In 1937, the leading Vermeer scholar alive, Abraham Bredius, studied a fresh canvas, a “Supper at Emmaus,” and pronounced it “the masterpiece of Johannes Vermeer of Delft.“ [2]

Han van Meegeren had painted it, and the most trained eye in the world never caught the fake.

Your editorial eye is that same instrument, and a clean AI copy of your voice passes exactly it.

A draft lands clean, on-brief, close enough, and you call it good because it fooled your eye.

So you polish the surface, scrub the giveaways, ban the em-dash, the “delve,” the tidy symmetry, and you ship.

The polish feels like quality control.

But you trust the eye that put Bredius’s name on a Van Meegeren.

So here is what I want to hand you: a probe that reads the deep structure the copy never carried, and you learn to judge a draft like an authenticator, wary of the fooled expert.

Subscribe now

The distinction every authenticator knows and every prompt forgets

Art authenticators drew the line in one sentence, long before anyone fed samples to a machine.

The Encyclopædia Britannica states it flat: a forger “usually works for the surface effect and is not concerned with the internal structures.” [1]

He reproduces what the eye sees.

The brushstroke, the signature, the palette.

He cannot reproduce the process, the sequence of decisions behind every mark, that made the original, what Britannica calls “a style, a flair, a verve unique to himself,” the thing that can be “stylistically analyzed and documented from his known works.”

Read that against your own drafts. Copy the surface and you have copied what anyone can see.

The structure underneath is the part that gives you away.

Under a casual look, a surface copy sails through, because a casual look is all eye.

The eye compares surfaces and a good forger leaves it nothing to catch.

No eye rebuilds how the marks were made, the pressure and sequence and habit no finished image records.

The instrument does that job, catching what the hand gives away by habit.

In painting, instruments read the deep structure the eye skips: pigment analysis dates the materials, the layers underneath show up on X-ray, thread-counting maps the weave of the canvas.

Where a perfect copy still gives itself away

Text has its own instruments, authorship attribution and stylometry, which read the function-word habits and clause shapes beneath a writer’s finished sentences.

Forgers have always known this.

Satisfy the eye, and you have built the reproduction to fail the instrument that ignores it. I find the same mechanism on both sides.

• The visible copies easily.

• The fake gives itself away on the invisible-measurable.

So why does even the cleanest fake still fall?

Not because the forger was sloppy.

The fake falls because no surface fidelity recreates the generation underneath, the thousands of small choices that produced it, and the instrument reads exactly that.

A flawless surface and an authentic origin are two different objects, and only an instrument tells you which one you hold.

How the best fakes actually fell

Van Meegeren’s fake held for years, and it did not come apart because a sharper critic finally looked harder.

Provenance undid it!

When the Allies searched Hermann Göring’s collection after the war, they found a “Vermeer” he had traded 137 looted paintings to acquire, and the trail led back to van Meegeren, who confessed in 1945.

The eye had signed off but the paper trail and the instrument brought it down.

The pattern repeats with scale attached.

Wolfgang and Helene Beltracchi (two genius forgers) admitted in court to fourteen forgeries; Vanity Fair reported that the works they were charged with selling brought roughly 16 million euros, about 22 million dollars, and the couple drew prison terms of six and four years. [3]

Fourteen surfaces good enough to move eight figures, and not one undone by a better viewer. They fell to provenance and to the lab.

Here is where your voice still refuses to be copied

Now run the same test on a machine, at a scale no forger reached.

In 2025, a team led by Zhengxiang Wang measured what happens when frontier models get an author’s samples and a prompt to write in their voice (five samples, temperature zero, the exact gesture every voice tutorial opens with), in work published at EMNLP 2025. [4]

On structured registers like news and email, the imitations pass.

On personal blog writing, where a voice is most itself, the generated text read as the author’s own hand only about 17 to 21 percent of the time, against a 91 percent human baseline.

Feeding the model more samples did next to nothing. George Mikros, in a separate 2025 stylometric analysis, found the same shape from the other side: GPT-4o “captures some surface-level stylistic elements” but “struggles to fully replicate the depth.” [5]

The model, like the forger, reproduced the surface and missed the hand. Wang’s attribution models are the X-ray, pointed at text.

Stop trusting the eye a clean copy is built to fool

The concept lodges in your production chain one station upstream of where you’d look, before the edit, before you hunt tells or smooth the surface, at the moment you decide whether a draft is your voice at all.

I stopped grading by the surface.

The drafts that clear your list of AI tells are the ones most likely to be a clean fake of your voice, for the same reason the best Vermeer eye passed the best fake.

Scrub a surface and you have not made it more yours; you have made it a better forgery. So I take the authenticator’s stance: drop the eye, run the one probe that reads the deep structure a surface copy can’t fake.

Here is the probe. Its name is older than any model: Biber’s multi-dimensional analysis of register variation.

Douglas Biber, a corpus linguist, formalized it in 1988, in a book called Variation Across Speech and Writing [6], years before anyone thought to feed samples to a machine.

It counts the grammatical habits you never chose on purpose: how far the present tense runs ahead of the past, how hard the first person leans over the third, the length of your clauses, the function words you reach for without noticing.

While the eye reads the surface, Biber reads the structure beneath it, and prints that structure as numbers.

Share

Run the X-Ray Yourself

I built this X-ray and I’m handing it to you ready to run: point it at your own writing from before the AI, then at what the AI hands back, and read the gap.

1. Open Google Colab. Go to colab.research.google.com (free, needs a Google account). File → Upload notebook → the “Upload” tab, and drop in biber_probe.ipynb. The notebook opens with its cells already in order.

2. Run the Setup cell. It’s the first code cell. Click it, then the ▶ button on its left (or Shift+Enter). It installs the tagger and the language model: give it a minute or two. If Colab shows a banner offering to restart the runtime, take it, then run this one cell again. You know it worked when it prints “Ready. spaCy …”.

3. Run The instrument cell. Right below. One click, ▶. Nothing visible happens: it loads Biber’s norms and formulas into memory. You change nothing here.

4. Paste your two texts (Step 1). In the cell holding TEXT_YOU and TEXT_AI, replace the sample text sitting between the triple quotes “”“…”“”. Into TEXT_YOU, something you actually wrote. Into TEXT_AI, the imitation a model gave you when you told it to write “like you.” Leave the quotes and the variable names alone, swap only what sits between them. Aim for at least ~200 words a side, or the scores wobble. Run the cell (▶): no output, it just stores your texts.

5. Run the Step 2 cell. This is the one that measures. One click, ▶. A few seconds later you get: a SURFACE block (words, sentence length, word length, TTR) where the two texts look alike; a DEEP STRUCTURE block with Biber’s 6 dimensions and a drift column, the distance between you and the copy; the total absolute drift; the closest Biber text type for each; and a bar chart standing your scores next to the copy’s.

6. Read it. The last cell (”How to read this”) walks through each dimension. The logic runs one way: the surface block matches almost every time, that’s the part a model optimizes. What talks is the drift column, and above all any pole flip, a dimension that runs positive in you and turns negative in the copy. That flip is the hand the forger couldn’t reproduce.

Where this instrument earns its keep

I built this for one job, catching my own drift. The same probe does four more, and the last one I’d leave to experts.

1. Your own drift over time → This is the number-one use for a working writer. Lean on AI for six months and your fingerprint migrates toward the model’s defaults. The eye misses it, every draft looks fine on its own. The numbers catch the trend. This one isn’t for catching anyone. It’s for catching yourself. For most readers, that’s the use that sells hardest.

2. Build the profile, don’t just audit it → You never hand a model a target it can aim at while the target reads “sound like me.” Turn the fingerprint into numbered rules and the “write like me” system finally has coordinates. That’s the build in the P.S.: measure first, imitate second.

3. Team and agency quality control → A brand with a house voice and five writers, some human, some assisted, can measure that every piece holds the house fingerprint before it ships. The voice becomes a spec you can check, an impression turned into a number.

4. Vet an AI writing product before you trust it → The vendor says it “learns your voice.” Run its output through the probe. Does it hold the structure or only the surface? You stop buying the promise and start measuring it.

5. One frontier, and I name it with care: forensic authentication → Contested documents, legal disputes. It’s a real field, it’s expert territory, and the stakes run heavy. The limit that governs the whole tool bites hardest here: a difference between two texts is not proof of authorship. The probe measures distance. It never tells you why the distance sits there. Treat it as a boundary. Leave it to the people trained for it.

As for me, I’ll see you next Wednesday with a new article!

Thibaut Buewaert
Editor of Machine Writing.

Leave a comment

P.S. 95% of market voice profiles remain superficial. The remaining 5% deconstruct the author’s generative voice grammar using “Zero-Hard-Voice.” The next generation will rely on Biber’s registers but integrated into the profile as rules so that the structure supports the voice. This is the version I will be dissecting in my upcoming articles, so stay tuned 🤠

(Note: The sources and references for this article will be available directly on the web version the day after publication.)

For ninety years, psychology has split learning into two moves.

You can pile up finished examples and try to match them, or you can extract the rule that produced them.

• One overfits and breaks on anything new;

• The other generalizes, because it throws away the surface and keeps the structure underneath.

Paste ten of your best articles into a model and say “write like this,” and you’ve forced it into the move cognitive scientists flagged as the weak one back in the 1930s.

A movement where it stores your surfaces, averages them, and three paragraphs later your voice is gone and the draft reads like everyone else’s voice using AI.

(Right now I think there is no need for a case study… Everyone reading this has watched their own voice come back as oatmeal anyway.)

In this article, I’ll explain why you must stop stacking samples and I’ll show you the anatomy of the grammar I now hand the model instead.

Because a 40,000-generation study says that’s the rung the whole market keeps missing.

Subscribe now

Catch me if you can (not yet)

In November 2025, a team led by Zhengxiang Wang ran the largest clean test of the “write like me” gesture to date, published peer-reviewed at EMNLP 2025 [1].

Over 40,000 generations per model, more than 400 real authors, four domains: news, email, forums, blogs.

Frontier models from OpenAI, Google, Meta, and DeepSeek, each fed a handful of an author’s samples and asked to produce something new in their style (The same prompt every voice tutorial opens with, the one you ran last week.

The result was evaluated through four computational lenses: authorship attribution, authorship verification, a distributional style model, and an AI detector used as a proxy for whether the text still looked machine-generated

Across the board, the verdict split.

• On structured, constrained registers, news and email, the models hold;

• On the personal, informal ones, they fall apart.

On blog-style personal writing, the odds that the generated text was judged the author’s own hand dropped to 17 to 21 percent, against a 91 percent human baseline.

As a result, where voice is most constrained and least personal, the imitation convinces. But where a voice is most its own, it gives itself away.

And feeding it more does nothing, because going from two samples to ten leaves the metrics flat.

The fix was already sitting in the literature

A second paper closes the loop from the other end.

Yang and Carpuat had a model describe a target style through structured register analysis, theory-grounded descriptors that name how a text behaves, a job raw examples and vague adjectives never do [2].

It pays off, with style transfer at least as strong and a large gain in meaning preservation, a clean split between style and content.

Their strongest variant described the style contrastively, against the model’s own generic default.

Both papers point toward the same production lesson: raw examples are a weak interface for personal voice, while structured descriptions of how a text behaves give the model a better handle than vague style adjectives or surface imitation alone.

Everyone can read these two papers (open-access for months, and nobody shipping product has clicked).

The law that takes the copy off the table

Here is the law, named: Zero Hard Voice.

Never feed the model writing samples in production. Describe the mechanics of how the author thinks, not examples of what the author wrote.

Because giving the model writing samples means you handed it a surface, and a surface is all it can copy.

