The Architecture of Voice
What Ariel and Ursula tell us about AI and ‘Voice’; and why a triangle and a circle tell even more!
This piece is the fourth in a five-article arc on AI, language, and what survives of you in your writing and other works:
Why the F-bomb is the most adaptable word in the English language…
The Architecture of Voice (this article)
The Ouroboros Closes (forthcoming)
I think you know, by now, that I love a story as an analogy for an important idea. Here’s one of my favorites and why it’s relevant to what’s worth preserving in cultural and professional content produced with AI tools.
Ariel gave up her Voice to Ursula.
That’s the part of the story everyone remembers. The mermaid trades her voice to the sea witch in exchange for legs and a chance of love with Prince Eric. But Ursula tries to use Ariels voice to steal Eric away.
You see, Ursula just puts on the voice. She wears it like a costume, strolls into the Prince’s life, and tries to bewitch him with it. The voice is ‘technically’ intact. The pitch is right. The timbre is right. The notes are perfectly placed. By every measurable property, the voice is that of Eric’s true love.
But Eric resists, and at first isn’t entirely clear why. He can’t quite say what is wrong. The voice sounds right. But something is off.
What Ursula has, you see, is the artifact. What she does not have is the grounding of who the person Ariel is.
You see, the voice was never just the sound waves, the timbre. It was the audible expression of who Ariel was, what she had seen, her longings, her passions, what she had risked, and her lived experience.
Strip the voice from the person, and what you have is a recording. A very good recording…but not the person.
In case anyone was still left wondering about this analogy, the Voice is yours; Ursula in this case is AI tools like ChatGPT, Claude etc; and Eric is your audience, your readers, your customers.
Just hold onto the metaphor for the rest of this article.
Why the standard model of competence is wrong
For most of my working life, I have watched organizations assess capability the same way. They measure what a person knows and what a person can do. Knowledge plus Skills. Two corners of a shape that nobody bothers to draw, because the shape is assumed as consisting of just those two elements.
That model is broken.
It is broken because it produces interchangeable workers, who ship interchangeable artifacts, into environments where the actual differentiator was always something else that the HR job role assessment and appraisal system never measured.
I saw this early in my nursing career in the NHS, where the textbook or procedure manual would tell you one thing, and the Charge Nurse of 20 years would teach another three better ways…and why.
In the years since, I’ve repeatedly seen the same thing on the commercial side of healthcare IT, where the difference between a competent commercial leader and a great one is almost never the content of their LinkedIn ‘skills’ list.
I have written previously about the Competence Framework, which I keep returning to because it is such an elegant mental model. In image form it consists of a triangle located inside a circle.
The Competence Framework
The triangle has three sides.
Knowledge: what you know.
Skills: what you can do (or tools you have).
Experience: what you have actually done, in specific environments, with specific stakes, with specific consequences, and with specific people, who you remember.
The three sides do not substitute for each other. A person with Knowledge and Skills but no Experience is, in a very real sense, untested. They own a map but they have not walked the terrain. The terrain (Experience) is the part that actually matters.
The circle around the triangle represents Context.
Context is the cultural, situational, relational, historical, and personal grounding that determines which Knowledge applies, which Skills are appropriate, and which Experience is relevant to the moment in front of you.
Context is accumulated over time by being exposed to, testing, and applying your Knowledge, Skills and Experience in different ways.
Context is your wins and losses, your lessons learned, and your scars gained.
Context is the thing that lets you ‘read the room’ before anybody has spoken.
Voice is Context.
In using AI to assist with writing and almost anything creative, Voice is not a style or preference switch, and it is definitely not a personality setting. Voice is the actual expression of accumulated Context. It is expressed through the choices a writer makes about which thoughts to surface and in what order to express the idea and/or emotions they are trying to communicate. It is how they flow their argument, with what emphasis, and what cultural backdrop they wish to draw from.
LLM and AI tooling can mimic style. It can record and produce a repeated preference. It can even approximate a personality of sorts (although that is a whole other conversation.)
What the AI can’t produce is Voice, because the model has no Context. It can never read the room!
Where the AI LLM fits in the Competence Framework
The AI LLM tooling almost fully occupies the Skills element. It knows a version of the rules. It can produce fluent prose. It punctuates correctly(ish). It structures arguments. It generates code that compiles. It drafts documents that read well at a glance. That visible output of one element of competence is, in fact, what the model is best at producing.
The model can also assist with Knowledge. It can surface references, summarize and compress large bodies of material into navigable form, and answer factual questions with (reasonable) accuracy. It is genuinely useful.
But it is not a substitute for actually knowing the ‘thing’, because what it has is a statistical compression of what other people have written about the ‘thing’, and that statistical association is not the same as ‘THE thing’. But it still remains useful nonetheless.
This is exactly the corner that the accusation culture I described in the first article in this series inhabits. The reader who suspects AI authorship is pattern-matching against the implied fluency at the intersection of Knowledge and Skills. They are reading the base of the triangle in the shape, and noticing correctly (just like Prince Eric), that the remainder of the triangle and circle might not be there.
One of the main areas where AI LLM tooling drops the ball is Experience.
There is no algorithmic shortcut to having done the work. Experience is, by definition, the accumulation of time spent in specific environments under specific conditions, with consequences that actually mattered at the time. It’s quite literally the old analogy about apprenticeship. Time served.
The model has ‘read’ about your environment but it has not been there. In real life the clock runs at one second per second, and there is no version of the LLM tool stack that can compress actual experience.
