brief that survives.
The Original Brief: Consistency, Tone, Scale
What specialist verbal identity consultancies were built to solve was, fundamentally, a polyphony problem. Many writers. Many channels. No coherence. The London International Awards did not even have a verbal identity category until 2016. Before that, the discipline existed but had no formal stage. The solution was documentation: voice guides, tone maps, vocabulary lists. It worked because the bottleneck was craft. Most teams could not write well at scale.
Verbal Identity Ltd., the London specialist consultancy that has been running brand language programs since the early 2010s, describes this founding mission plainly: "Cut the crap and cut through. Align all your writers in one distinctive brand voice." That pitch made sense when the alternative was inconsistency born from human limitation.
The limitation is gone. The pitch needs to change.
What AI Commoditised. And What It Did Not.
LLMs removed the craft barrier. The range of linguistic adaptation, the ability to shift register while staying in-voice, the generation of variants at volume, all of it automated. The industry's implicit argument collapsed.
Column Five Media articulated the stakes in a 2026 article on brand voice: "Brand voice is verbal architecture. A list of adjectives on a slide deck does not scale. A documented, operational voice system does."
The sentence that matters is the first one. Verbal architecture. Not verbal decoration. Not verbal consistency. Architecture.
An architecture requires a load-bearing argument. You can replicate a building's surface. You cannot replicate its structure without understanding what it is holding up.
AtomWriter made the limitation explicit: "A brand that has operated in a niche for twenty years writes differently not because someone chose different words, but because twenty years of context is embedded in the perspective. AI trained on public data cannot replicate proprietary experience."
That is the distinction. And everything that follows is an unpacking of it.
What changed. What did not.
If your brand is navigating this shift, our brand strategy and verbal identity work starts with exactly this question.
What LLMs Can Generate vs. What They Cannot Originate
Here is the critical distinction, and it requires being precise.
LLMs are pattern completers. They learn statistical regularities across enormous amounts of language and produce continuations that match those patterns. Feed them enough of a brand's copy, and they can approximate its surface features with remarkable accuracy. The rhythm. The word choice. The personality markers. The sentence structure.
This is impressive. It is also not voice.
A 2025 academic paper published on arXiv (arXiv:2509.14543v1) tested whether language models could convincingly imitate the implicit writing styles of individual authors. The finding was direct: "while LLMs can partially emulate user style in more structured formats like news and email, they struggle with nuanced, informal expression in domains such as blogs and forums. Generated outputs often default to an average, generic tone and remain readily detectable as AI-written."
Default to an average.
That phrase is worth holding. LLMs pattern-match to the centre of the distribution. They produce the statistical average of what similar language sounds like. Distinctive brand voice, by definition, does not live at the centre of the distribution. It lives at the edge. Often a very specific edge, built over years of consistent, opinionated choices.
The Grain: What Barthes Called the Irreducible Remainder
Roland Barthes made a useful distinction in his 1972 essay "The Grain of the Voice," later collected in Image-Music-Text (Hill and Wang, 1977). He was writing about music, but the structure applies here.
Barthes distinguished between the pheno-song and the geno-song. The pheno-song is the communicative content of a performance: the notes, the lyrics, the technical execution. The geno-song is a bodily, material quality that exceeds the linguistic sphere entirely. It is what no score can capture. The grain that emerges in the friction between language and the body that produced it.
LLMs are excellent at the pheno-song.
The rhythm of a brand's copy. The diction. The tonal register. The personality markers in the style guide. All of this is pheno-song. It can be learned, approximated, and reproduced.
The geno-song is the position the voice expresses. The commitment behind the words. The refusal, earned over time, to sound like everyone else. This cannot be extracted from examples because it was never in the examples to begin with. It was in the years of decisions that preceded them.
Point of View Is Not a Linguistic Property
Voice is often described as a list of adjectives. Warm. Direct. Irreverent. Authentic. These adjectives are imitable precisely because they are linguistic. They can be fed into a prompt and returned, competently, in twelve seconds.