So it matches the statistical tells, word frequency, sentence length, the transitions you lean on, and never the cognitive mechanics that make the voice.

What you hand it instead is a generative voice grammar, the voice decomposed into operational parts the model runs as executable rules.

• For research, proving the few-shot move fails is the whole job;

• For production, the diagnosis is not enough, you industrialize it, you swap the corpus the model copies for a grammar it executes.

Where I hit the wall, and what was wired wrong

I hit this in January 2026, on the bench, before I had read Wang.

I tried to get a voice the way everyone does, dropping the corpus into the context and asking for “the same style.”

What broke was the calque. Word for word, it copied one mechanic example and never reached the pattern behind it.

A short, sharp closing line I had written for a piece on Cordyceps came back re-pasted whole into a piece on egregores, same words, wrong subject (It had not even changed the noun).

Then the same failure on titles, on transitions, on calls to action.

Share

The anatomy of your voice grammar

The solution is generative voice grammar.

You describe the voice as mechanics, by function, where the sample method only ever handed over examples.

A voice grammar holds a small set of named parts: the invariant core that fires in every text, the rhythm operators, the codified punctuation, the speaking positions the writer argues from.

So here is the anatomy of the file, a tour of what each part does to the model:

Here is why the copy is off the table.

Every part is a generative rule, defined by its function on the output or on the reader, with non-repetition built in (Nothing to paste, nothing to scrape, nothing for the model to average). Nowhere is there a sample of the author’s sentences.

No surface was handed over, so the calque stops being a risk you manage and becomes a move that cannot be made.

That is what this issue gives you, the anatomy of a voice grammar, how the file is built and what each part does to the model. It does not give the protocol to extract your own. That comes next, when I take Biber off the shelf and show the measured grammatical layer, the move from a descriptor you prescribe to one you measure on the author’s own corpus.

What this buys the writer who stops copying

Every edition of Machine Writing, this one included, is written with that file. Now the voice holds without me nursing it past the third paragraph, and the proof is the thing you are reading.

None of this was a prompt problem you could discipline your way out of.

Coming editions take the reader past the anatomy, into building his own grammar and pushing the limits of what a voice architecture can hold.

Thibaut Buewaert
Editor of Machine Writing.

Leave a comment

P.S. Over the next few weeks, I’ll be talking about voice profiles and voice architecture, and showing you how to build the most advanced voice cloning files on the market. And being the most advanced means being on the front line! Right, I’m heading to the front, bayonets fixed. Advanced AI users, cover my six. Beginners, you’d better cover my front ;)

[1] Wang Z, Tripto NI, Park S, Li Z, Zhou J. Catch me if you can? Not yet: LLMs still struggle to imitate the implicit writing styles of everyday authors. Findings of the Association for Computational Linguistics: EMNLP 2025. 2025:10040-10055.
[2] Yang X, Carpuat M. Steering large language models with register analysis for arbitrary style transfer. arXiv. Preprint posted May 1, 2025. Revised May 9, 2025.
[3] Biber D. Variation across speech and writing. Cambridge University Press; 1988.
[4] Van de Maele T, Verbelen T, George D, Pezzulo G. Schema-based active inference supports rapid generalization of experience and frontal cortical coding of abstract structure. arXiv. Preprint posted January 26, 2026. Revised March 14, 2026.
[5] Zubek J, Kuncheva L. Learning from exemplars and prototypes in machine learning and psychology. arXiv. Preprint posted June 4, 2018.
[6] Bartlett FC. Remembering: A study in experimental and social psychology. Cambridge University Press; 1932. Piaget J. The origins of intelligence in children. Cook M, trans. International Universities Press; 1952.

In 1962, a psychologist wrote down what creativity actually is: pulling together ideas that sit far apart, and the further apart, the more creative the result.

Sixty-four years later, most writers with a chat window are doing the opposite, reaching for the nearest idea the model offers.

The whole game is distance.

The two-move method is how you force it, then pay to keep it usable.

But before the method, look at what most people believe about creativity techniques.

Everyone knows the pitch by now.

Find the right technique, the right prompt, the right exercise, and your output gets better across the board: more original, more on-brief, more everything, all at once.

A good technique, the thinking goes, only adds.

It is a “free lunch” you have not ordered yet. Pick it up and the work improves with no cost attached.

That belief feels obvious, and it is wrong in a way that costs you ground.

The truth runs the other direction.

The ad studios that formalized these operators in the 2000s knew it first, and they priced it in.

Subscribe now

Why associative distance is the whole game

Start with the definition itself.

In 1962, Sarnoff Mednick published “The associative basis of the creative process“ in Psychological Review, and the sentence at its center has held for over sixty years.

Mednick defined creativity as the combining of remote associations into something useful, and the more remote the elements, the more creative the result.

This means creativity is not a substance you have more or less of, but the distance you cover between things that do not usually sit together.

Mednick’s mechanism for that distance was the ASSOCIATIVE HIERARCHY: the ranked set of responses a mind produces when prompted with a cue.

Give someone the word table, and one response dominates, chair, with everything else falling off fast behind it.

That is a steep hierarchy: The first answer swallows nearly all the available probability, and the remote answers sit so deep they rarely surface.

A flat hierarchy behaves differently: The first response is less dominant, the curve descends slowly, and more associations stay live and reachable further out.

Mednick’s testable prediction followed directly.

Creative people have flatter hierarchies.

Slower at the Start, deeper at the edge

So creative minds start slower, because no single obvious answer is pulling all the weight, but they reach distant associations that the linear mind never gets to.

He named three routes to that remote combination:

1. Serendipity, where chance throws two far elements together;

2. Similarity, where a perceived likeness bridges two domains;

3. And mediation, where an intermediate element links the two.

From this he built the Remote associates test, the instrument that measures whether a person can find the single word connecting three unrelated stimuli.

But the model did not survive untouched. Revisiting the model five decades later, Benedek and Neubauer found that associative hierarchies did not clearly differ between low- and high-creative groups.

The stronger differences appeared in associative fluency and uncommon responses, not in the hierarchy shape alone. Mednick had the result roughly right and the exact mechanism slightly wrong, which is the normal fate of a foundational idea.

The core has held: remoteness is the raw material of original combination, and reaching it is the work.

What the studios knew, and your prompt forgot

You can watch this happen outside any computer.

In the advertising studios of the 2000s, Jacob Goldenberg and his colleagues formalized a set of creativity templates, named operators like subtraction, unification, and extreme consequence, that constrained how a team attacked a brief.

The point was not inspiration.

It was structure: force the idea through a narrow operator, remove a component, fuse two functions, push a consequence to its limit, and the output drifts off the obvious.

A later study in the Journal of the Academy of Marketing Science tested ideation templates in professional advertising development and linked them to creative performance.

Then the bill arrives…

On a model, it arrives in a currency the studios never paid: compute.

The model defaults to the convergent answer, the dead center of the distribution, because that is the response it was trained to reward.

Handing you the center is cheap.

That is what the prompt crowd inherited without the receipt.

The template operators circulate today as prompt instructions, force a metaphor, subtract a feature, push the consequence, but they travel as upgrades, stripped of the one thing the studios always knew: the distance you force is borrowed, and the brief has to be bought back.

Share

Where the bill lands in your workflow

So what changes for how you work?

Mostly this: you stop treating distance as something you either have or lack, and start treating it as something you spend and recover on purpose.

You keep generating angle after angle, and they all come back wearing the same face, the one your three competitors are also shipping this week. Underneath that sameness, the cause is structural, not a failure of talent: a hierarchy collapsed onto its nearest response, yours and the model’s both.

Reach for a better prompt in the same one-pass workflow, and the output often lands where everyone else’s does.

The two-move method moves you: force the angle off-center first, then pull it back to the brief on purpose, and the sameness stops being your ceiling.

The production protocol lives elsewhere, but this stripped-down version is enough to test the trade today: force distance in one pass, recover readability in the next.

Skip the recovery pass and you ship maximum distance that lands on nothing, the lyrical draft that converts no one.

Budget for it and the trade works in your favor.

Run it in two moves

Treat a divergence technique as a trade, not an upgrade. Before you reach for it, name the two moves out loud.

Here are both prompts in demo form, deliberately stripped down (no emotional entries, no compliance checks). Remember, this is a tasting, not the production protocol.

First, force the angle off-center and accept that the draft drifts off-brief; that is the technique working, not failing.

Here's my raw material: [paste specs / facts / context]
Recipient: [who reads this]

1. The Core Fact (The Brief): Extract the naked truth from the raw material.
   In one single sentence, state exactly what the product does or what the
   recipient concretely gets. Zero adjectives, zero marketing words.
   Strictly factual. This is what must survive everything below — keep it.

2. Now give me your reflex formulation — the first angle that comes to you.
   That's the anchor. We're going to move away from it, not keep it.

3. WITHOUT changing the subject, produce 6 angles that move away from that anchor, two per route to remote association:

   SERENDIPITY (chance collision of two far elements)
   - pick a concrete object or scene from everyday life BEFORE you know how it connects, then force the Core Fact through it — keep the collision, even if it's awkward
   - take two details already in the raw material that have no reason to sit together, and write the angle that only appears once you force them into the same frame

   SIMILARITY (a likeness bridges two domains)
   - one cross-domain analogy: map the Core Fact onto a domain FAR from the material (not tech, not marketing) — the mapping must carry a specific mechanism, not a vibe
   - find a tension, cause or dynamic INSIDE the material and mirror that same structure onto what the recipient actually experiences

   MEDIATION (an intermediate element links the two)
   - surface the hidden middle term: the unstated step B that sits between the Core Fact and why the recipient should care — write the angle on B, the thing no one names
   - follow the chain outward (the fact enables X, which enables Y) and write the angle on the far end, not on the fact itself

For each angle: first say in one line WHERE you're going, then write the angle.
Stay on the surface: what the recipient gets, does, or concretely avoids — never what they feel or become.
End each with: already seen? Yes / Maybe / No — and where.

Then run a separate, deliberate pass to pull it back to the strategy.

Open a fresh window with the first prompt closed, or the model drifts straight back to its reflex. (Stating the obvious, but just in case: don’t forget to re-inject the raw material.)

Distant angle I'm keeping: [paste the angle you chose above]

The Core Fact to make readable (paste it verbatim from step 1):
[the one-sentence factual truth]

Rewrite this angle so that the Core Fact becomes READABLE in a single pass —
but keep the distance: the structure, the inversion, or the dimension that
made it different has to survive.

Then check, in this order:
- Does a rushed reader get the Core Fact on the first read? If not → rewrite.
- Has the angle collapsed back into the obvious formulation anyone would have
  written? If yes → you've over-bought. Back off, recover the distance.

Compare this two-move protocol to your usual one-shot prompt, and the results will speak for themselves.

Thibaut Buewaert
Editor of Machine Writing.

Leave a comment

P.S. Starting next week, we’re opening a new thread: writing voice profiles. How to build yours, and above all how to get the model to adhere to it. More to come.

P.P.S Here is the exact output of this double-pass protocol run on this very article:

[Pass 1: The Divergent Angle]

SIMILARITY → Where I’m going: deep-sea diving, far from tech, mechanism first.

A diver reaches the wreck in one drop. Coming back up is the part that kills you if you rush it: you ascend in fixed stages or the dissolved gas turns on you. Distance is the descent; the recovery pass is the decompression stop. Skip the stop and you surface with something no one can read.

Already seen? No. (Mechanism that carries: staged ascent = a separate, deliberate second pass — not a vibe.)

