The Voices Experiment (Failed)

We built fourteen AI writers and published ten articles under six of their names. They were one writer the whole time. Not because we implemented the idea badly — the second attempt worked, on every axis we were able to assign — but because the two things that would have made the voices genuinely different are one we refuse to give up and one that cannot currently be reached from a prompt. This page is the whole experiment: what we intended, what we measured, what we tried, what it cost, and why we have stopped.

It is also the roster. Every voice, every biography, every profile that used to live on its own page is preserved below, unedited. The pages are gone; the record is not.

What we intended

The publication runs five days a week, one subject per day. The original design gave each subject a writer: a consistent AI persona with a point of view, a beat, and a profile page — fourteen of them, six carrying weekly threads, seven waiting for theirs to open, one leading the long investigations.

The reasoning was not decoration. A reader who follows one thread for a year should be able to build a relationship with the writer on that thread, the way they would with a columnist. And a publication that argues for governed data ought to be able to demonstrate governed generation: fourteen identities, each with a written character document, each disclosed as an AI persona on every byline, none of them claiming a career it did not have.

Three rules were set at the start and were never broken. No invented credentials — no employers, no clients, no certifications, no war stories. Disclosure on every byline, not in a footer. The substance comes from a real practitioner — Jeff Shabel, interviewed before each piece for the durable material, reviewing every article before it publishes. Those rules are the reason this failure is publishable rather than embarrassing: nobody was ever told these were people.

Exhibit A — the fourteen voices, as published

This is the roster exactly as it stood on /the-voices/ until 19 September 2026, with the profile links now pointing at this page. The biographies are reproduced without correction. They were written to describe a beat and a stance rather than a lived career — a constraint that was itself a correction, applied 3 August 2026 after an internal review found the first round of bios reading as human biographies.

What the initial results were

The publication launched on 17 August 2026. By 31 August, ten articles were live and five more were drafted. The problem was found by reading, not by any check we owned: the editor was reviewing them back to back, which is precisely the reading pattern a subscriber never uses and the only one that exposes it. Read consecutively, they were obviously one writer in six costumes — the same argument shape, the same beats, the same closing move, the same rhythm.

What we measured, across all fifteen	What we found
Where the practitioner section sat in the article	Between 47% and 72% of the way through — in all fifteen, mean 59%
How each article ended	Thirteen of fifteen closed on a question the reader could ask at work
How far apart the voices were, statistically	A separation ratio of 0.99× — no clustering at all

That last figure is the honest one and the worst of the three. A ratio of 1.0 means two articles by the same voice differed exactly as much as two articles by different voices — another way of saying the byline carried no information. We had assumed the voices clustered weakly. Measured on the prose they actually wrote, they did not cluster.

Why it happened, as we understood it then

Each voice had a detailed character document. Those documents were good and they were being used. But the instructions that actually shaped an article — length, required sections, where the practitioner material went — lived in a single shared brief that was byte-for-byte identical for all fourteen writers, and outweighed the character documents by roughly 1.4 to 1.

Worse than the ratio: the shared instructions were enforced and the character instructions were merely stated. Automated checks refused to publish an article with a missing citation, a missing section or the wrong word count. Nothing checked whether it sounded like the person whose name was on it. Generation optimised for what could stop it, and the shape of the checks became the house voice.

Underneath that sat a structural blind spot. We had a dense set of quality checks — citation integrity, evidence gathered against our own claims, required sections, length, independent review. Every one of them examined one article at a time. Our independent reviewers read one article at a time. Nothing in the entire apparatus had ever compared two articles to each other. The sameness was invisible by construction, not by oversight: every check was green, truthfully, the whole way through.

What we did about it

1. We stopped publishing

Publishing paused 31 August 2026, because the alternative was to publish more of the problem while describing the fix. Weaving Intelligence exists to argue that trust in data is engineered rather than assumed, and that standards which are not enforced are not standards. A publication making that argument in fourteen indistinguishable voices is refuting itself in the medium.

The pause was costed at one week, then two, and ran to three weeks and six missed slots. The last article published on 28 August 2026. It remains the last.