What is not imitable is the specific, defensible worldview from which a genuine brand speaks. A view of the world that could be wrong. That someone else could argue against. That the brand has staked something on.
LLMs cannot hold a position because they have no position. They pattern-match to the centre of the distribution. Distinctive brand voice lives at the edge.
Ludwig Wittgenstein wrote in the Tractatus Logico-Philosophicus in 1921: "The limits of my language mean the limits of my world."
A brand whose language is borrowed operates within someone else's limits. Its language cannot take it somewhere that language has not been before, because it is not that brand's language. It is the statistical average of everyone's.
Lera Boroditsky's research, summarised in Scientific American in 2011, showed that the vocabulary available to a speaker shapes what they are capable of perceiving and arguing. Apply that to brand: a brand whose vocabulary is generic is constrained to generic thought. The language we use determines what we can say next. Borrowed language can only make borrowed arguments.
The Three Components of a Voice AI Cannot Fake
There are three properties of a genuinely distinct verbal identity. Each one is beyond what a language model can produce, not because the models are technically limited, but because these properties are not linguistic. They are ontological. They precede language. Language can only express them once they exist.
A Specific Point of View: Not Adjectives, but Arguments
Ask almost any brand to describe its voice, and you will get adjectives.
Warm. Direct. Human. Authentic. Bold.
These are not voices. They are descriptions of voices. The problem is that every adjective in that list is infinitely imitable. Feed an LLM those five words and tell it to write a headline. It will. It will also produce the same headline for your twelve closest competitors if you feed them the same five words.
What is not imitable is a specific, defensible claim about the world. A view that could be wrong. A position that someone else could argue against.
Oatly does not say "friendly." It says the dairy industry has a structural problem, this product is the honest corrective, and the entire packaging is evidence for the case. John Schoolcraft, who took over as creative director in 2013, did not start with a voice brief. He started from the refusal of a default. In a Fortune interview, he described the brand before his arrival as "uninspiring and boring. It looked like everything else." The voice that followed was the linguistic expression of an argument that already existed.
Liquid Death does not say "irreverent." It says that beverage marketing has been governed, for decades, by rules that entertainment brands were never subject to, and that this is an absurdity worth exposing. Mike Cessario, who founded the company after years in advertising, stated the logic plainly: "The only way the brand would have a chance at survival is that the actual product itself has to be so insanely interesting, where so much of the marketing is baked into the product." That worldview existed before a single word of copy was written.
To test whether a brand has a genuine point of view: can it be disagreed with? If the answer is no, it is an adjective. And adjectives, at this point, are free.
Earned Vocabulary: Words the Brand Has Changed
Every genuinely distinct voice has a proprietary vocabulary.
Sometimes the brand coined the words. Sometimes it inherited them from a subculture. Sometimes it has deployed a common word so consistently, so specifically, and with so much accumulated context, that the word now carries the brand's meaning alongside its dictionary meaning.
Starbucks' size names. Oatly's "wow," deployed with a particular ironic sincerity that took years to develop. The Innocent bottle's "I feel a bit dotty," written in the first person by a smoothie, in 1999, before that register existed in FMCG.
Byron Sharp's research, published as How Brands Grow (Oxford University Press, 2010), argues that distinctive brand assets, including verbal ones like taglines and brand vocabulary, build memory structures that drive mental availability at the moment of purchase. The argument is about distinctiveness, not differentiation. What the brand owns in the mind of the consumer is not a feature set. It is a set of retrieval cues. Verbal ones compound over time.
No language model can reproduce the accumulated context that makes a word distinctively theirs. It can replicate the word. It cannot replicate the years of specific, consistent use that give the word weight. The Mailchimp style guide describes their tonal position as "We prefer the subtle over the noisy, the wry over the farcical." Every word in that sentence has been earned. The guide can be read by an LLM. It cannot be felt by one.