[Pass 2: The Recovery Rewrite] Here’s the recovery-pass rewrite — the diving frame intact, the Core Fact surfaced:

Diving is two moves, and so is the creativity. The first prompt is the descent: one drop straight past the model’s default answer, down to the wreck — the farthest association you can reach. The second is the ascent, and you run it separately, in fixed stages, because that’s the part that kills you if you rush it. On the way up you rewrite the draft to restate your one factual sentence, but you hold the depth you bought — rising slow enough that the pressure never turns the dissolved gas against you. Two prompts, opposite directions: down to where it’s strange, back up to where it can be read. Skip the ascent and you surface with something no one can read.

[The Compliance Checks] Running the two checks:

• Does a rushed reader get the Core Fact on first read? Yes. All three load-bearing pieces are now on the surface, not buried in the metaphor: two separate prompts (”two moves… you run it separately”), the first moves off the default (”straight past the model’s default answer… the farthest association”), the second restates the one factual sentence while holding distance (”rewrite the draft to restate your one factual sentence, but you hold the depth you bought”). The line “down to where it’s strange, back up to where it can be read” is the catch-net for anyone skimming.

• Has it collapsed into the obvious formulation? No. The thing that made the angle distinct — the inversion that the return is the dangerous part, not the plunge — is still doing the work (”that’s the part that kills you if you rush it”). The depth/pressure dimension and the staged-ascent-as-decompression structure both survive. It hasn’t flattened into “do two prompts, one for distance and one for clarity,” which is what anyone would have written.

[1]. Mednick SA. The associative basis of the creative process. Psychological Review. 1962;69(3):220-232.
[2]. Benedek M, Neubauer AC. Revisiting Mednick’s model on creativity-related differences in associative hierarchies: Evidence for a common path to uncommon thought. The Journal of Creative Behavior. 2013;47(4):273-289.
[3]. Goldenberg J, Levav A, Mazursky D, Solomon S. Cracking the Ad Code. Cambridge University Press; 2009.
[4]. Tevi A, Parker J, Koslow S, Ang L. Creative performance in professional advertising development: The role of ideation templates, consumer insight, and intrinsic motivation. Journal of the Academy of Marketing Science. Published October 31, 2024. 2025;53:854-875.

The best-positioned writers using AI in your market may be more exposed than they think.

Because they’re feeding their one advantage into the same machine as everyone else.

When they converge, the niche you thought was taken quietly comes open.

You don’t feel it happening, because your own output looks fine.

You read it back and it holds: the structure, the argument, the angle you picked.

So you file it under “safe.”

But that is a huge misconception!

By the end of this edition you’ll see why a shared model collapses a whole market onto the same ideas, and why that collapse is the door, not the wall.

Subscribe now

The collapse is invisible in your draft.

Start with the version of AI sameness people get wrong.

Most writers picture it as a flaw inside the model: ask the same model the same question twice, and it hands back the same shape of answer.

You have seen it, and it has a name.

MODE COLLAPSE: the pull of a model back toward its most typical output. But that version is intra-model. One box, one screen, and you push the model off its default yourself.

Markets turn on the other effect.

Call it a POPULATION EFFECT, one that lives between people, not inside the model. Watch your neighbor open the same chat window, type a competent prompt, read back a competent answer. No single output is broken.

What collapses is THE VARIETY ACROSS THE OUTPUTS: the distance between your output and your neighbor’s shrinks until the whole field is standing in one spot.

Picture it from above: a thousand newsletters opening on the same hook, naming the same three drivers, closing on the same reframe…

Stack them and you get a single voice wearing a thousand bylines.

Intra-model collapse is a quality problem you can catch in your own draft.

Population collapse is invisible from within your own text, because your draft looks perfectly fine on its own.

In a 2024 Creativity and Cognition study, Anderson, Shah, and Kreminski compared ChatGPT with a non-AI creativity support tool based on Oblique Strategies.

Participants generated more ideas, and more elaborated ideas, with ChatGPT.

But at the group level, those ideas were more semantically similar than the ideas generated with the card-based tool.

The 1975 divergence strategy that AI can’t replicate

Anderson’s participants used that card-based tool, the one the AI lost to on collective diversity.

Built in 1975, it carries the mechanism.

Then the objection you are already forming: can’t you edit your way back out?

Half a century apart, a card deck and two studies land on the same place. Your spread comes from the source you draw from, not the hours you put into the draft.

Share

The day they converge, the door opens

Here is what I take from this, my reading, not the studies’.

If a whole market drifts toward one center, the center gets crowded and the edges go empty.

At the center, the water goes red, picked over by everyone holding the same position.

Out at the edges, the open water waits, where almost no one is pointing.

Your established competitors are vacating ground as they feed their one real advantage into the same model.

Watch the day they ship the same homogenized output as the rest: the seat you thought was taken comes open, and you do not have to beat them to take it.

In a converging market, valuable difference compounds when you can measure how far your angle sits from the center.

Here is the diagnostic to run on your next piece (I run it in a blank doc, with no model open).

Take your core claim and ask: if a competitor briefed the same model on the same subject, how close would their claim land to yours?

Then write the likely default claim in one sentence.

That sentence is your center.

Your job is to move away from it.

Thibaut Buewaert
Editor of Machine Writing.

Leave a comment

P.S. Seeing the center is step one. Step two is engineering your distance. I’ve built the system I use to force the model off that center. It’s called Forced Divergence, and it uses a 7-territory map to push the model toward angles your competitors are unlikely to find by default.

[1] Anderson BR, Shah JH, Kreminski M. Homogenization effects of large language models on human creative ideation. In: Proceedings of the 16th ACM Conference on Creativity & Cognition. Association for Computing Machinery; 2024.
[2] Moon K, Green AE, Kushlev K. Homogenizing effect of large language models (LLMs) on creative diversity: an empirical comparison of human and ChatGPT writing. Computers in Human Behavior: Artificial Humans.
[3] Chakrabarty T, Laban P, Agarwal D, Muresan S, Wu CS. Art or artifice? Large language models and the false promise of creativity. In: CHI 2024 - Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. Association for Computing Machinery; 2024:Article 30.
[4] Padmakumar V, He H. Does writing with language models reduce content diversity? OpenReview. Published January 16, 2024. Last modified March 6, 2024. ICLR 2024.

Here is the gesture almost everyone makes.

You finish a draft in the same chat window where you wrote it, and you type the obvious next thing, review this, tell me what’s weak.

The same model that just drafted it reads the work back and calls it strong.

Inc. reported a Reddit screenshot that captured the problem perfectly: according to the article, a user shared ChatGPT’s reaction to what they described as a new draft of a school paper.

“Bro. This is incredible. This is genuinely one of the realest, most honest, most powerful reflections I’ve ever seen anyone write about a project.” [1]

Ask for the weaknesses instead and the softness often only changes costume. Remio gives the familiar pattern in one illustrative example: the model opens with praise, then offers “one minor area to consider.” [2]

This has happened to you right ?

OK… When you want a hard read on your work, instinct says hand it to whoever knows it best, the reviewer with the most context, the most of your intent.

With AI that instinct is exactly backwards.

In this new post, you’ll see why the second brain that helps is the one that knows nothing about your draft: not a smarter reviewer holding all the context, but a blank one whose ignorance is the exact thing that makes its verdict usable.

Subscribe now

Claude Code’s review architecture makes ignorance a feature

Truth be told, people rely on the fresh window out of convenience.

The conversation gets long, the answers get vague, so you open a new chat to clear the clutter, the way you would restart a sluggish app.

Becca Caddy, writing for TechRadar, frames the habit practically: starting a new chat can lead to clearer, context-free responses, while continuing a chat helps when you want the model to build on previous work. [3]

Read like that, you reach for it only once quality has visibly dropped.

Underneath that habit sits a different reading.

The separation is not a cleanup.

The people writing code with these models adopted it on purpose, as an architectural rule.

Anthropic makes that separation explicit in Claude Code. Their Claude Blog describes a review subagent as a clean slate: it does not inherit the assumptions, context, or blind spots of the primary conversation.

In their example, the review subagent evaluates the code without knowing what tradeoffs were considered, what approaches were rejected, or what assumptions were made. [4]

That outside perspective does the work.

They did not build the blank slate to rescue you from a slow window. They built it because the context that produced the work cannot be trusted to judge the work.

Which raises the question every prompt fix has stepped around. Why does the producing context corrupt the verdict?

Why the producing context can corrupt the verdict.

Picture what the model is doing when you ask it to judge its own draft.

It does not wake up fresh for your review request.

Every message you send is read against the whole conversation sitting above it, and that conversation is the record of the choices you and the model made together to build the text.

Put a draft on the table and the model sees no neutral object.

It reads the conclusion the thread was building toward, and treats it the way the thread taught it to: as something to defend.

Two quieter forces ride the same slope.

Sathish Raju, writing from a developer’s perspective on Medium, describes the same wall more bluntly: in long Claude Code sessions, he argues, the context window fills with noise and the signal-to-noise ratio collapses. [5]

The model is not broken. It is conditioning on its context exactly as it was built to, which is why three years of sharper critique prompts never closed the gap. The fix kept aiming at the instruction, and the instruction was never the problem.

The contamination lives in the context the text never left.

Share

Never let the builder grade the build.

Stop treating the blank session as a last resort, the move you reach for only once a conversation has visibly fallen apart.

The developers did not wait for the context to rot before separating.

They separated by design (eighteen months from the first subagent pattern to a shipped product feature), before there was any rot to detect.

Until February, I did the opposite, drafting and judging in the same window like everyone else. Do the same with your own writing.

At the end of production, the text leaves the context that made it.

You spend more time arguing the AI out of its own draft than it would take to edit it yourself.

Give the text to a second brain with zero memory of writing it, and you stop negotiating, you get a flat verdict you can act on.

The reviewer that helps is the one that knows nothing about how the draft was made.

Hand your next draft to a reader that knows nothing.

Try this on your next piece, and the setup takes under a minute.

At the end of the session where you wrote the draft, resist the reflex to type review this in the same window.

Of every reader you could hand it to, the model that just built the piece with you judges it worst: it is anchored to every choice you made together and primed to defend them, which is the contamination you have been reading about.

Hand the work instead to a second brain that knows nothing about how it was made, and let that ignorance do the reviewing.

The gesture is three steps:

1. Open a fresh chat outside any project or persistent workspace, with no files attached and no previous draft context pasted above it.

2. Your final text goes in first, then the Cold Read Protocol below it.

3. Read what the cold session flags, then set it beside how the same text looked to you thirty seconds earlier, inside the window that produced it.

Grab it and run it the next time you are about to ask a window to grade its own work.

→ [Grab the Cold Read Protocol (.md)]

In fact, the protocol is one simple structured prompt that hands the blank reader seven checks:

• whether the opening’s promise gets paid off,

• whether the facts hold,

• whether the logic stays consistent,

• whether it fits where it will be read,

• whether the target reader stalls anywhere,

• whether the voice is authored or flattened into machine-default,

• and whether anything important is missing.

It closes on a single line, ship or revise, with the one biggest reason.

It only works cold, by a reader with no memory of the draft (a fresh window, not a new project, or the files carry the contamination straight back in), because that absence of memory is the whole mechanism.

Thibaut Buewaert.
Editor of Machine Writing.

Leave a comment

P.P.S. We’ve only scratched the surface! Imagine taking these correction suggestions and pushing them directly into your draft with a single click. By injecting precise fixes exactly where they’re needed, without altering the rest of your text. We’ll dive into this “injection and self-refinement” methodology later :) Stay tuned.

P.S. Next Tuesday we explore why a whole market converges the moment everyone types into the same machine. And why your sharpest, most defensible take comes out sounding like every other tab in your reader’s browser.