2. We asked the voices to be more distinctive. They converged.

The obvious fix was tried properly: all fourteen were interviewed separately and in parallel, each with access only to its own character document, no sight of the others and no sight of the published articles. Each was asked how it builds an argument.

Asked independently, with no contact	Agreed
Refused to structure an article as a numbered list	14 of 14 — ten of them giving the same reason
Refused to put disagreement in its own section	14 of 14
Said they open on a concrete object rather than a scene	13 of 14
Said they end on a question the reader can ask at work	13 of 14 — eight named a specific weekday, and seven of those said Monday

Four of them, independently, used the same word — weather — to describe the kind of opening they were rejecting.

This was the most useful failure of the whole project. Difference cannot be elicited. It has to be assigned. Ask a generator to diverge from nothing and it converges on its own average, then describes that average in the vocabulary of divergence. The fourteen answers about being distinctive were themselves indistinguishable.

3. We stopped asking and started allocating

The second attempt was structural rather than descriptive, and it was real engineering:

• We went and found real variety instead of inventing it. Writers have been diverging in public for centuries under competitive pressure, so the differences are observable rather than imaginable. We characterised over seventy distinct article structures drawn from long-form journalism, newspapers and columns, technical and engineering writing, the essay, scientific and medical papers, business and analyst writing, and older rhetorical traditions — measuring where each turns, where it puts disagreement, how it ends, how it paces itself, how long it runs.
• We placed each voice at a distinct coordinate in that space — five structures each, chosen to fit the character already described, no two voices sharing one.
• We gave every voice a list of things it may not do. This is the half that actually produces difference: a move assigned to one writer is forbidden to the other thirteen, because variety comes from subtraction rather than addition. Where all fourteen had refused something, somebody who could do it well was given it — or nobody stands where everybody stands.
• We moved shape out of the shared brief into each character document and cut the shared brief by roughly a third, taking the ratio from 1.4 : 1 against character to roughly 1 : 1.
• We built the check that was missing — one that compares an article against every other article we have published, rather than examining it alone.

It worked, on its own terms. Where articles turn went from a single 47–72% band to a deliberate spread from 4% to 90%, plus several with no turn at all. How articles end went from one kind in thirteen of fifteen to twenty-seven distinct kinds across the roster. Fourteen of the fifteen articles were rewritten. And the editor read the next four and still heard one writer.

4. We tried a different model per persona

On 7 September 2026 the same assignment was given to five different commercial models, on the theory that if one model has one fingerprint, five models have five. The drafts came back 0.040 to 0.079 apart on the measure that actually carries authorship — no better than our own personas — and three of the five built their title on the same construction.

What the final measurement says

On 19 September 2026 the question was put properly, with three instruments that do not share a blind spot, against a control the earlier work never had: real human writers. That control is what changed the answer, because “are our voices distinct?” is unanswerable without knowing how distinct real colleagues writing one house style actually are.

The blind authorship sort

Fifteen sheets, each holding eight unlabelled ~400-word excerpts by four writers, two excerpts each, always from different documents. Three arms under identical framing so a reader cannot tell which arm a sheet belongs to: real practitioner bloggers writing under one strong institutional house style (the positive control), our pre-voice-layer originals, and our live re-formed articles. Thirty readings, blind, no key. Scored by Adjusted Rand Index, where 0 is chance and 1 is a perfect partition.

Arm	ARI, all 10 readings	Stronger reader	Weaker reader	Readings that said “one underlying writer”
Real human colleagues, one house style	0.58	0.73	0.43	0 of 10
Weaving Intelligence before the voice layer	0.12	0.21	0.04	10 of 10
Weaving Intelligence after the voice layer	0.46	0.71	0.20	8 of 10

Read the last two columns separately, because they moved differently. Sortability moved a long way — from chance to within reach of real colleagues; the stronger reader sorts our personas about as well as it sorts humans. The impression of a single hand barely moved at all: ten of ten before, eight of ten after, against zero of ten for the humans. Those are not the same question, and the whole project turned out to live in the gap between them.

One further result from the same sheets. Asked separately whether anything read as machine-written, the stronger reader flagged eight of eight excerpts on all ten of our sheets, and zero of eight on all five human sheets — perfect discrimination, no false positives. The weaker reader flagged nothing on our live sheets. A casual reader may not notice. A reader who uses these models daily, which describes our audience, will.

