The Literary DNA: why generic AI erases your firm's style — and how to preserve it

Style Is the Asset

A firm's style has concrete components: the restraint of short, dry filings, versus long doctrinal architecture; opening with the strongest argument or laying the procedural groundwork first; a deferential or assertive tone toward the bench; vocabulary, sentence rhythm, the way precedent gets cited. None of this is decoration. Together, it forms a signature recognizable to judges, clients, and opposing counsel — and that signature carries reputation built over decades.

The problem is where that voice lives: in the heads of a few people. It's tacit knowledge in the sense philosopher Michael Polanyi gave the term in 1966, when he observed that "we can know more than we can tell." The partner who writes brilliantly can almost never state the rules they follow; the young lawyer absorbs them by proximity. When the base of the pyramid narrows and day-to-day contact shrinks — the scenario diagnosed in Edition 2 — the channel through which style used to transmit loses throughput. An asset worth decades of reputation is left exposed to walk out the door at any moment, with no copy left behind.

Stylometry: Measuring What Goes Unsaid

Stylometry is the quantitative study of writing style. Its core insight is counterintuitive: what most distinguishes an author isn't the big, conscious choices, but small, involuntary habits — the typical length of their sentences, the connectives they favor without noticing. These are stable traits, hard to fake. They function like a fingerprint.

The discipline is old and serious. In 1887, physicist Thomas Mendenhall proposed, in the journal Science, characterizing authors by the distribution of their word lengths. In 1964, statisticians Frederick Mosteller and David Wallace resolved the disputed authorship of twelve of The Federalist Papers by analyzing the frequency of function words — articles, prepositions, conjunctions — attributing them to James Madison. In 2005, a group led by Hans van Halteren and Harald Baayen empirically demonstrated the existence of a stylome: each individual's measurable stylistic signature, in a deliberate parallel with the genome. The literary DNA metaphor isn't rhetorical license.

That this signature is identifiable carries weight even in forensics. The most famous case is the Unabomber's: for almost two decades the FBI pursued the perpetrator of a bombing campaign with no fingerprints and no biological DNA; what identified him was language — forensic linguist James Fitzgerald recognized writing patterns in the manifesto that matched old letters. Style is distinctive enough to identify one person among millions.

The Pasteurization of Voice

Language models produce, at every step, the statistically most probable text. That mechanism is insensitive to truth — hence the hallucinations discussed in Edition 3 — but it's also insensitive to individuality. The most probable text is the most average one: it gravitates to the center, to the conventional. And tuning by human preference pushes the prose even further toward that center, rewarding the inoffensive. The result tends to be competent and lukewarm: it reads without stumbling and is forgotten without effort.

This isn't speculation. In 2024, researchers Vishakh Padmakumar and He He, of New York University, published a controlled study in which writing with a preference-tuned model reduced content diversity and increased the similarity between different authors' texts. The machine pulled the writing of distinct people closer together, making them sound alike.

For a firm, the translation is uncomfortable: if every firm adopts the same generic AI, everyone's filings drift toward the same average voice, and a differentiator that took decades to build dissolves at the speed of adoption.

From Problem to Solution

The question that follows is whether AI can preserve voice instead of flattening it. It can. If stylometry knows how to measure how an author writes, it's possible to measure a firm's style from its own body of work and use that profile as a constraint that guides AI-assisted drafting — instead of letting the model fall back on the average, you hand it the firm's fingerprint as a rule to follow. Output then chases the most probable text for that firm, not for the world.

That's different from pasting a template filing into a prompt and asking it to "imitate": an isolated example frays over the length of a document and doesn't transfer between filing types; a measured, structured profile is a persistent constraint that generalizes. LawAgent implements a solution of this kind — internally, the Stylometric DNA — always under human review. We keep the mechanism proprietary: it's a competitive differentiator, and a firm's style profile is sensitive material that belongs to the firm. The choice available today is no longer "generic AI or no AI at all": there's an AI that writes in your voice, not everyone else's.

From Tacit Knowledge to Institutional Asset

In the language of knowledge management, the SECI model developed by Ikujiro Nonaka and Hirotaka Takeuchi (1995) describes how knowledge converts between tacit and explicit. Socialization (tacit→tacit) is exactly training by osmosis: the junior absorbing the partner's style through proximity — and it's the channel the diamond narrows. Externalization (tacit→explicit) is converting that knowledge into something recordable and reusable. Stylometric codification is precisely the externalization of a firm's written voice.

Once style becomes a structured asset, it behaves like one. It survives departures and retirements; it accelerates onboarding, steering a newcomer's drafts toward the house voice from day one; it standardizes quality across a larger team without flattening it to the market average, because the standard is the firm's own voice. And it solves an increasingly common problem in the Brazilian market: in firm mergers and integrations, which voice prevails stops being a tacit power struggle and becomes an explicit decision.

There's a limit worth stating plainly: externalization is never complete. Part of what makes a great lawyer's writing excellent resists codification — the judgment of when to break one's own pattern, the read on a specific judge. The codified profile is a floor and a scaffold, not a ceiling. It preserves the signature and accelerates training; it doesn't manufacture a great lawyer.

Authorship, Confidentiality, and Ethical Limits

Who is the author? The filing is signed by a human lawyer, who reviews it and assumes full responsibility — the AI never signs or files anything (a principle from Edition 3). Codification shapes the form; it doesn't displace authorship.

Whose style is it? When a firm codifies the voice of a partner who later leaves, who owns the profile — the institution that built it, or the individual whose habits it captures? There are legitimate, competing perspectives here, and the Brazilian tradition of authors' moral rights (Law No. 9.610/1998) makes the question less obvious, even though legal filings are functional works. We don't claim to settle the controversy; the prudent recommendation is to address it explicitly in partnership agreements, before it turns into litigation.

Style serves the argument, never the other way around. The most serious ethical risk is making a weak argument sound as if a strong firm had written it. Substance — citation verification, whether authority is still in force, the underlying reasoning (Edition 3) — governs the content and comes first; style is a layer applied on top of already-verified material, never a substitute for it.

Confidentiality and the LGPD (Brazil's data protection law). The raw material for extraction is the firm's body of work, which contains client data. The profile is derived from the firm, belongs to the firm, and stays locked to it, under strict tenant isolation; the data doesn't train shared models, per CFOAB Recommendation No. 001/2024. And preserving style is a proposition about identity and craft — not about litigation outcomes: in line with OAB Provimento No. 205/2021, we don't claim that codifying voice wins cases or raises success rates.

The diamond-shaped structure narrows the base and thins out osmosis; generic AI flattens voice toward a faceless average. Caught between the two pressures, a firm's most distinctive asset — the way it argues — risks dissolving through neglect. A firm's signature is an asset to be governed deliberately, not an outcome to be hoped for.

This content is for informational and educational purposes only and does not constitute legal advice.