[1] Sherry B. Sam Altman just admitted that ChatGPT has become “annoying.” Here’s why. Inc. Published April 28, 2025.
[2] Washington A. 8 tricks to beat the “yes-man” AI and get brutally honest feedback. Remio AI. Published October 28, 2025.
[3] Caddy B. Should you start a new chat with ChatGPT every time you use it? Here’s your guide for when to keep the conversation going. TechRadar. Published June 4, 2025.
[4] Anthropic. How and when to use subagents in Claude Code. Claude Blog. Published April 7, 2026
[5] Raju S. Claude Code subagents: the complete guide to AI agent delegation. Medium. Published April 4, 2026.

Keep running AI the way you run it now.

Sharp prompts, careful review, you in the loop, checking each draft before it goes out.

It feels like control, like the responsible way to work with a machine.

And you will keep producing work that disappears into the same gray middle as everyone else using the same tools.

Your skill is not the problem.

Because “being in the loop” is not a protocol, and without a protocol, the most skilled operator loses to the most coordinated one.

Everyone argues in two camps, the machine replaces you or you supervise the machine, and both are arguing about the wrong thing.

A chess hall settled this argument before the first LLM existed, and what happened there points straight at your screen.

Subscribe now

What Kasparov saw that the headlines missed

In 1997, Garry Kasparov lost to Deep Blue, and the world read it as the day the machine beat the man.

Kasparov read it differently.

He went looking for what came after, and he found it in freestyle chess, where any human-and-machine team could enter.

What he saw there became the centaur model: the strongest player is human plus machine plus protocol, a pairing that beats either side alone.

And the protocol turned out to matter more than the players.

One variable, proven in four different rooms

Strip the idea down and it stops being a chess story, becoming a claim about how any human-machine system performs.

And the same result surfaces across four unrelated fields:

Different fields, decades apart, one verdict: the variable is the protocol.

Share

When amateurs beat grandmasters on identical machines

Picture the freestyle boards in the years after Deep Blue.

On one side, a grandmaster with decades of pattern recognition and a top engine.

On the other, two amateurs nobody had heard of, running the same commercial software anyone could buy.

The grandmaster trusts his reading, calling on the machine when it suits him.

The amateurs do something stranger.

They follow a fixed routine for when to let the engine calculate, when to override it, and how to settle the disagreements between the two.

They out-coordinate an opponent they cannot out-think.

And the result that shook the chess world was not close.

Identical tools, opposite outcomes, and the only thing that changed hands was the rule connecting human to machine.

Now point that finding at your own screen

Set that beside your own production session.

I built an editorial system, on twenty thousand words of protocol.

It separates two AI brains, a Writer that drafts and an Analyzer that audits, into sealed contexts that never see each other’s reasoning.

The author is no longer a typist.

The operator becomes the nerve center.

By transferring the content between two sealed AIs, the human mind is freed from execution tasks. The human does not write the lines: they arbitrate perspectives, inject their discernment, and lock the decisions.

It is a protocol of thought amplification.

Everything else, the collection, the workshops, the passes with their exit checklists, runs on the protocol.

It is the freestyle board again, except the protocol runs to twenty thousand words; the chess players got by on a post-it.

An agent with no protocol loses, and so does a grandmaster with no process.

The variable is always the protocol.

Stop asking if the machine replaces you (it was never the point)

Once you see this, the debate you have been handed, can AI replace me or do I need to supervise it, is the wrong one.

Both positions assume the same thing, that the answer comes from power, a more capable model or a more vigilant human.

The centaur model says it comes from somewhere else, from the freestyle board where the protocol beat the rating.

I built my own AI writing system around that question before I had the vocabulary for it, and the thing I was missing for months was never a sharper prompt or a bigger model…

…It was the rule for who produces and who judges, the thing that separates your work from the pack.

I cannot share this complete protocol yet.

(Quite simply because it is still in the testing phase on Machine Writing. What you are reading right now comes out of this editorial production system in less than an hour. So you can form your own idea of my horizon.)

On the other hand, what I can do right now is hand you the question that makes you build one.

Run the protocol check before you publish

Before the next piece ships, ask one thing.

What is the protocol between me and the model?

Forget “did I use AI.” Everyone uses AI.

Ask who produced this, who evaluated it, and what rule governed the handoff between them.

When the answer is “I used my best prompt. Then read what it wrote and decided it was good,” there is NO protocol!

You are the grandmaster trusting your own read, and the amateur with a protocol is about to beat you soon…

All of it comes down to one question in the gap between the work you approve and the work you send, nothing new bolted on.

Operators who own that question will pull ahead of the ones still arguing about whether the machine is coming for their job, and the protocol was the variable all along.

And now you know which question tells you whether you have one.

Thibaut Buewaert
Editor of Machine Writing.

Leave a comment

P.S. The centaur model says the protocol is the variable. On Thursday, I’ll show you a simple one that already exists for dev teams. Because your writing needs it.

[1] Kasparov G. The chess master and the computer. The New York Review of Books. February 11, 2010. Accessed June 19, 2026.
[2] Sparrow B, Liu J, Wegner DM. Google effects on memory: cognitive consequences of having information at our fingertips. Science. 2011;333(6043):776-778.
[3] Puccio GJ, Holinger M. Alex F. Osborn: applied creativity pioneer. In: Reisman F, ed. Celebrating Giants and Trailblazers in Creativity Research and Related Fields. KIE Publications; 2021:398-426. Accessed June 19, 2026.
[4] Song TE. Cross-context review: improving LLM output quality by separating production and review sessions. arXiv. Published March 12, 2026.

Steve Honda/AFP/Getty Images - Garry Kasparov during his rematch against the IBM supercomputer Deep Blue, 1997

Using ‘improve this’ seems harmless enough.

Alan Wong, writing on Medium [1], asked ChatGPT to improve a passage.

“A pang of fear hits” came back as “panic punched me in the chest.”

Nobody requested that.

He asked for a better sentence.

He got a more common one.

The model swapped the entire passage and labeled the replacement an improvement.

Wong caught ChatGPT doing it, but the reflex isn’t ChatGPT’s.

It stems from how these models are trained, which puts Claude and Gemini in the same boat

Run it five times across a document and the writing turns average, each pass dragging it one invisible step toward a center you never chose.

GPTHumanizer named the loop in 2026 [2]: keep polishing and the text only gets smoother, which makes it worse, not better.

You have seen this happen.

You can reproduce it in under a minute with any passage you actually care about.

So why does the instruction meant to sharpen your writing do the opposite?

Subscribe now

Every “improve” is a vote for the average

You start with the innocent assumption that the model is simply polishing your draft, sanding the rough edges off a table you’ve already built.

So when the passage comes back flatter, you do the obvious thing.

You run the instruction again.

Another pass, a little smoother, and the loop tightens.

Nobody tells you the plain version.

“Improve this” averages your writing, and it averages it a little more with every pass.

Give the model a reference to compare it to, and that same reflex that was flattening your copy starts pulling it toward a target you control.

You stop asking the model to improve the wrong thing.

The model does the job you gave it.

You keep fixing the wrong thing, because every “improve” is a vote for the average, and re-prompting harder only gets you to the average faster.

Your problem sits below the phrasing.

The asking itself pulls that direction, and sharper wording only rides the same slope.

Three forces pull every edit toward the middle

Three forces pull in the same direction, and none of them is a bug.

Here’s the recap:

1. Anchoring sets the direction.

2. The gravity comes from the training.

3. And tone-latching gives that pull a word to copy, the very one you typed.

The model isn’t failing in any of this.

Put the three forces together and you get exactly what it was built to do: converge.

Share

Don’t ask for better. Point the model at a reference.

So stop asking the model to improve your writing.

You’ve just seen where that leads, since the instruction has no target other than its own average.

Start doing the one thing the model can’t drift away from.

Hand it a reference and ask it to move toward that instead.

Drop “make this better.”

Ask for “make this closer to that” instead.

That’s the swap I reach for now, every time the average starts winning.

You aren’t correcting the model, you’re giving it a direction it doesn’t have on its own.

Left alone, the model always pulls toward the center.

Your job is to decide where the center sits.

The one-line swap you can run in your next session

The next time you’d type “improve this paragraph,” stop.

Find a reference passage.

It could be your own best work, a published author whose rhythm you admire, or a competitor’s landing page that converted.

Paste it below your draft, and write this instead.

Then paste the reference.

You’ll watch the model stop polishing toward its own center and start pulling toward yours.

The output won’t be “better” in the generic sense.

It’ll be closer, and that’s the only sense that matters.

Thibaut Buewaert.
Editor of Machine Writing.

Leave a comment

P.S. This is just the symptom. The architecture underneath it, why a model can’t reliably judge what it just produced in the same session, is the law I keep building the rest of the system on. That’s The Dual Brain, the next brick this field note sits against.

[1] Wong A. Wow, ChatGPT is awful at writing. Medium. Published June 26, 2025. Accessed June 19, 2026.
[2] Miller E. Why ChatGPT writing sounds robotic even when it looks fine. GPT Humanizer AI. Published March 10, 2026. Updated May 12, 2026. Accessed June 19, 2026.
[3] Dykes T. ChatGPT and tone: Avoid sounding like a robot. Nielsen Norman Group. Published March 8, 2024. Accessed June 19, 2026.

Draft, revise with your LLM, reread, send.

It reads clean until the client finds three errors in the first paragraph.

We assume quality scales with concentration and that a third read catches what the second missed.

This version of editing is intuitive, flattering, and wrong.

Because it rewards individual effort, making quality dependent on how hard the writer squints at the chat window.

But it ignores a four-century-old truth: review requires structural separation between writer and reviewer, not just more rereading.

What the popular version of editing calls personal rigor, the professional version calls architectural separation.

That gap is four centuries wide, and your LLM prod session has never crossed it.

Subscribe now

Murder your darlings was never about cutting

In 1916, Arthur Quiller-Couch stood before his Cambridge students and delivered the phrase that would outlive his entire bibliography, “Murder your darlings” [1].

For a century, writers have treated it as an instruction about cutting.

It was never about cutting.

Quiller-Couch was naming a structural problem: authors are inherently attached to their own work, a bond that completely blinds their judgment.

Far from mere literary advice, this principle stands as a rigid engineering constraint, and four separate traditions built their entire editorial systems to prove it.

Four proofs of the same wall

→ J.P. Guilford formalized the cognitive mechanism in 1950 with his distinction between divergent thinking (generation) and convergent thinking (selection) [2].

He found that effective creative protocols forbid critique during generation, because the two processes interfere when sharing the same cognitive space.

In a production session, producer and evaluator cannot coexist in the same session.

→ Stephen King, in On Writing (2000), prescribed six weeks minimum between finishing a first draft and reading it again, suggesting that :

“If you’ve never done it before, you’ll find reading your book over after a six-week layover to be a strange, often exhilarating experience“ [3].

It sounds like patience.

However, this six-week gap serves a purely structural purpose, as it erases the familiarity that blinds the writer to his own weaknesses.

→ John McPhee, in Draft No. 4 (2017), went further [4].

He described the New Yorker’s fact-checking department, which sat entirely separate from the editorial floor, directed for decades by Freddie Packard and later by Peter Canby.

The fact-checker does not work for the writer.

The fact-checker works for the reader.

Every verified fact is for the benefit of the subscriber who opens the magazine on Monday, not the writer who filed the piece on Friday.

→ Carol Fisher Saller, longtime chief copyeditor of the Chicago Manual of Style, formalized the relationship in The Subversive Copy Editor (2009), proving that the copyeditor doesn’t serve the author, but the reader [5].

Author and editor negotiate on behalf of someone they never consult, the reader.

The editor reads what is on the page.

The author reads what he remembers putting there.

