Building a rigorous Brand Voice Framework takes time, and most organisations stop short. They settle for tone, a list of banned words and a few signature phrases. Three pillars, four at best.
Our Brand Voice Framework covers nine: the mission statement, values, positive traits, traits to avoid (the ones the brand refuses to be associated with), personality, archetypes, the emotions to evoke, narratives, and finally tone and language.
Why a nine-pillar framework is still incomplete
Each pillar documents a distinct dimension of brand identity. Together they form a system rather than a list.
That depth is rare, and it is necessary. It is still not sufficient.
No tenth pillar would close the gap. Something cross-cutting remains undocumented in almost every framework in use, the strongest included: the way the brand reasons. What a brand says and how it says it are usually covered. The intellectual route by which it reaches its conclusions is not.
That route is precisely what LLMs standardise first. Even a nine-pillar Brand Voice Framework does not protect it explicitly unless it documents the reasoning arcs that run through each pillar and set it in motion.
What AI does to the nine pillars when the arcs are not documented
The trouble with a well-built nine-pillar framework is that it creates an illusion of protection.
The fault does not lie with the framework. Its very solidity is the trap AI exploits: the surface holds, so nobody on the team sees the drift coming.
Three researchers at the University of Southern California (Sourati, Ziabari and Dehghani) set out the mechanism in a review published in 2026 in Trends in Cognitive Sciences. LLMs reproduce the dominant styles in their training data and marginalise other ways of reasoning. Reinforcement learning, which tunes a model to human preferences, reduces stylistic variability. Chain-of-thought reasoning, optimised for linear and explicit inference, discourages more intuitive styles, which are also harder to model. The review reports a telling case: on a classification task where a few examples broke the rule, step-by-step reasoning made GPT-4o four times slower to learn the correct answers. The model generalised the rule and overlooked the exceptions.
In practice, pillar by pillar:
- Mission statement: AI can rephrase your statement in a smoother, more “reasonable” register. It keeps the skeleton and lets the density evaporate. No alarm goes off, because the result is formally correct, perfectly readable and cognitively interchangeable with your competitors’ statements.
- Values: it mentions them without embodying them. It produces texts about values rather than texts written from them. The difference is invisible on the surface and structural underneath.
- Positive traits and traits to avoid: it knows your prohibitions if you hand them over. Without understanding why certain postures are excluded (the world view, the relationship to expertise, the relational ethic they imply), it sidesteps them by accident, cleanly, without tripping a single stylistic alert.
- Personality: it can reproduce the markers of a personality. It cannot reproduce its internal coherence, the way a character trait changes the reasoning as well as the tone.
- Archetypes: the Sage and the Creator are categories it knows. It can name, describe and illustrate them. What it cannot do is reason from the archetype. Thinking like a Sage means building an argument by progressive revelation rather than by linear demonstration. AI produces the form and misses the internal logic.
- Emotions: it leans towards the positive, the reassuring, the motivating. Less standard emotions, such as brutal lucidity, productive discomfort or tension held open, are structurally under-represented in its output. The smoothing is mechanical, not intentional. It has a counterpart on the reading side: Haller, Bolliger and Jäger (ACL 2024) show that the predictions of language models match the reading times of readers with lower verbal intelligence far better than those of the highest performers.
- Narratives: this is the most exposed pillar, and the hardest to defend, because degradation there is slowest and least visible. A narrative of meaning cannot be reduced to a formulation. It rests on a specific narrative logic: a way of building tension, of setting up the founding conflict, of leaving unresolved what should stay unresolved. Without that logic documented, AI produces narratives that are formally correct, emotionally reassuring and intellectually empty. The expert reader senses it, says nothing, and does not come back.
- Tone and language: this is the pillar AI handles best. It can reproduce a register with high apparent fidelity, which is exactly what makes it misleading: the surface holds while the substance drifts.
The four reasoning arcs: the missing cross-cutting layer
Reasoning arcs are not a tenth pillar. They run across the nine existing pillars and set each of them in motion. Documenting the arcs means documenting the internal logic that keeps every pillar working once AI is in the loop.
Arc 1. Analogical reasoning
It transposes a logic from one domain to another to produce an idea nobody expected to find there. It is rigorous rather than decorative, because it carries a demonstration instead of an image.
Link with the pillars: it directly activates archetypes (the Sage reveals by analogy, not by catalogue) and narratives (meaning is often built by comparing what is with what could be in another system).
What AI does with it: it produces generic, overused analogies. “Like the conductor of an orchestra”, “like the foundations of a house”: images that signal the form without carrying the substance.
What your framework must document: the analogical domains that constitute your voice. For WeAreTheWords, these are architecture, engineering, operating systems and infrastructure. They signal a positioning, and they are not interchangeable.
Arc 2. Reasoning through tension
It starts from a real contradiction, two truths valid at the same time, and builds a position without trying to reconcile them neatly. It lives inside the paradox.
Link with the pillars: it sits at the heart of traits to avoid (refusing association with the smooth posture that resolves everything) and of emotions (productive discomfort, lucidity without anaesthetic).
What AI does with it: it resolves tensions. Trained to produce coherent, balanced answers, it turns a productive paradox into comfortable nuance.
What your framework must document: the fundamental tensions your brand holds without resolving them. In a mid-sized industrial company, the tension between “producing faster with AI” and “not hallucinating technical rigour in a compliance file” is a business risk before it is an editorial nuance. A framework that leaves that specific tension undocumented lets AI settle it on your behalf, cleanly and quietly, with nobody noticing that the decision has already been made.