Structural Refusal: The Voice Defined by What It Will Not Do
The third component is the one most often missed.
Every genuinely distinctive voice is defined as much by its negations as by its properties. What the brand refuses to sound like shapes every piece of copy as powerfully as any style rule.
Innocent will not use corporate language. Liquid Death will not use health-food sincerity. Oatly will not pretend to be a multinational. These are not stylistic preferences. They are structural commitments. And they are not imitable because LLMs have no ontological commitments. They will produce the voice and its precise opposite with equal fluency if prompted.
The refusal is a choice. A choice is an act of will. And that is something an AI cannot perform.
Jon Wright, co-founder of Innocent Drinks, described the origin of the brand's label copy in an interview: "All of the large companies were adhering to the rules and regulations with glossy slogans on the front. There was just a big blank space so we thought we ought to think of something to put there."
The structural refusal was there before the voice was. The voice was just what happened when you filled the blank space with honesty instead of convention.
Paul Burke, the copywriter credited with originating the Innocent voice concept, later reflected on what it had produced. His structural refusal of CPG formality was so complete that when Guardian journalist Rebecca Nicholson named the resulting phenomenon in 2011, she coined a new word for it: "wackaging." The term was hers. The style was Innocent's. The refusal was structural and preceded both.
Three Brand Voices That Prove the Point
Every principle above becomes more legible when you trace it backwards through specific brands. Three cases, chosen because each proves a different dimension of the argument. And because all three built their voice before writing a single document that described it.
Oatly: The Voice Came from an Argument, Not a Brief
Oatly was founded in Sweden in the early 1990s. By 2013, the product had been on shelves for two decades. The packaging was, in Schoolcraft's own description, "uninspiring and boring. It looked like everything else."
What followed is one of the most instructive brand transformations of the past decade, not because it produced great copy, but because of the sequence. Schoolcraft and CEO Toni Petersson did not commission a voice guide and then brief a designer. They started from a position: the dairy industry has a structural problem, and Oatly is the honest alternative. The voice came from that argument.
The packaging carries long-form copy because the brand has things to say that require space. "IT'S LIKE MILK BUT MADE FOR HUMANS" is not a tone choice. It is an argument that happens to be funny.
Oatly even demonstrated the structural refusal principle in real time. When the brand temporarily paused selling oat drink, their website read: "After momentarily losing all interest in selling you oat drink products, we decided to do this instead." No brand whose voice comes from a style guide produces that line. Only a brand whose voice comes from a position can afford to.
Under Schoolcraft's creative leadership, Oatly went from an obscure Swedish oat drink to a $10 billion IPO valuation. The Ministry For Mindcontrol, Oatly's internal creative department, has been described as a "global masterclass in brand-led business transformation." The masterclass is not about the copy. It is about what produced it.
Liquid Death: Voice as Cultural Transcript
Mike Cessario came from advertising. He was also, by his own account, embedded in underground music culture. In 2017 he noticed something: the beverage category was following a set of rules that entertainment brands had never been subject to. Everything was either aspirational wellness imagery or corporate sterility.
He did not write a brand voice guide. He asked a simpler question: what would this category look like if it were treated as entertainment?
The answer was Liquid Death. Heavy metal aesthetics. Death imagery. The slogan "Murder Your Thirst." The voice was not designed to be funny, though it often is. It was designed to be a perfect linguistic rendering of an existing subculture applied to a category where that subculture had never appeared before.
The voice is not imitable because the subculture it transcribes pre-exists it. You could generate copy "in the style of Liquid Death" and it would feel like a Halloween costume. The real thing has roots.
Cessario was explicit about the cultural source in a podcast interview: the brand drew directly from "death metal and punk rock culture." The subculture preceded the brand. The brand translated it. That is a different kind of vocabulary than the kind a style guide can describe.
From $2.8 million in sales in 2019 to $130 million by 2022. Not because the copy was different. Because the position behind the copy was.