Some will object that the tools can accelerate the acquisition of Experience by giving the user faster access to the knowledge and skill patterns that other practitioners have surfaced. There is a kernel of truth in that assertion, BUT it is a dangerous and slippery belief.
The dangerous part is the assumption that accelerated Knowledge substitutes for Experience, when in fact accelerated Knowledge, and improved Skills untethered from time, is precisely the recipe for the Builder’s Mirage.
The Builder’s Mirage is the illusion of competence, produced without the underlying thing being present.
A document that reads as though a strategist wrote it, but with no strategist behind the keyboard.
A clinical summary that sounds like a clinician produced it, but with no clinical hours with hands on patients behind the words.
A piece of writing that sounds like a writer wrote it, but with sanitized, empty statistically predicted conclusions underneath.
This is the same dynamic I described in the Alignment Files in my discussion about AI generated synthetic user personalities and their use in Product Management. The model produces an approximation of an experienced practitioner by pattern-matching across the recorded outputs of actual practitioners. The approximation reads convincingly, until you put it next to the real thing, at which point the shallow summarization becomes visible to anyone who has done the actual work.
A synthetic strategy report sounds like a strategist, until a strategist reads it; a synthetic clinician sounds like a clinician, until a clinician reads it; a synthetic writer sounds like a writer, until a writer reads it.
Knowledge can be acquired. Skills can be trained. But Experience can only be done. Context can only be lived.
Why Context is the part that cannot be synthesized at all
Knowledge and Skills can accelerate. Experience accumulates at one second per second. Context, the outer circle, is more stubborn than any of the other three.
Context is not a thing the writer acquires through deliberate effort. It is the residue of a life lived.
It is Mrs. Armstrong’s classroom in Ayrshire in 1973, and the way she taught a child to hear a sentence before writing it down. It is the ward round where you learned that the patient’s quietness was the warning sign nobody had named in the textbook. It is the meeting where the data said one thing and the room knew and felt another, and you learned, slowly, to trust the room, because the data was probably incomplete. “Spidey sense” is real, and it lives in Context.
Context is what tells you, when the Skills produces a fluent paragraph, whether the paragraph is right. Not technically and grammatically correct. That it’s right.
The model has no access to this, because the model has not lived. It has read records of other people having lived. The records are not the same as the actual living, and any practitioner who has done the actual work will tell you the same thing in their own vocabulary, because they have all noticed the same gap.
This is why the writer with deep Context can use the AI tools without surrendering to them, and the writer with thin Context struggles to maintain Voice in a rush to be productive.
The AI tools amplify whatever Context the writer brings. If it’s thick and rich, the AI amplification produces useful and sometimes valuable work, and the writer’s Voice survives the process intact. If the Context is thin, the amplification produces fluent, hollow emptiness. It has volume, but with a superficial confidence that fools nobody who has actually been there, done that and got the T-shirt..
The question this leaves open
Ursula couldn’t hold the Prince with Ariel’s voice, because the voice without Ariel was a hollow thing, and Eric could feel it.
Your readers can feel it too in your writing and your output. They might be impressed by a single article you publish that delivers an insight for them. They might nod their head at a several well articulated posts. But eventually, just like Eric, they will sense that something is off.
But if Voice in your writing remains the expression of accumulated Context, and Context cannot be shortcut, then the question is not whether to use the AI LLM tools. I’d argue the real question is how DO we use them without losing the grounding of your Voice?
That is the AI tooling integrator’s question, and it is what I’ll tackle in the fifth and final, article in this short series.
Keep Haverin
SM
This piece is the fourth in a five-article arc on AI, language, and what survives the dominant signal:
Why the F-bomb is the most adaptable word in the English language…
The Architecture of Voice
The Ouroboros Closes (forthcoming)
Stuart Miller is Managing Director of Haverin Consulting LLC, a Scot living in Minnesota, and co-host of the Haverin About podcast with his daughter Bethany Miller-Urroz.
I so enjoyed this piece. There is so much discourse around “voice,” and this is, bar none, the clearest explanation I’ve seen articulated on the subject. There are so many AI camps now that I’m starting to lose track.
I absolutely love the framework you’ve presented. In my own writing, I use four different AIs and generally follow one of two approaches.
Sometimes I write a completely rough draft and then ask each AI to produce its own version. I then consolidate those drafts into a new document and ask each AI to synthesize the strongest elements into another draft. From there, I choose the version I like best and begin refining it: word choice, phrasing, additional examples where needed, until it feels polished to my satisfaction.
Other times, I’ll spend two hours in conversation about the topic I want to write about, exploring ideas, pushing back, and testing alternative perspectives before asking for a draft. If I’m reasonably satisfied, I move into the same refinement process. Before I finish either method, I always ask for blind spots and pushback on the thesis, then address those concerns in some way within the piece.
In the end, I can easily spend 20 hours on a 1,200-word article. And if someone asked me to describe my voice, I’d probably say: “well-articulated ideas,” or at least that’s what I aim for.
Well done, Stuart.
Hi Stuart--am learning so much from you. What I've been reacting to, the falsity of tone and approach, in most of marketing writing, you have formulated its roots via impeccable logic.
I noted your article in my Must Reads roundup for the week in Telehealth & Telecare Aware here: https://telecareaware.com/a-must-read-potpourri-the-math-of-ai-data-center-builds-healthcare-ai-failures-telehealth-in-schools-hippocratic-ais-problems-the-loss-of-empathy/