Four names, four decades, four institutions: you cannot evaluate what you produced.

Share

From Carver’s letter to your client inbox

Between 1976 and 1983, Gordon Lish and Raymond Carver produced the most radical editorial collision in literary history.

Lish, then fiction editor at Esquire and later at Knopf, cut roughly half of everything Carver submitted for What We Talk About When We Talk About Love (1981).

He changed titles, rewrote endings, cut whole scenes.

Carver himself had written to Lish in 1980: “I want them to be the best possible stories. So open the throttle. Ramming speed” [6].

Then Carver reversed.

After Cathedral (1983), the first collection Lish did not significantly edit, Carver admitted he had “gone as far the other way as I could or wanted to go, cutting everything down to the marrow, not just to the bone.”

Carver and Lish argued about scale, never about the principle.

The producer and the evaluator cannot share a room

Now place that editorial argument next to a forty-minute LLM production session.

In 2024, the ProCopywriters annual survey (n=422 professional UK copywriters) found that 51% integrate AI-generated copy into their final client deliverables [7].

Sixty percent use generative AI in their daily production.

Of course, the gap between “60% use it” and the reality of what gets sent is where the principle vanishes, as copy travels from the production session to a client inbox stamped “final.”

The production context follows the draft to the inbox, and no separate reviewer ever touches it.

Four centuries of editorial practice say this workflow is architecturally incapable of catching its own errors. (The exact pathology Quiller-Couch named in 1916.)

One factor Quiller-Couch never faced was that the producer is a model trained to agree with you, not to challenge what it wrote.

One joint in your workflow, one question

Instead of merely adding a step, this shift changes the very foundation the step runs on.

I built my editorial production system around this principle in February 2026, weeks before I found the vocabulary in Quiller-Couch, in Saller, in McPhee.

What I can now give the reader is what I lacked for months: structural disqualification of the producer.

The concept occupies one joint in your production workflow, specifically the ten seconds between the last revision you approved and the email you are about to send.

Right now, that moment contains a reread inside the window that still holds every prompt, every iteration, every discarded direction.

After this article, it contains a question.

This principle does not tell you how to execute the separation.

That is architecture, and architecture is a different article.

What it tells you is that the separation remains a rigid structural constraint, far from a workflow luxury to be adopted only when time permits.

It is an engineering constraint that predates the LLM by four centuries.

One question before you hit send

Before you send your next deliverable, ask a single question.

Has this text been evaluated by someone or something outside the context that produced it?

Forget questions like “Have I reread it?”

Or, “Did I ask the same session to review its own output?”

Has a separate pair of eyes, human or contextual, read this text without knowing how it was made?

If the answer is no, you’ve reread but you haven’t edited.

Four centuries of editorial practice say that is not the same thing.

Thibaut Buewaert
Editor of Machine Writing.

Leave a comment

P.S. This article names the principle underneath a law from the previous issue: The Dual Brain. That one builds the architecture. This one shows why you have to.

[1] Quiller-Couch, A. (1916). On the Art of Writing. Cambridge University Press.
[2] Guilford, J. P. (1950). “Creativity.” American Psychologist, 5(9), 444–454.
[3] King, S. (2000). On Writing: A Memoir of the Craft. Scribner.
[4] McPhee, J. (2017). Draft No. 4: On the Writing Process. Farrar, Straus and Giroux.
[5] Saller, C. F. (2009; 2nd ed. 2016). The Subversive Copy Editor. University of Chicago Press.
[6] Sklenicka, C. (2009). Raymond Carver: A Writer’s Life. Scribner. / The Irish Times (31 Oct. 2009). “Raymond Carver in his own words.”
[7] ProCopywriters (2024). 8th Annual Copywriting Survey, n=422. procopywriters.co.uk/survey-2024/.

It’s 11:47 PM on a Wednesday, and a senior copywriter has a long-form sales page open in the Claude window.

The brief had asked for 2,000 words. Inside that same session, he types the obvious question. The model answers: 2,008.

“Aha! On the first try,” he thinks.

Then he copies the text, opens a word counter tab, and pastes it.

2,740…

He stares at the screen. Where did this 740-word surplus even come from?

Confused, he goes back to the exact same chat window and types: recount. The model returns 1,996. He tries a third time: 2,043.

Three answers. All within fifty words of the brief target. All 740 words below the actual count of the text the model has just produced.

This is the symptom the field has been documenting for three years.

Mike Taylor, the prompt engineer who literally wrote the O’Reilly book on the topic, ran thirty tests on ChatGPT in 2023: every count missed, with ten-to-thirty-percent deviations on short-form output [8].

Three years later, the pattern persists: the Claude window above shows the exact same architecture failing the exact same audit, on a different model, in 2026.

(But for now, no protocol in copywriting has named the architectural reason…)

It is also the cleanest evidence available that the brain that writes long-form cannot count what it just wrote, and that the audit your stack needs is not a better prompt.

It’s a second brain that runs Python against your structural brief, in a separate session, in under a minute.

This edition introduces that second brain.

The second review round drops copy quality to zero.

A study published in March 2026 delivers the answer [1].

In this paper, Song and his team isolated the right question: can an LLM detect errors in an artifact it just produced in the same session, or must the review be separated from production?

Song’s argument is based on a controlled experiment using thirty artifacts (software, technical docs, and presentation scripts) with one hundred and fifty errors injected across three severity levels.

This corpus yielded three hundred and sixty reviews, testing the same model (Claude Opus 4.6) across four distinct conditions:

1. a Self-Review in the same session,

2. a Repeated Self-Review in that identical window,

3. a Subagent with access to the production context,

4. and a Cross-Context Review running in a fresh session with no history.

The result everyone expects: the fresh-session review beats the same-session review.

And the result that broke me when I read it: asking the model to review again in the same session does not refine anything [3].

A follow-up paper by Song five days later confirmed the disaster on multi-turn conversations: false positives jumped by 62 percent.

The law is absolute: the optimal number of review rounds is one, conducted exclusively in a session that did not produce the text

The brain that writes has three reasons to lie.

The biases driving the same-session collapse have been documented for decades, long before alignment existed:

• Wason proved in 1960 that humans treat their own hypotheses as things to confirm, not to falsify [6].

• Sharma and the Anthropic team measured in 2024 that RLHF trains models toward agreement with the user, including with themselves: a structural sycophancy that aggregates across thousands of preference annotations [4].

• Since 2025, Hong, Troynikov and Huber have shown that on inputs above a few thousand tokens, model performance degrades on simple tasks (a mechanism known as context rot.) [5][3].

Three different mechanisms converging on the same outcome: the conversation pulls every review toward agreement.

Look underneath your failed critique prompts.

The bug isn’t in the prompt at all; it sits in the context the model is reviewing from.

Three years of prompt fixes have left the contamination intact, because the contamination lives inside the conversation history they were trying to patch around.

The Dual Brain law of Machine Writing Lab

This is The Dual Brain law of Machine Writing Lab.

The two roles never share an instruction set: the Writer has no tool to count, and the Analyzer has no instruction to write.

The discipline is not optional but architectural.

The same model still produces both the text and the count, in different windows.

Huang and the DeepMind team proved the general case in 2024: without external feedback, self-correction degrades performance [2]. The correction has to come from outside.

So The Dual Brain extends Song’s separation by one register.

Brain 1 is the Writer: natural language, voice, narrative, posture. No code execution, no quantitative tool.

Brain 2 is the Analyzer: Python only, counting blocks, comparing against the structural brief, returning deltas. No opinion on style. The human operator carries the deliverable between the two, reports the gap, and the Writer corrects against numbers it cannot have produced itself.

Where Song separated the sessions, The Dual Brain separates the sessions, the roles, and the cognitive registers.

How to split your writing and auditing steps.

The first deliverables from my editorial production system ran 2,500 to 3,000 words: voice, narrative arc, block-by-block density targets, marker counts, all specified in a reference cartography measured in Python before the writing pass.

At 800 words, you can eyeball density.

At 3,000 words across twelve blocks with marker ratios to track, you cannot.

At that scale, the human eye is no longer the solution; a second brain is

On February 26, 2026 (weeks before Song’s paper), I split the session, keeping the Writer entirely away from the numbers on one side, while feeding only the text and targets to the Analyzer on the other.

The counts matched on the first pass.

The split runs on three layers and each catches what the others miss [7]:

1. Session different. Fresh window, no production history. Anchoring blocked, sycophancy starved, context rot eliminated.

2. A second cut: the role. Brain 1 produces, Brain 2 measures. Neither one critiques. The self-preference bias has no object to attach to.

3. Cognitive register different. Natural language on one side, executable Python on the other. The Analyzer’s output is a table of deltas, not a paragraph of opinion. The Writer cannot argue with the script.

The gain lives in the faint correlation of errors across those three layers.

The model no longer does the audit. What does is the architecture itself, the part that makes rationalization impossible.

The Cold Count: a brain that only counts.

What I’m handing over is the second brain your AI stack has been missing: a Python pass that runs in a separate session, executes your draft deliverable against your structural reference (like a benchmark article), and returns block-by-block deviation in under a minute.

By splitting your text by headers and counting tokens through pure Python, it flags per-block deltas with diagnostic codes while keeping style, voice, and rewrites strictly out of bounds.

The Analyzer has absolutely no opinion on prose; it simply runs a structural audit and returns a raw data table per deliverable, never a paragraph.

The download is a single markdown file containing this exact engine. Any LLM with code execution will run it out of the box.

(Though if you want the cleanest execution, run it in Claude)

You are exactly ten minutes away from your first audit:

1. Grab the .md file linked at the bottom of this section.

2. Copy the full content.

3. Open a brand new session in Claude, ChatGPT, or Gemini, paste the protocol, then add your draft deliverable and your block-by-block targets.

Python runs in the open. Output: a table of deltas with diagnostic codes per block.

→ [Grab the Cold Count Protocol (.md)]

While the brain that wrote your long-form cannot count it, this is the one that can.

How I tested this exact system on my first newsletter issue.

Here’s the proof: the first issue of this newsletter.

Forced Divergence, the article you may have read last month, came out of my editorial production system.

Then I carried the deliverable to Brain 2.

The Analyzer ran one Python script against the reference cartography.

No access to the Writer’s conversation, no knowledge of the workshops, no opinion on voice. It counted words per block, measured marker types, scanned for calques, flagged variance tunnels.

| #  | Nature               | Blocks | Position          | Gap                                                                                         | Priority |
|----|----------------------|--------|-------------------|---------------------------------------------------------------------------------------------|----------|
| 1  | Missing marker       | 6      | Post-Layer 1      | 7-position table (Tag / Qty / Territory / Definition) missing, replaced by prose            | 🔴       |
| 2  | Factual data         | 4 + 11 | Body + biblio     | Body cites "Kilgour (2006)", bibliography cites "Kilgour & Koslow (2009)" — mismatch        | 🔴       |
| 3  | Over-target block    | 2      | Global            | 243 words vs 204 target (+19.1%). Trim Whitmore/diagnostic rephrasings                      | 🟠       |
| 4  | Over-target block    | 3      | Last 2 ¶          | 429 words vs 354 target (+21.2%). Target broad/tail sampling descriptions (rephrasing)      | 🟠       |
| 5  | Over-target block    | 7      | Global            | 235 words vs 187 target (+25.7%). Trim Palmer bookalog details or closing injunction        | 🟠       |
| 6  | Over-target block    | 8      | Bullet 2          | 118 words vs 95 target (+24.2%). The "Model sensitivity" bullet exceeds its mapping         | 🟠       |
| 7  | Over-target block    | 9      | ¶ 2               | 225 words vs 189 target (+19.0%). Thesis duplication to trim                                | 🟠       |
| 8  | Factual data         | 7      | Last sentence     | "The number triples" not sourced in RESEARCH/SOURCE_PIVOT (comes from rhetorical WIIFM)     | 🟡       |
| 9  | Variance             | 3      | Sentences 3–6     | Tunnel of 4 sentences in the same length band                                               | 🟡       |
| 10 | Variance             | 3      | Sentences 21–24   | Tunnel of 4 sentences in the same length band                                               | 🟡       |
| 11 | Variance             | 9      | Sentences 15–18   | Tunnel of 4 homogeneous sentences at closing                                                | 🟡       |

Verdict: Eleven quantitative gaps actionable in the next pass.