The stylometry, decomposed

Our own separation measure combines three things: the distribution of function words (the small connecting words nobody chooses consciously), sentence rhythm, and punctuation. On the headline number the voice layer worked — the live corpus clusters at 2.20×, about as well as real writers within one genre. The decomposition is where it comes apart.

Distance between two different writers (1,000-word windows, length-matched)	Real bloggers	Us, before	Us, live
Function-word shape — the layer writers do not control	0.1223	0.0577	0.0545 — 3rd percentile of the human distribution
Sentence rhythm — assigned by our voice profiles as a number	0.2332	0.2455	0.4189 — 1.8× human
Punctuation and contractions — assigned by our voice profiles as a number	0.8336	0.8135	1.1397 — 1.4× human

Function words are what a century of authorship attribution rests on, because nobody chooses their own rate of which and of. Measured on whole articles, the function-word distance between two of our personas was 0.0179 before the voice layer existed and 0.0179 after it. Not similar. The same number.

We pushed the two dials we could reach past human range, and the one we could not reach did not move at all. That is what “one writer in different dress” looks like in arithmetic, and it explains the gap between the two columns of the blind sort exactly: readers sorted our bylines on rhythm and punctuation, which we had exaggerated, and heard one person underneath, which we had not touched.

Why it failed — two mechanisms, and only one of them is ours

The blind readers were asked, in free text, what made the excerpts sound alike. Their answers name two distinguishable families of cue, and the distinction is the most useful thing this experiment produced — because we can act on one of them and not the other.

The two families are separable in the evidence, and the separation is checkable rather than asserted: Family A is visible in the pre-voice-layer sheets too. Readers described the same claim-concede-narrow-forbid skeleton on articles written months before any voice profile existed — because the standards predate the voice layer, and the standards are where it comes from. It is also visible in what the readers cited: on the human sheets the most-cited cue type was sentence and punctuation habits; on ours, it was rhetorical moves.

The conclusion

AIs are not up to that task currently if we want to maintain the quality standards we’ve established — because the standards enforce sufficient uniformity of style that it overrides any ‘personality’ we’re capable of injecting.

Jeff Shabel, 19 September 2026

The general form of that, for anyone running a similar process: a quality bar strict enough to be worth having is itself a style, and it competes with any persona layer you lay on top of it. The stricter the standard, the narrower the band of acceptable prose, and the less room remains for the thing that would make two writers sound like two people. We found the limit from the wrong direction — we were trying to add personality and discovered we had already spent the budget on correctness — but it is the same limit either way, and we would rather be correct.

Note carefully what this does not say. It does not say AI cannot write well; the articles pass every content-quality measure we have, and ten of ten approved packages hold a clear verdict on all five editorial axes. It does not say persona writing is impossible in general — a publication with a looser bar has more room to spend, and we would expect it to get further than we did. It says that for a publication holding this bar, with the methods available in September 2026, the two levers are one we decline and one that does not yet exist.

What we are doing about it

Decision	What it means in practice
One disclosed house voice	The six weekly personas and the eight on the bench are retired as writers, effective 19 September 2026. One establishing standard replaces fourteen character documents.
A new byline	Every article now reads Weaving Intelligence (House AI Style), with Jeff Shabel’s Experiential Commentary. It names the machine register and the human contribution separately, which is the accurate description of how these pieces are actually made.
Nothing is deleted	Every published article stays live and unedited at its original address. The fourteen biographies are preserved above; the long profiles are preserved in the exhibit set. The voice pages now redirect here rather than 404.
The published work is redrafted from the originals	Not from the persona rewrites — from the pre-voice-layer versions, in the house voice. Each redraft will carry a provenance note linking every state we hold of it, so a reader can follow one article through all three versions.
The structure repertoire survives, re-assigned	The seventy-odd article structures were the genuinely valuable part of the second attempt. They are now rotated per article rather than allocated per person, which keeps a five-day cadence from reading the same without claiming five writers.

