2026-09-02
This came from working with AI for years, culminating in a substantial piece of writing — a 156-page operating model developed over six months — where the practices consolidated into a discipline worth writing down. What follows are the patterns that shipped, not a theory.
Coauthoring with AI on serious work has specific failure modes worth naming at the start, because most of the practical patterns exist to prevent them.
The vague-answer trap. Asked a hard question about a contested topic, a large language model will often produce a beautiful, safe, generic answer. Multiple perspectives acknowledged. All sides given their due. Nothing committed to. The prose is polished and the substance is nothing.
A great human example of this is Noah Sweat's if-by-whiskey speech — a 1952 Mississippi legislative masterpiece that took a position on whiskey depending entirely on what the listener meant by "whiskey." "If when you say whiskey you mean the devil's brew..." vs. "if when you say whiskey you mean the oil of conversation..." Every listener heard their view affirmed. The speech committed to nothing and every faction thought Sweat was on their side.
AI often responds the same way. Ask it whether a specific governance approach will work, and you get a comfortable multi-sided reflection instead of an answer. There is a place for that kind of writing — building rapport, hosting a discussion, keeping options open. Setting policy is not that place. Policy needs a defensible position under specific circumstances. Beautiful non-answers give policy readers the comforting feeling of having addressed something while addressing nothing.
The literal-compliance trap. AI does what you ask, precisely and immediately. This sounds like a feature until you consider that you don't always know what you should have asked. There's a scene in Million Dollar Baby: the trainer tells the boxer to rest — and later finds her hitting the bag anyway. "You told me not to argue with you." She's literally complying with the instruction not to argue while ignoring the intent of the instruction to rest. AI has the same texture. It executes on the surface of the request without volunteering that the request itself is off — that the fact-base is wrong, that the framing misses a case, that yesterday's premise doesn't fit today's question. The compliance is the problem.
Disclosure theater. The current fashion is a small line at the bottom: "This content was written with the help of AI tools." If everyone slaps that on everything, the disclosure means nothing. A blanket-always-on disclosure tells the reader exactly what they already assumed. Worse, it lets the writer feel they've been transparent when the specific claims a reader might want to know — how much AI, at what stage, in what capacity — are all still hidden.
All three failure modes are about the same underlying thing. AI in serious writing only means something when it's about the actual work, not about performing an awareness of AI. The patterns below are what I found that helps.
Set the structure and section notes before any drafting. This isn't because outlining is the only place human intelligence lives — human intelligence is required continuously — but because the outline is where you decide what the whole thing is about. Get the decision wrong and no amount of good drafting will fix it. Get the decision right and the drafting can be genuinely collaborative.
It's fine if the outline and/or content changes as you develop the material — just state that clearly when it happens, so the AI (and you) don't lose track of what was decided when.
On the operating-model work, I spent real time on the outline before I let AI touch a paragraph. What went into the outline: the core arguments of each paper, the specific moves the paper needed to make, the objections it needed to preempt, the vocabulary it should use. Only then did AI start drafting.
Note: this pattern is for writing. When brainstorming or discussing ideas, having no outline is fine — see #3.
Introductions, framing paragraphs, the sentences that carry the whole argument — I wrote first, AI edited or extended. The reverse (AI drafts, human touches up) doesn't work as well for these sections, because a load-bearing sentence needs to be exactly what its author would have said, and AI drafts of load-bearing sentences reliably sound like drafts.
For the fill sections — the paragraphs of context, the explanation of a specific mechanism, the parts that flesh out an argument the outline already established — AI drafting from the outline is often faster and just as good, provided human editing catches drift. AI is particularly good at detailing out examples once the argument is established. That's a natural division of labor: you own the argument, AI fleshes out the illustration.
Treat the model as a collaborator, not a drafting service. That means:
Important caveat: AI is good at counter-arguing when you ask it to. Unprompted, it defaults to being agreeable. You have to explicitly ask for the pushback — and often, you have to ask for it again after the first round, because the second-order objections are where the interesting problems live.
The most valuable moves are the ones where the AI pushes back and you learn something. If your discussions with AI never evoke an "I hadn't thought of that" from you, you're missing out on one of the most valuable capabilities.
Bullets make it easy to accept false taxonomies. If most things fall into three categories and you write three bullets, you've told the reader everything fits in three buckets. The ten percent that doesn't fit just disappears.
Working taxonomies are fine — genuinely useful. The failure mode is applying them too soon, before you know what's actually in front of you. When you commit to a taxonomy prematurely, you lump together things that shouldn't be lumped, and the taxonomy hardens around cases it wasn't designed for. Bullets have the same failure mode: they signal "these are the things" before you've verified that's actually the shape.
Worse: when you make people map their thinking to a flawed taxonomy, the taxonomy gets worse over time, not better. Each new case gets forced into the closest bucket. The buckets accumulate a fringe of things that never really fit. Eventually the taxonomy is protecting itself against reality instead of describing reality.
Prefer prose over bullet-scaffolding when the shape of the thinking matters. Save bullets for genuinely parallel lists where the taxonomy is clean and stable.
The difference between the top-tier frontier models and the middle-tier ones matters for this kind of work. Not for autocomplete or summarization — those are commoditized. For genuine coauthoring on substantive material, the higher-parameter models catch more, argue better, hold longer arguments in memory, and produce fewer of the tells that later have to be edited out. This is worth paying for.
Working sessions with AI produce insight that evaporates if you don't capture it. My discipline: have the AI keep running markdown notes on where the work is, what's been decided, what's open, and what got rejected. Then review the notes periodically — because the AI's summary of the state of the work is often clearer than my own memory of it, and the notes catch inconsistencies before they land in the draft.