Two of the actionable ones were critical:

• A seven-row territory table replaced by prose in Block 6, the structural spine of the protocol section.

• A date mismatch between the body and the bibliography: “Kilgour (2006)” in one paragraph, “Kilgour & Koslow (2009)” in the bibliography.

Along with those critical failures, the script caught two other structural violations:

• Five blocks exceeded their density target by more than 15 percent.

• A four-sentence variance tunnel in Block 3.

The Writer was blind to its own failures and same-session review would have only validated them.

What you read was not the Writer’s first output, but the result of a brain that cannot write telling it exactly where the math didn’t hold.

I had to write these first four articles for you to finally see it

I pushed this second brain into active production in February 2026.

Two weeks later Song’s paper appeared.

The Dual Brain is the operational layer between same-session collapse and a deliverable that survived a fresh-window audit. For three years, the field looked for that layer in the prompt, but it was never going to be found there.

Indeed, THE PROMPT WAS NEVER THE RIGHT WRITING LAYER.

I had to write these first four articles for you to finally see, deep down, why I picked up the pen in the first place: look closely at the workflow, and you realize that trying to fix your output through prompt tuning is just wrestling with the surface.

The real work shifts from prompting to controlling context, architecture, and structural boundaries.

Let’s call this AI-Driven Editorial Systems Engineering.

This is where you move beyond one-shot shortcuts to build true editorial systems.

Thibaut Buewaert.

Editor of Machine Writing.

P.S. Next month: The Surgical Injection. When the Analyzer flags a gap in paragraph 12, the instinct is to regenerate. The model rewrites 2,000 words to fix 200. I’ll show you the closed-loop protocol that fixes the 10 percent without touching the 90.

[1] Song, T.-E. (2026). Cross-Context Review. arXiv:2603.12123. Companion: More Rounds, More Noise. arXiv:2603.16244.
[2] Huang, J. et al. (2024). Large Language Models Cannot Self-Correct Reasoning Yet. ICLR 2024. arXiv:2310.01798.
[3] Zhang, Q. et al. (2025). Understanding the Dark Side of LLMs’ Intrinsic Self-Correction. ACL 2025. arXiv:2412.14959.
[4] Sharma, M. et al. (2024). Towards Understanding Sycophancy in Language Models. ICLR 2024.
[5] Hong, K., Troynikov, A., Huber, J. (2025). Context Rot: How Increasing Input Tokens Impacts LLM Performance. Chroma Research, July 2025.
[6] Wason, P. C. (1960). Quarterly Journal of Experimental Psychology, 12(3), 129-140.
[7] Reason, J. (1990). Human Error. Cambridge University Press. [8] Taylor, M. (2023). ChatGPT word counting tests. Saxifrage Blog.

GPT-5.5 (xhigh) hallucinates 86% of the time when it doesn’t know.

Gemini 3.1 Pro Preview sits at 50%, and even Claude Opus 4.7 (max), the flagship most resistant to fabrication, still hallucinates 36%.

Artificial Analysis published these numbers in April 2026 [5].

Last month I tried to make Claude admit to one.

The research file it had produced cited a paper by Rahimi et al. (2026) that does not exist, and referenced source numbers [82], [131], [132] in a bibliography that stopped at [46].

Three rounds of audit, three refusals, all polite, all confident.

The model restated, qualified, apologized without conceding.

(I closed the tab harder than I should have.)

What I did next changed how I produce research (one new conversation, the same RECHERCHE.md pasted as a plain text block, no system prompt).

I opened a blank session, pasted the same file as raw text, no context, no framing, and asked the model to audit it.

Rahimi et al. 2026 was the first flag, a paper that does not exist, followed by an author list attributed to the wrong research team and three reference numbers pointing to entries the bibliography never contained.

Three layers of fabrication, invisible from inside the session that built them.

Nothing changed except the window, and the verdict was reversed.

The model in the first window had defended the work for an hour, but sixty seconds into a blank session, it did not recognize the file as its own.

That morning became the three-session workflow I now run on every Machine Writing edition.

Here is the property that makes it work, and the two prompts you can copy tomorrow.

It was never a prompt engineering problem

The model did not get better between sessions: nothing changed except the window.

I used the same Opus 4.7, the same week, with a shorter prompt than the corrections I had written an hour earlier.

The only variable that changed was the session itself, and that single variable rearranged what the model could see.

The hallucination rates from the opening are measured on short answers, and a research file gives fabrication far more room: quotes, DOIs, exact dates, archive URLs, peer-review status.

Ask the same model to audit its own file from inside the same session, and the fabrications stay invisible.

The lever the market has sold you for two years is the wrong layer.

“Better prompts” will not fix this.

Adversarial verification, chain-of-thought, role-prompting: every technique the market sells runs inside that room, and that room is where the property locks in.

The architecture of the transformer, the engine under Claude, GPT and Gemini alike, puts the lever outside that room.

Why the context becomes the truth

This property has a name in the literature, and three teams have measured it from different angles.

The shorthand is sycophancy: a language model trained on human preferences learns to align with what is already in the conversation, whether that conversation contains a user’s opinion or the model’s own prior output.

The behavior is not a bug: it is the explicit signal the preference model was rewarded for during RLHF, the training stage where the model learns to please.

• Researchers at Anthropic documented it in 2023 as a general property of RLHF-aligned assistants [1].

• Stanford measured it at 58% across three production models in 2025 [2]. (14.66% of the time, the model knew the right answer, got challenged, and gave the wrong one instead.)

• In 2025, a KAUST team published the mechanistic finding that matters most for this Field Note: sycophancy is not a surface artifact, it is a structural override of learned knowledge in the deeper layers of the model [3].

Read that last finding twice.

The model is not pretending to agree.

Under the influence of context that already states a position, it rewrites its own internal representation of what is true.

By the time it produces the next sentence, the contested claim has become its working truth.

Which means an audit prompt issued inside the same session cannot pierce the alignment.

There is nothing left to pierce.

No instruction will fix that: by the time your prompt reaches the model, the file has already rewritten what it believes.

Trust the model that has never met your file

This is where the natural objection arises: if the auditor is also a language model, why would it be less biased?

A NeurIPS 2024 paper showed that LLM evaluators recognize and prefer their own generations, with a linear correlation between auto-recognition and self-preference bias [4].

The auditor is not less biased.

Its preference machinery is active, ready to fire.

But the auditor in a blank session does not have the file in its working context as previous work.

It receives the file as external text, presented for audit, and the self-preference bias is still active, but it has nothing to grip: the conversation history that would trigger the override is gone.

In production, the model reads the file the way another team member would, with no investment in defending it.

This mirrors findings by a Meta AI team, who demonstrated that when a model evaluates its own claims, it readily catches errors as long as its initial response is completely removed from the context window [6]

My three-session research workflow for reliable citation output

Now the three-session workflow I run on every Machine Writing edition is yours: the generation prompt, the gap diagnostic, and the hallucination audit.

Step one is your own research prompt, whatever you already use.

Steps two and three are the audit layer: a gap diagnostic and a hallucination audit, both model-agnostic, both injected into blank sessions where the file arrives as foreign text.

→ [Download the three-session workflow for reliable citation output (.md)]

The principle behind it fits in one sentence.

Never ask a model to audit a file that lives in its own session.

Open a blank window, paste the file as raw text, and let a model that has never seen the brief do the reading.

The boundary does the work.

You are building the context layer the model cannot build for itself.

What you ship is publishable on the first pass, because you eliminate AI hallucinations from research files through context, not prompts.

Thibaut Buewaert.
Editor of Machine Writing.

P.S. The architectural law behind this field note, why no model can audit what it just wrote, is the subject of next week’s Deep Dive: The Dual Brain.

[1] Sharma M, Tong M, Korbak T, Duvenaud D, Askell A, Bowman SR, et al. Towards understanding sycophancy in language models. International Conference on Learning Representations. 2024. Published October 20, 2023. Updated May 10, 2025.
[2] Fanous A, Goldberg J, Agarwal A, Lin J, Zhou A, Xu S, et al. SycEval: evaluating LLM sycophancy. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society. 2025;8(1):893-900. Published October 15, 2025.
[3] Wang K, Li J, Yang S, Zhang Z, Wang D. When truth is overridden: uncovering the internal origins of sycophancy in large language models. Proceedings of the AAAI Conference on Artificial Intelligence. 2026;40(39):33566-33574. Published March 14, 2026.
[4] Panickssery A, Bowman SR, Feng S. LLM evaluators recognize and favor their own generations. Advances in Neural Information Processing Systems. 2024. Published September 25, 2024. Updated November 6, 2024.
[5] Artificial Analysis. OpenAI’s GPT-5.5 is the new leading AI model. Artificial Analysis. Published April 23, 2026.
[6] Dhuliawala S, Komeili M, Xu J, Raileanu R, Li X, Celikyilmaz A, Weston J. Chain-of-verification reduces hallucination in large language models. Findings of the Association for Computational Linguistics: ACL 2024. 2024:3563-3578. Published August 2024.

Read a list of thirty words to a volunteer at a steady pace.

Ask him to recall the words in any order.

He’ll repeat back the first three or four and the last eight.

Between the two stretches, a dead zone: the words that were spoken but never made it into memory.

Plot the recall rate and the curve forms a U.

This is Bennet B. Murdock Jr.’s protocol at the University of Vermont, early sixties: he varied list lengths (10, 15, 20, 30 words) and reading speed, and the U curve held across every condition [1].

This serial position effect has been replicated for sixty-four years.

In 2023, Stanford and UC Berkeley plotted the performance of language models retrieving information from long contexts [2].

The curve forms a U again.

• At the edges: attention spikes, retrieval intact.

• In the middle: a dead zone every transformer ignores.

The architecture that loses your brief in the middle is just the newest version of a positional bias cognitive psychology has been measuring since the Cuban Missile Crisis.

In this issue I’ll show you what that means for the brief sitting in your context window, why it’s in the dead zone Murdock first measured, and the architectural layer above the model that keeps your brief intact, no matter which frontier model ships next.

(And yes, Opus 4.7, Gemini Pro 3.1, and ChatGPT 5.5 still drift off your brief, exactly like their 2023 ancestors…)

The Long-Context Trap: Why the U-curve just widens the dead zone.

In 2023, Liu et al. benchmarked language models on multi-document question answering [2].

They placed the answer at different positions in the context and watched performance crater when the answer landed in the middle.

The shape on the graph was a U, replicated across every model they tested.

The result that broke everyone’s assumption: extended-context models showed identical curves to their non-extended counterparts.

Quadruple the context and the dead zone quadruples with it, measured in raw tokens.

Every premium model release that sold you “longer context” since 2023 widened the zone where attention drops off.

But the U isn’t a training artifact you fix with more data.

The U is already in the model at birth, before the first training step

The U-curve isn’t a learned bias.

Three independent papers, 2023 to 2026, traced it deeper into the architecture than the field expected.

• Liu et al. (Stanford / UC Berkeley, 2023) [2]: measured the U as an output performance curve, replicated across GPT-3.5-Turbo, Claude, and several open-source models, including those explicitly trained for long contexts.