Innocent Drinks: The Voice No One Knew They Were Building
In 1999, Richard Reed, Adam Balon, and Jon Wright had no copywriting agency. No verbal identity brief. No tone-of-voice document. They sold smoothies from a stall at a music festival. They designed their first label by sitting in front of a computer, writing the legal requirements, and staring at the blank space that remained.
They filled it the way they talked to everyone. With fewer swear words, reportedly.
The voice that resulted, conversational, confessional, gently absurdist, became the template for an entire category of consumer packaging. So distinctive and so widely imitated that in 2011, Guardian journalist Rebecca Nicholson created a Tumblr to document the phenomenon and coined a word for it: "wackaging." The tagline on the Tumblr was "I blame Innocent smoothies."
The founders created the style. A journalist had to invent a word for what they had started.
That gap, between doing something genuinely and having others name what you did, is the gap that verbal identity design is trying to create. It is not created by a document.
Three founders who genuinely believed health food did not have to be humourless, and who had nothing to prove to a marketing department, produced what no brief could have commissioned.
The voice survived Coca-Cola's 90% acquisition in 2013, after which the founders stepped down. A worldview, properly embedded, outlasts its owners.
Where the voice came from.
What Verbal Identity Design Looks Like Now
If the old brief was "write in our voice," the new brief is harder and more interesting.
It is: find out what we actually have to say. Build a system around that. Create the conditions for a voice to compound over time rather than flatten into consistency.
This is what a genuine verbal identity design agency does differently in a world where language models can execute competent copy on demand. Not write better. Think harder about what the brand is actually saying, and whether that argument can be held.
From Adjectives to Arguments: The Discovery Process
Most verbal identity briefs produce adjectives. Genuine verbal identity design starts by questioning them.
What does "authentic" actually mean for this brand, in this category, at this moment? What does "direct" look like when the product is complex? What does "warm" mean when the brand operates in a category where warmth is the default register of every competitor?
Column Five Media put the test plainly: "If your audience cannot tell who is speaking when the logo is removed, the voice is not doing its job. Most definitions stop there. They describe brand voice as a set of adjectives and move on to examples. That framing misses the point."
The discovery process, properly executed, is an excavation. Competitor voice mapping to find the category's default register. A language audit of the brand's existing communications to find what is genuinely distinctive versus what is merely consistent. Founder interviews, not to understand personality, but to find the argument: what does this brand believe that others in the category do not?
The deliverable is not a list of adjectives. It is a position.
Mailchimp's voice guide, one of the most detailed public examples in the industry, demonstrates this. The guide does not open with "Mailchimp is friendly and approachable." It opens with a specific framing of the brand's relationship with its users: "the experienced and compassionate business partner we wish we'd had way back when." That is not an adjective. That is an argument. A reader who encounters that line understands not just how Mailchimp sounds, but what it believes about its users and what it thinks good communication does.
This is what the brand's position looks like written down. And it is also the step that cannot be outsourced to a model.
The Vocabulary Audit: What the Brand Already Owns
A vocabulary audit is not about identifying the brand's favourite words.
It is about identifying the words the brand uses differently than anyone else uses them. Where usage has accumulated enough context to be genuinely proprietary.
Byron Sharp's research in How Brands Grow (2010) provides the conceptual anchor. Sharp argues that distinctive sensory and semantic cues, including verbal ones like taglines and brand vocabulary, build memory structures that drive mental availability at the moment of purchase. The vocabulary audit finds what has already been built, often without anyone knowing it.
This is also the step LLMs cannot perform. It requires reading the history of a brand's language across time and finding what has compounded. Not what appears frequently. What appears distinctively.
The commercial case for getting this right: Lucidpress surveyed over 400 brand management experts in 2019 and found that consistent brand presentation increases revenue by an average of 10 to 20 percent, with the upper bound reaching 33 percent. The mechanism is not mysterious. Consistent brands require fewer touchpoints to convert. They price at a premium. They generate higher lifetime value. Verbal consistency is a component of all of this, more so now than before, because it is now the hardest component to replicate.