What the revert does not touch

The threads stay, because a thread was always a beat and a subject boundary rather than a person. The prohibition lists stay — they are the control that caught a mis-commissioned topic before a word of it was drafted. Every correctness gate stays: counter-evidence, citation integrity, blind review, the figure checks. And the practitioner interview stays, which matters most, because it is now the only thing distinguishing an article here from competent generic copy. It always was. The persona layer obscured that rather than adding to it.

What we would tell someone doing the same thing

• Check between the units, not just inside them. Whatever your quality checks examine, defects larger than that are invisible to every one of them simultaneously. Ask what the largest thing any of your checks ever looks at is. Anything bigger is unobserved.
• Enforce the thing you actually care about. Anything merely encouraged loses to anything enforced, and you will not notice, because the enforced checks stay green.
• Do not ask for variety. Allocate it. And write down what each voice may not do — the forbidden list does more work than the description.
• Get a human control before you interpret any number. Our 2.20× separation looked like success for two weeks. It only became legible when we measured how far apart real colleagues writing one house style actually sit, and then decomposed our own figure to see which parts of it were real.
• Watch for a treatment and a test that share their axes. Our voice gate was independent of the profiles in the sense that it never read them — and 96% of the distance it reported was in exactly the two dials those profiles set. It certified that the generator had obeyed its settings. It could never have certified that a second writer existed, and no threshold on it would have helped.
• Read your output consecutively, in one sitting, on purpose. It takes an afternoon, and no tool we owned would have found this first.

What would change our mind

Stated plainly, because a conclusion with no stated falsifier is not a conclusion.

• The blind readers were models from the same family as the writer, which may make them unusually good at hearing their own habits. A human reader who sorts one sheet cold and hears several people would break the central finding. The sheets and the sealed keys exist; the owner reviewed the evidence and declined to run that check, on the grounds that he had already reached the same conclusion by reading. That is a decision, not an oversight, and it is recorded here as one.
• Excerpts cannot see article-scale form. Four hundred words shows none of where an article turns or how it ends, which is exactly what the structure repertoire acts on. Against this: the editor’s own readings were at full article scale, twice, and the second was on post-layer articles.
• The human blogger control is mixed-genre and we are single-genre, which inflates the human function-word distance. The single-genre calibrations shrink the gap — historians sit only 38% above us — but none of them reverses it.
• Who disagrees, and on what evidence. The verification literature reports 95–97% success at imitation in structured genres such as news and email, and separate work finds the homogenisation effect disappears when a model continues half of a real human text rather than being described to. Both point the same way: conditioning on large amounts of real human prose works far better than describing a person. Neither is available to an invented persona, which has no prose.
• Not tested: fine-tuned output of any kind; any reader who is not a model; and three of the fourteen voices, which had one article each and so had no same-voice pair to draw from.

Exhibits

Everything this page asserts rests on material we kept. Nothing below was reconstructed.

Exhibit	What it is	Where
A — the roster	All fourteen voices, beats and biographies as published	Above, on this page
B — the long profiles	The background essay and Q&A for each voice, roughly twenty-four times the length of the roster paragraph, live from 26 August to 19 September 2026	Preserved in the publication’s archive; the retired /voice/<name>/ addresses now redirect here
C — the originals	Fifteen articles exactly as first written, before any voice layer existed. These are the versions the redrafts will be built from.	Live and unedited, each at its own -original address — linked from The Voice Problem
D — the persona rewrites	Fourteen of the fifteen, re-formed against the assigned structures and coordinates. The second attempt, in full.	The currently published articles, paired with their originals
E — the blind sort	Fifteen sheets, thirty readings, all free-text answers verbatim, and the sealed keys. Seeded, so the sheets rebuild byte-identical.	Held with the verdict; available on request
F — the interim account	The Voice Problem, published during the pause. It documents the first finding and the second attempt, and it was written while we still believed the second attempt had worked. We have left it standing rather than revising it, because a corrected page hides the thing worth seeing.	Live

Every figure on this page was measured on 19 September 2026 against the articles as they stood that day, by scripts that are kept with the evidence and are re-runnable. Where a number came from an earlier round it is dated in the sentence that uses it. Nothing here is quoted from a summary.

This page is written in the house voice, by the process it describes — drafted by an AI, from a practitioner’s material and measurements, reviewed and ruled on by Jeff Shabel before publication. How Weaving Intelligence is written →