Arc 3. Contextual, situated reasoning
It starts from concrete experience in a specific setting and draws a more general rule from it. It works bottom-up, the reverse of generic top-down reasoning.
Link with the pillars: it anchors values in reality (they are demonstrated from experience rather than declared) and activates narratives (the brand’s role is illustrated through precise client configurations, not through abstract principles).
What AI does with it: it starts from general principles and works down to illustrative examples. It has never worked in your sector and cannot simulate your real frictions.
What your framework must document: your brand’s reference settings, meaning its sectors, types of client, observed configurations and recurring frictions. For a marketing department in the energy sector, that is the approval cycle with five stakeholders, the subject-matter expert available two hours a month, the technical content that has to convince the regulator and the purchasing committee at once. These configurations appear in no training corpus. They are your structural advantage, and left undocumented, they vanish the moment AI enters the loop.
Arc 4. Reasoning by deliberate divergence
It questions the dominant frame explicitly. It names what everyone presupposes without saying so and offers an angle that consensual thinking does not generate.
Link with the pillars: this is the arc most directly tied to the mission statement (a stance always diverges from something) and to traits to avoid (refusing association with soft consensus means choosing divergence as a posture).
What AI does with it: it reproduces dominant positions. It is trained not to offend, to validate consensus and to avoid divisive stances. Deliberate divergence is, by definition, what it does least well.
What your framework must document: the market beliefs your brand explicitly contests, and the proprietary formulations of those counter-positions. They are your strongest markers of authority, and the most exposed to dilution.
How to integrate the arcs into your existing framework
No rebuild is needed: three targeted operations on what already exists.
1. Activation review
For each of the nine pillars, identify which reasoning arc activates it in your best historical content. A pillar without an identifiable arc is exposed to homogenisation.
2. Documenting the invariants
For each arc, set down three to five examples drawn from your founding content, the analogical domains and reference settings that constitute your voice, the tensions you hold and the positions you contest. This work cannot be delegated to a model. It is human editorial work, which is why it produces lasting value.
3. Rules of engagement in the workflow
Define where AI may come in, arc by arc.
- Factual rephrasing of situated reasoning already framed by a human: can be delegated.
- Constructing a deliberate divergence: cannot be delegated.
- Densifying an analogy already chosen: can be delegated.
- Choosing the analogy itself: cannot be delegated.
The multidimensional check prompt
Run it as a systematic check before publishing any content with a high symbolic charge.
Act as a reasoning-diversity controller for [BRAND NAME], working from its documented reasoning arcs.
Analyse this text on five axes, scoring each from 1 (heavily smoothed by AI) to 5 (strongly distinctive):
- Rhythmic variability (a mix of short and long sentences, deliberate breaks)
- Lexical dispersion (key terms maintained, registers mixed)
- Emotional richness (non-standard emotions present, tensions left unresolved)
- Contextual anchoring (precise sector references, situated reasoning)
- Originality of the arc (an identifiable analogical, tension-based, situated or divergent mode)
For each axis, justify the score and propose a targeted revision if it falls below 4.
Verdict: “Acceptable for publication by [BRAND NAME]” or “Needs rework: excessive homogenisation”, with rewriting priorities.
The prompt is an instrument of cognitive governance: it brings the arcs into daily work.
What this changes for your editorial system
A nine-pillar Brand Voice Framework with documented reasoning arcs answers the question conventional frameworks leave open: what in our voice cannot be delegated to AI, and why?
Tone can be imitated, and a proprietary lexicon can be reproduced. What cannot be delegated is the forms of inference that make up your intellectual identity. They are the product of your history, of the archetypes you genuinely embody, of narratives built from real frictions. No training corpus contains them.
Documenting the arcs does not guarantee distinctiveness. It guarantees that you know where distinctiveness has to be protected. That is a different thing, and already a great deal.
Frequently asked questions about reasoning arcs
What is a reasoning arc?
A reasoning arc is the intellectual route by which a brand arrives at what it says. It describes the shape of the brand’s inference, upstream of tone and vocabulary. We distinguish four: analogical reasoning, reasoning through tension, contextual and situated reasoning, and reasoning by deliberate divergence. They run through the nine pillars of the Brand Voice Framework rather than adding to them.
Why does AI threaten a brand’s way of reasoning first?
Because no review against the voice charter checks it. Models reproduce the dominant styles in their data, favour linear inference and resolve tensions instead of holding them. Tone and vocabulary can be checked line by line; a way of reasoning cannot. A generated text can therefore respect the whole charter and reason exactly like a competitor’s without anyone noticing.
Which tasks can be handed to AI once the arcs are documented?
Those that work on reasoning a human has already framed. AI can rephrase a factual passage drawn from documented experience, or densify an analogy already chosen. Choosing the analogy, constructing a deliberate divergence or deciding which tension the brand holds remains human work, because these are precisely the operations models perform least well.
→ Want to build this dimension into your Brand Voice Framework? See the Clarity engagement
Sources
Sourati, Z., Ziabari, A. S. & Dehghani, M. (2026). The homogenizing effect of large language models on human expression and thought. Trends in Cognitive Sciences. → https://www.sciencedirect.com/science/article/pii/S1364661326000033 (preprint: https://arxiv.org/abs/2508.01491
Santurkar, S. et al. (2023). Whose Opinions Do Language Models Reflect? ICML 2023. → https://arxiv.org/abs/2303.17548
Haller, P., Bolliger, L. S. & Jäger, L. A. (2024). Language models emulate certain cognitive profiles: An investigation of how predictability measures interact with individual differences. Findings of ACL 2024. → https://aclanthology.org/2024.findings-acl.469/