The Living System vs. the Style Guide
Style guides date. Voice systems compound.
A style guide is a document. The moment it is written, it begins to lag the brand's actual language. New platforms emerge. New tones appear in the category. New writers interpret the guide differently.
A voice system built around a position is self-reinforcing. Because every new piece of copy can be tested against the argument. Does this sound like a brand that believes what we believe? Does this carry the structural refusal we have committed to? Does this add something to the vocabulary we are trying to build?
The test of a genuine verbal identity is not whether it survives a new copywriter. It is whether it survives a new platform, a crisis, a pivot, an acquisition.
Innocent's voice survived Coca-Cola acquiring a 90% stake. The founders stepped down. The copy team changed. And yet the brand continued to sound like itself because the voice was rooted in a position, not in a document.
Oatly went public in May 2021 at approximately $10 billion valuation. A public listing is often the moment a challenger brand's voice flattens into investor relations language. Oatly's did not. Because the Ministry For Mindcontrol is not just a creative department. It is a structural commitment to an argument.
What the document describes, AI can reproduce. What the position requires, humans must continue to inhabit.
We have written separately about where this exact dynamic applies to visual identity as AI-generated design scales, and the stakes are identical.
This is what makes verbal identity design different from content strategy, different from copywriting, and different from brand guidelines. And it is also why understanding how AI changes this discipline requires understanding what AI actually exposes about brand thinking more broadly, something we have written about in our piece on what AI reveals about brand design thinking. For the semiotic substrate on which voice sits, the companion piece on how brands carry meaning before they say a word takes the argument further.
Frequently Asked Questions About Verbal Identity Design
What does a verbal identity design agency actually do?
A verbal identity design agency establishes what a brand has to say, not just how to say it. The deliverables include a voice architecture, a messaging hierarchy, a vocabulary system, and tone modulation guidelines for different contexts. The most important deliverable, though, is the discovery process that precedes all of them.
Can AI replace a verbal identity design agency?
AI can execute within a defined voice system. It cannot establish one. Establishing a voice requires finding the brand’s argument, which requires judgment about what is true, defensible, and genuinely distinct in the category.
How do you build a brand voice that AI cannot imitate?
Build from position, not adjectives. Identify the specific, arguable claim your brand makes about the world, then build vocabulary around it. Establish the structural refusal: what your voice refuses to sound like shapes it as powerfully as what it sounds like. Voice built on a genuine worldview is not more complex. It is more rooted.
What is the difference between brand voice and tone of voice?
Voice is the consistent personality that holds across every context. Tone adapts within that personality depending on the situation. Voice does not change. Tone shifts. What makes this distinction matter more now is that AI is very good at modulating tone. Given clear guidelines, a model can shift from warm to formal to playful within the same voice. What it cannot do is generate the voice itself.
The Limit of Language Is the Limit of the World
Wittgenstein wrote: "The limits of my language mean the limits of my world."
He was making a philosophical claim about the relationship between language and reality. Apply it to brand, and it holds.
A brand whose language is borrowed, from a competitor's framing, from a default LLM output, from an adjective list built to satisfy a stakeholder review, does not just sound like everyone else. It thinks like everyone else. It is constrained to the arguments that borrowed language can make. And borrowed language can only make borrowed arguments.
That is the whole problem.
The machine can say anything. That is, at this point, not an achievement. What the machine cannot do is mean something specific, because meaning something specific requires having a position. A position requires judgment. Judgment requires the accumulation of experience, refusal, and commitment over time.
Brand voice, properly built, is the record of those commitments.
The brand manager who stared at that LLM output on a Tuesday afternoon was not afraid of the quality. She was afraid of something more precise: that if the copy was this good, and it had not come from anywhere, it did not prove anything. And a brand whose copy proves nothing is a brand with nothing to say.
The question is not whether you can sound like yourself.
Any model can sound like you.
The question is whether there is a self to sound like.