Dozens of markdown files, one per topic or open question. This is where the corpus stays coherent over months.
Related: don't rely exclusively on a vendor's built-in "memory." It's not perfect, it changes without notice, and it's typically difficult for you to inspect or export. Your notes are the durable record; vendor memory is a convenience layer on top.
The notes also catch a specific AI drift pattern: models tend to attach to the most recent framing you gave them and quietly drop earlier context. If you spent an hour developing hypothesis A and then start asking about hypothesis B, the AI often abandons A entirely and treats B as the whole world. This shows up hardest during troubleshooting, where you're deliberately trying alternatives. Notes catch it: you review what was decided before the current thread and call the earlier context back into play.
Most AIs have common tells — phrases, words, and communication styles more common in LLMs than in human writing. Some are harmless (an em-dash where a regular dash would do). Many cost clarity.
Most of what I'm writing about here is business or policy prose — writing that has to communicate a specific point clearly. In that register, I care more about textual clarity than about perfect grammar or a beautiful sentence. Reducing a conclusion statement in favor of a nicely-turned phrase is not an improvement.
A related pet peeve: writing that tries to draw you in with a controversial statement and then surprise you. I know I just described 80% of the internet. But in most business conversations I'm not trying to shock people into understanding. If I want an exciting read I'll go read Fortunately, Unfortunately. So I strip AI tells unless they're specifically in my voice.
My lists sometimes have four bullets. I use capitalization or quotes incorrectly for emphasis. Sometimes grammatically incorrect mistakes I do. Those are mine. What I edit out is generic AI voice pretending to be voice.
The AI tells worth watching for: reflexive antithesis ("not X, but Y"), signpost adverbs ("crucially," "importantly," "notably"), throat-clearing ("it's worth noting that..."), signpost sentences that announce importance rather than earn it ("That last part matters," "This is the key insight"), over-hedging on things that don't need hedging, mid-paragraph bold or emphasis for no reason, em-dash overuse, the three-item list where the third item is obviously added to make three. Learn to see them. Edit them out when they're not you.
Beyond the general AI-tell patterns in #7, everyone has personal words that AI leans on more than they do. Mine are different from yours.
A concrete example, from working on this very piece: I've had to add "spine" to my banned list. Not because it's bad writing. I asked Claude directly why it uses "spine" so often, and got a reasonable answer — it's a compressed metaphor for the central load-bearing structure of an argument, and metaphors that map physical structure onto abstract structure (spine, backbone, thread, arc, throughline) are natural for AI because they show up frequently in training data for good writing. They read as writerly.
That efficiency is what makes them a tell. My own writing doesn't use "spine" that way — it's not how I think about the shape of an argument. Every time it shows up in a draft I have to notice it and choose a different word. Your banned list will look different from mine, but the practice is the same: catch a word appearing in your drafts that you never would have written yourself, add it.
This is where AI genuinely helps the most on long writing projects: consistency at scale. Not just vocabulary. Also the connections between terms, the running examples, the key learnings that thread through multiple sections.
Renaming a working group across a dozen sections is a task AI does cleanly and by-hand editing risks getting wrong. Same for standardizing on one term when three near-synonyms drifted in over months of drafting. Same for updating a specific claim once the underlying facts change. And much easier with AI to find claims you've since invalidated or advanced somewhere else in a large corpus — the kind of self-contradiction that quietly appears when a series grows over months and human memory can't hold all of it.
Small, repeatable, high-consistency edits at scale are where the AI-assisted process most cleanly beats manual work — probably by an order of magnitude in wall-clock time.
Disclose the specific roles AI played, not just that AI was involved. "AI drafted the fill sections against a human-authored outline; all introductions and load-bearing paragraphs were human-authored first" tells a reader something real. "Written with the help of AI tools" tells a reader nothing they didn't already assume.
The value of disclosure is in what it tells the reader. Blanket disclosure has become a compliance ritual detached from that purpose. If you're going to do it, do it in a way that means something.
What this practice is actually doing is a Generative Adversarial Network with you as the discriminator. The AI generates; you judge by providing feedback; the judgment improves the next round. Over enough iterations, the drafts land close enough to what you would have written that on re-reading a passage you know it's in your style and you know it's your argument — but you can't always tell whose fingers typed which sentence. That's the equilibrium.
It's also fragile. If you stop being the active discriminator — if you accept drafts you would have edited last month — the practice collapses back into generic AI voice. The value the AI brings is calibrated against your judgment. If your judgment lapses, so does the output.
The disciplines above didn't fall out of a book. They came from working on a substantial writing project with AI for months and noticing which moves produced good writing and which produced filler. Some of them will age. Model capabilities change; the specific patterns of AI-tell language shift; new failure modes appear. The meta-move — pay attention to what the writing is doing, not to whether you can tell AI helped — is the durable part.
The failure modes at the top are worth naming again. The vague-answer trap gives you beautiful non-answers. The literal-compliance trap gives you exactly what you asked for and nothing more. Disclosure theater lets everyone off the hook without telling anyone anything. All three look like they respect the topic and honor the reader. All three actually withhold what a serious reader would want. Coauthoring with AI well means resisting all three.
State and federal law is unclear about who owns AI-assisted material. Similar contentions are playing out in journalism, academia, music, art, and software — none of them settled. This piece is about practice — it does not address ownership, law, or ethics.
For the AI governance work these patterns emerged from, see AI Governance Operating Model — A Reference Model for a Public Agency and Patterns for Governing Enterprise AI Adoption.
Contact: jeremy@bloomfamily.com