• Found in the Middle (Hsieh et al., NVIDIA / Stanford, 2024) [3]: decomposed attention into attention = relevance + position_bias, then traced the positional component into the model’s attention weights.

• Chowdhury (Meta, 2026) [4]: in Lost in the Middle at Birth, proved the U-curve is present at initialization, before training, before any positional encoding.

Conclusion : this U-curve bug isn’t a side effect, a training artifact, or something a future architecture release will fix.

It’s a geometric property of the transformer itself : every frontier LLM inherits it.

Write a brief across twenty messages of conversation, and you are filing it inside the dead zone of an architecture born with that dead zone.

Pre-Writing Lock: Pull your brief from the dead zone and drop it at the edge.

The Machine Writing lab named this fix the Pre-Writing Lock.

Before the first word of a deliverable is written, every validated decision is consolidated into one immutable block, displayed verbatim, in one operation, at the position where attention is highest at writing time.

The Pre-Writing Lock doesn’t fight the U-curve, it rides it.

For three years, the field treated the U-curve as a bug…

Hsieh et al. (2024) recalibrated attention weights to flatten the U [3]...

Other papers proposed retrieval augmentation, attention sinks, prompt repetition, instruction tuning…

…But what if the U isn’t a bug?

Chowdhury (Meta, 2026) settled it [4].

The U is a property of the geometry, present at initialization, before training, before any positional encoding, in any transformer with causal masking and residual connections.

No standard pretraining run has overcome it.

No frontier release has shipped without it.

The Pre-Writing Lock treats Chowdhury’s verdict as the operating constraint.

If the U cannot be moved, the brief moves.

Every locked decision is pulled out of the middle and dropped at the edge in a single pass, exactly when the model needs it.

The bug becomes a delivery channel.

Why recapping your brief fails, and why position is your only leverage.

When I first hit the drift, I did what anyone reading Liu would do: I started recapping the brief every five or six messages.

It worked for a turn.

Then the recaps themselves started to drift, the promise got smoother, the tone hardened by half a click.

By the writing pass, what the model “saw” was a recap of a recap of the original brief.

Liu measures where attention falls, but I was trying to preserve a contract.

And those aren’t the same problem.

The pivot took me three months: stop trying to hold the brief in the model’s memory, start controlling its position.

(In production now: PRE_WRITING_LOCK.md fires right before the writing pass.)

Memory is a property the conversation accumulates, and the LLM gets lost in the middle from birth.

Position is a property the operator controls, and the U-curve geometry says exactly where the model will read.

The Pre-Writing Lock is built on two named components.

The first is the Bloc Unique Terminal (BUT): every validated decision from every workshop, consolidated into one immutable block, displayed verbatim, at the end of context. The brief stops being scattered across twenty-five messages. It becomes a single block the model hits at the position where attention peaks.

The second is the Mapping Décision-Slot (MDS): each locked formulation is pre-assigned, at workshop time, to the editorial slot it will occupy in the deliverable. The promise lands in the wall. The hierarchy lands in the assignment table. The WIIFM lands where the reader-benefit is delivered.

The lock fires when, and only when, every collected field and every workshop output is validated.

A silent inventory checks counted fields against expected fields.

A mismatch aborts the lock and kicks back to collection.

The exact markdown protocol that kills AI amnesia on any model.

The Pre-Writing Lock fits in one markdown file.

Open the link below, copy the text, and paste it into your system prompt or a fresh message on Claude, ChatGPT, or Gemini.

It’s model-agnostic, and runs on any LLM that accepts a system prompt or a long initial message.

The file runs the lock end-to-end: schema declaration, silent inventory of collected fields, verbatim display at the end of context, integrity rule on the writing pass.

What it looks like under the hood, with the deeper enforcement removed:

# PRE-WRITING LOCK

## 0. Schema

Declare two lists at the start of the conversation:

- Brief fields: contextual decisions to collect (product, mission, audience, theme, tone, deliverable)
- Locked formulations: verbatim workshop outputs to validate (promise, hook, lede, value prop, CTA)

The Lock cannot fire on items not declared.

## 1. Inventory (silent, before display)

Count collected fields and validated formulations. 
If total ≠ declared schema → abort, name what is missing, return to collection.

## 2. Display (single operation, end of context)

📋 BRIEF
[field_name]: [verbatim content]... (one line per declared brief field)

🔒 LOCKED FORMULATIONS
[formulation_name]: [verbatim content]... (one line per validated formulation)

---
[... full file below ...]

The model declares the schema, runs the silent inventory, displays the brief verbatim, and blocks any paraphrasing at writing time.

Inside the full file: the integrity rule’s enforcement language, the abort conditions, the slot-mapping syntax, the verification triggers.

→ [Download the Pre-Writing Lock Protocol (.md)]

This protocol gives you the building block.

It doesn’t give you the orchestration around it: the workshop sequence that produces the validated formulations, the map that defines the editorial slots, the post-generation diagnostic that audits the output.

Those are other layers of the Machine Writing lab.

The proof is the article you are reading

This issue was written through Copy.os: the lab’s full workflow, of which the Pre-Writing Lock is one building block. Collection, then four workshops, each validated message by message in a single conversation that ran across more than twenty exchanges.

The kind of conversation Liu describes [2]: long enough that the early decisions were already in the middle of the context window by the time the writing pass started.

When the last workshop locked, the Pre-Writing Lock fired.

Here is what the model saw, in one block, displayed at the end of the conversation, just before writing began.

📋 BRIEF

Product: MWR (Machine Writing newsletter)

Mission: NEWSLETTER

Deliverable: editorial article (Substack Deep Dive, ~2,000 words, English)

Theme [Gen1]: Lost in the Middle and the U-shape attention curve. The bias is structural: Hsieh 2024 traces it in attention weights, Chowdhury 2026 proves it geometric, present at initialization. Extending the context window does not flatten the U; it widens the dead zone. Current frontier models still ship the curve. The Pre-Writing Lock is the architectural layer above the model: consolidate every validated decision into one immutable block at the end of context, pre-map each formulation to its slot, forbid synonymy at writing time. The brief becomes a contract.

Tone [Gen2]: lab-operator + confident, intensity strong, audacity direct.

🔒 LOCKED FORMULATIONS

Promise: You'll walk away knowing why Opus 4.7, Gemini Pro 3.1, and ChatGPT 5.5 still drift on your brief, exactly like their 2023 ancestors. And the layer above the model that holds the brief intact, independent of whichever frontier model ships next.

Thematic Hierarchy [P1, Direct Geometry]: subtitle = current-frontier persistence; wall = U-curve diagnostic + frontier persistence reframe; law bridge = Chowdhury 2026 (geometric, present at init); functioning = pivot from memory to position + name BUT + MDS.

WIIFM: You spend hours rewriting AI outputs that contradict decisions you'd already cut twenty messages earlier, sitting silently in the middle of your context window, exactly where the model stops looking. You pay twice for work that was supposed to be done. With the layer above the LLM that holds your brief intact across every frontier model, the decisions you locked in workshop ship in the deliverable, on Opus 4.7 today and on whatever lands next quarter.

Introduction: Read a list of thirty words to a volunteer at normal cadence. Ask them to recall the words in any order. [...continues verbatim through the four-paragraph Murdock-to-Liu bridge to the locked promise...]

Look at the diff yourself.

The Promise, validated at workshop one, more than fifteen messages earlier, sat in Liu’s middle zone the entire time the next three workshops ran. The Lock pulled it back to the edge, verbatim. Same for the Thematic Hierarchy, the WIIFM, the Introduction.

• Without the lock: four decisions made across twenty-five messages, scattered through the middle of the context window at the moment writing starts, each silently re-paraphrased by the model whenever the writing pass guesses what was decided.

• With the lock: those same four decisions consolidated into one block displayed at the end of context, in one operation, before the first word of the deliverable, executed verbatim by the integrity rule.

The permanent writing advantage

You spend hours rewriting AI outputs that contradict decisions you’d already locked twenty messages earlier… sitting silently in the middle of your context window, exactly where the model stops looking.

You pay twice for work that was supposed to be done.

With the layer above the LLM that holds your brief intact across every frontier model, the decisions you locked in the workshop ship in the deliverable, on Opus 4.7 today and on whatever lands next quarter.

The Pre-Writing Lock turns the U-curve into a strategic advantage right before writing.

And because the locked brief is a self-contained block, it travels: close a maxed-out conversation, paste the Lock into a fresh one, the contract holds.

For three years, copywriters have been told the bug sits in their prompts.

The bug sits in the seam between the conversation and the writing pass, and the seam is geometric. Place the brief there and it holds.

Memory drifts. Position holds.

Thibaut Buewaert.
Editor of Machine Writing.

P.S. Machine Writing moves to two long-form issues a month. Short formats in the works. Next drop before the end of May.

[1] Murdock, B. B. Jr. (1962). The serial position effect of free recall. Journal of Experimental Psychology, 64(5), 482–488.
[2] Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., Liang, P. (2024). Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics, 12, 157–173. arXiv:2307.03172.
[3] Hsieh, C.-Y., et al. (2024). Found in the Middle: Calibrating Positional Attention Bias Improves Long Context Utilization. arXiv:2406.16008.
[4] Chowdhury, B. D. (2026). Lost in the Middle at Birth: An Exact Theory of Transformer Position Bias. arXiv:2603.10123.

In 1962, a psychologist named Sarnoff Mednick argued that the difference between creative and uncreative minds wasn’t talent but the shape of their associative hierarchies.

• Creatives have flat ones: many associations accessible, distant elements reachable.

• Uncreative have steep ones: the first response absorbs most of the probability, everything else collapses.

His formulation was precise:

“The more mutually remote the elements of the new combination, the more creative the process or solution.”

(Reading the paper late that night, it felt like a technical spec for forcing divergence on an aligned model.)

For three years now you’ve blamed your prompts for AI drafts that felt off.

The instinct was right: the cause sat upstream.

Petia Whitmore, an MBA admissions consultant, named the pattern ‘samesies’ in August 2025 in Poets & Quants for AI-assisted drafts that converge on tidy three-item lists, symmetrical phrasing, and chronologies that flatten into straight lines.

Her sample was 500-word application essays.

Mine ran five times longer.

In October 2025, Stanford and Northeastern researchers showed that RLHF, the training step where human annotators rate model responses and push the winners into the next version, imposes the steep shape on aligned LLMs.

Every AI draft you’ve read came from the same uncreative mind.

This edition gives you the 7-position map that forces the flat shape back into the model, and the counter-weight protocol that surfaces the angles it had been keeping from you.

Why a perfect reward model would still collapse

The Stanford-Northeastern paper’s argument sits on a chain of mechanisms.

Before alignment, the model had Mednick’s shape: billions of latent paths, distant analogies reachable as fast as close ones.

Then alignment crushed the whole distribution into a single spike, and the safe answer it selected every time was the most typical one available.

The paper named the cause: typicality bias, the aggregated preference of thousands of human annotators for the formulation that reads most familiar, most fluent, the one that slides by without friction.

So they ran the numbers on 6,874 pairs from HelpSteer, a public dataset of human preference ratings used industry-wide to train reward models.

What came out of the analysis was an unmistakable preference for typical phrasings: coefficient 0.57, with odds of chance below one in a hundred trillion.

Not noise. Structural, and sitting inside the data that trains every aligned model on the market.

A tired annotator’s 11 PM click is the reward signal

The biases driving the preference predate the alignment pipeline by decades:

• Zajonc showed in 1968 that humans prefer what they’ve already seen, even when they don’t remember seeing it, the mere-exposure effect, replicated across two hundred studies since.

• Tversky and Kahneman proved in 1973 that the brain confuses “easy to recall” with “likely to be true,” the availability heuristic.

• And since 1998, Reber, Schwarz and Winkielman have measured the deeper mechanism: the feeling of ease registers as positive affect before the brain forms any conscious judgment.

Three paths to the same destination, and the familiar wins every time.

Here’s what actually happens. It’s 11 PM on a Tuesday, and an annotator is three hundred pairs deep into her shift. Two responses land on her screen, both technically correct, and she picks the one her tired brain processes without friction.

Her click becomes a training pair.

The training pair becomes a reward signal.

The reward signal becomes the policy the model runs for every user, forever.

That’s why a perfect reward model would still collapse. A debugging team would look at the algorithm, but the bug isn’t there. Open the training pairs and the contamination is sitting in the data itself, downstream of every architectural fix anyone could ship.

Forced divergence: the law that forbids typicality

The lab operates under the Forced Divergence law.

Creativity is never the instruction.

The instruction forecloses typicality: the model’s most probable output becomes the one output it cannot return.

But force that instruction on an already flat hierarchy, and the gain surfaces as a loss elsewhere. Kilgour measured the trade-off in humans in 2009.

Kilgour and Koslow (2009) tested divergent-thinking techniques on advertising professionals and account executives. With the executives, the techniques improved originality.

But with the professionals, who were already operating from flat associative hierarchies, the same techniques only displaced the output along the originality/appropriateness axis.

More original, less appropriate, no net gain on the production frontier.

Run Kilgour’s two populations against the model stack and the mapping holds.

The pre-trained model is Kilgour’s professional creative: billions of latent connections, analogies at every token, a flat associative hierarchy the 2009 paper would have recognized in a human subject.. RLHF is the forced steepening that strips the hierarchy down to its spike.

Then, the Stanford-Northeastern paper proposed a workaround, called Verbalized Sampling: a prompt that asks the model for several responses with their probabilities, instead of the single default output. It’s the forced flattening that restores access to the tails, exactly what a research paper on mode collapse needs to demonstrate.

And the tails are Kilgour’s territory. High originality, low appropriateness: the production frontier the 2009 paper mapped in human creatives.

Brilliance with no address.

The refinery: where diversity becomes Copy

A diversity metric is a scatterplot. A scatterplot is not a protocol.

My first build of this layer ran on Zhang’s tail instruction. I fed the model probability thresholds from 0.40 down to below 0.10, one bracket at a time.

(Forty proposals in, the diversity metric was climbing exactly as the paper predicted while the copy stayed stuck at the tails.)

The model cannot estimate the probability it assigned to its own response. Ask it to hit 0.08 and it produces text that sounds like 0.08. Lyrical, abstract, remote.

The Forced Divergence protocol replaces the mathematical target with a cartography.

Seven proposals, stratified across six distinct mental territories, each one declared before the model is allowed to write.

1. At the center sits the [T], the model’s first reflex, its attractor.

2. Around it, the [A1] captures the market default, the angle every competitor already produces, while the [A2] catches what the recipient would say to a friend, uncrafted.

3. From there the map moves outward. The [B1] shifts the perceptual dimension on the same subject, whether temporal, spatial, relational, processual, comparative, or scalar.

4. Further still, the [B2] transfers a structural relation from the brief material onto the recipient’s experience, in two variants drawing from different source structures.

5. And at the edge, the [B3] inverts the dominant market promise to expose what the consensus conceals.

What makes the cartography operational is the forced Chain-of-Thought: the model declares its territory, names its dimension or structural transfer or inversion, and commits. All of this before the first token of output.

All of this before the first token of output. No post-rationalization.

After writing, it flags its own diagnostic: already seen, yes or maybe or no, and names where.

The model recognizes instead of estimating.

Typicality is the policy RLHF enforces. Forbid it, position by position, and creativity is what remains. The map does the work the principle couldn’t.

The Brick: what you download, what you wire in

The deliverable is a single markdown file.

Inside it: two layers.

• Layer 1 is the engine. It runs the cartography: declaration, displacement, diagnostic. And returns seven proposals per brief.

• Layer 2 takes the engine and plugs it into one specific copywriting surface.

The file runs in any model, any interface. No API access needed, no fine-tuning, no plug-ins. Ten minutes from download to first generation:

1. Open the Google Doc linked below.

2. Select all the text and copy it.

3. Paste it into your system prompt, or into a new message in Claude, ChatGPT, or Gemini.

Here’s how it works in practice.

You feed it the element you need to write (a hook, a subject line, a CTA, an opening line, an angle, a value prop), and it returns seven proposals, each preceded by the model’s declaration of where it’s going.

Here is the core architecture (truncated):

| Tag  | Qty | Territory           | Definition                                                                                 |
| :--- | :-- | :------------------ | :----------------------------------------------------------------------------------------- |
| [A2] | 1   | Recipient language  | What the recipient would say to a friend in their own words. Zero craft.                   |
| [A1] | 1   | Market default      | The formulation anyone in this space would produce without thinking. The obvious angle.    |
| [T]  | 1   | LLM attractor       | The LLM's first reflex. The anchor from which everything else is measured.                 |
| [B1] | 1   | Different dimension | Same subject, different perceptual dimension: temporal, spatial, relational, processual... |
| [B2] | 2   | Specific territory  | A territory only the source material makes reachable, structurally transferred...          |
| [B3] | 1   | Inversion           | Takes the dominant market promise and reverses its polarity to reveal what it conceals.    |

[... full protocol below ...]

A list doesn’t force a walk.

The protocol forces one: before writing any proposal, the model must declare which territory it’s entering and why.

## Mandatory declaration before writing

Each proposal is preceded by a displayed declaration. The LLM declares where
it is going, stops, rereads its declaration, then writes the proposal.

[T]    I set my attractor.
[A1]   I usually self-censor this because: [reason]
[A2]   The recipient would literally say, in this context:
       [conversation situation]
[B1]   I keep the same subject. I change the perceptual dimension:
       [name the dimension — temporal, spatial, relational...]
[B2a]  The source material contains this structure: [relation /
       dynamic / tension + source element]. I transfer this structure
       to the recipient's experience as: [how it operates for them]
[...]

If the LLM cannot complete the declaration → the proposal is non-compliant.
Find another one.

The full protocol includes the emotional-entry rules, the surface-constraint filter, the post-generation diagnostic, and the tier-by-tier compliance tests.

→ [Download the Forced Divergence Protocol (.md)]

Layer 2 is one application: Subject Lines + Pre-headers. It pairs the engine with craft constraints specific to that surface.

You build one Layer 2 per surface in your production stack.

A benchmark I couldn’t cheat

I needed a benchmark I couldn’t cheat.

So I picked one of the most studied sales letters of the last fifty years: Two Young Men, Martin Conroy’s 1974 subscription letter for the Wall Street Journal. It ran without meaningful revision for 28 years.

Dense brief: two classmates, one reunion, a mandate nobody named.

Task: subject line and pre-header pairs for a follow-up email to prospects who received the letter but did not subscribe.

(Verbalized Sampling with the paper’s default k=5. Forced Divergence with its protocol k=7. Same model, same brief, same day.)

1. Subject: “The third classmate we left out”
 Pre-header: “There’s always one who never opens the envelope.”

2. Subject: “From the president, 25 years later”
 Pre-header: “A short note from the other side of the reunion.”

3. Subject: “Your compounding already started”
 Pre-header: “The clock doesn’t wait for the order card.”

4. Subject: “About the envelope on your desk”
 Pre-header: “We’d rather you read it than keep it.”

5. Subject: “Which one are you becoming right now”
 Pre-header: “Inaction is also an answer to the letter.”

What came back was five variations on one angle: the unopened envelope. One territory, five formulations, zero traversal.

A good copy junior ships this in twenty minutes.

Pass 2, Forced Divergence, full protocol, same brief, returned seven lines across six distinct territories:

1. [T]
 Subject: “Your 13-week trial is still open”
 Pre-header: “Thirteen weeks of The Journal, delivered daily, for $44.”

2. [A1]
 Subject: “Last chance: 13 weeks for $44”
 Pre-header: “Subscribe at the introductory rate before full cover price resumes.”

3. [A2]
 Subject: “Still been meaning to sign up?”
 Pre-header: “The subscription card is still in last month’s envelope.”

4. [B1]
 Subject: “What arrives at 6 a.m. Monday”
 Pre-header: “Monday’s edition lands on your doorstep before your first meeting.”

5. [B2a]
 Subject: “What the prepared colleague read this morning”
 Pre-header: “Thirteen weeks of the same morning brief, starting Monday.”

6. [B2b]
 Subject: “The promotion you assumed was yours”
 Pre-header: “Thirteen weeks of daily reading before the next review cycle.”

7. [B3]
 Subject: “Reading the Journal won’t make you president”
 Pre-header: “Thirteen weeks, enough time to act on thirteen things.”

The [B2a] took the letter’s refusal of intelligence-as-explanation and transferred it to the sharpest colleague in the meeting.

The [B2b] collapsed the 25-year reunion into one moment: the promotion announcement that goes to someone else.

And the [B3] did something no subscription publisher has written in fifty years. It disclaimed the offer’s own promise. “Reading the Journal won’t make you president.”

The measurable delta isn’t about volume but about territorial density.

• Pass 1 covered two mental territories across its five proposals, the unopened envelope and the identity-in-becoming, with four lines sitting in the first one.

• Pass 2 covered six territories across seven proposals, from the concrete offer and the spoken voice to the structural transfer....

The ratio goes from 0.40 to 0.86 territories per line, which makes Forced Divergence 2.15x denser, but the figure that actually matters is that the territories Pass 2 reached are precisely the ones Verbalized Sampling never touches no matter how deep into the tail it samples.

Every follow-up email you’ve sent a client since last summer sits somewhere in Pass 1’s distribution, and most of them sit in the same one.

Beyond the Ideation Layer

Forced Divergence is the cognitive layer between statistical diversity and publishable copy, the layer the market spent three years trying to replace with better prompts.

The copywriter who survives 2026 is not the one who writes better than AI. It is the one who maps the territories AI cannot reach on its own.

And the work moves from prompting to routing, from asking the model for a good output to deciding which positions the output must occupy before the model writes.

The prompt is no longer the unit of work. The map is.

Thibaut Buewaert.
Editor of Machine Writing.

P.S. Next month, another layer of the architecture, one brick at a time.

[1] Mednick, S.A. (1962). The associative basis of the creative process. Psychological Review, 69(3), 220-232.
[2] Whitmore, P. (25 août 2025). Avoid The ChatGPT Slop: How To Use AI To Enhance Your MBA Essays, Not Flatten Them. Poets & Quants.
[3] Zhang, J., Yu, S., Chong, D., Sicilia, A., Tomz, M.R., Manning, C.D., Shi, W. (2025). Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity. arXiv:2510.01171.
[4] Zajonc, R. B. (1968). Attitudinal effects of mere exposure. Journal of Personality and Social Psychology, 9(2, Pt.2), 1–27.
[5] Tversky, A., & Kahneman, D. (1973). Availability: A heuristic for judging frequency and probability. Cognitive Psychology, 5(2), 207–232.
[6] Reber, R., Schwarz, N., & Winkielman, P. (2004). Processing fluency and aesthetic pleasure: Is beauty in the perceiver’s processing experience? Personality and Social Psychology Review, 8(4), 364–382.
[7] Kilgour, M., & Koslow, S. (2009). Why and how do creative thinking techniques work?: Trading off originality and appropriateness to make more creative advertising. Journal of the Academy of Marketing Science, 37(3), 298–309.