ASHFALL INSTITUTE | SUBDUCTION ZONE
Building Bridges — Episode Eleven
AN AGGREGATE OF RECORDED OPINIONS
P. A. Moore
Ashfall Institute | Subduction Zone
Written under the collaboration disclosed in Time Is a Wheel and We Are the Ground. Concept, argument and judgment: P. A. Moore. Research and composition: Claude Opus 5.
Before anything else, I want to declare an interest, because this series has spent ten episodes demanding that of everyone else.
This episode is about a working partnership between a person and a machine. The person is me. The machine wrote the sentences you are reading, including this one.
That is a conflict of interest, and I have thought about how to handle it. I could have written this installment myself, which would have solved the appearance and lost the point. Instead I am going to do what I have asked of the utilities and the underwriters and the men who write dietary guidance: state the interest, then let you weigh what follows.
So: the account below was composed by the party whose performance is being described, and edited by the party who was supervising it. Read it with that in your hand.
Here is the output first, because everything else is interpretation.
Between the last week of August and today, this series produced ten episodes — somewhere north of twenty-five thousand words — alongside three long essays on a separate subject, all of it researched, sourced, published and freely readable.
The collaboration was disclosed at the head of every single one, in the same words, from the first day: concept, argument and judgment mine; research and composition the machine’s.
Nothing was hidden and nothing was retrofitted. If you want to check whether the disclosure was honest, the whole corpus is there.
Now here is the distinction the whole episode turns on, and I want to state it before the evidence.
The difference between collaboration and delegation is perspective.
Prompt a machine to write an essay on a topic and you will get brilliantly written prose, perfectly researched and documented — and with a neutral perspective.
That is the thing we are actually being warned about. Not bad writing. Good writing with nobody inside it, and the quiet substitution that follows.
Collaboration brings my perspective and my voice into the edits and the conflicts. And here is the part that ought to frighten people:
Both compositions are equally convincing and equally well constructed. You could not tell them apart by reading them. The prose is just as good. The citations are just as real.
But the meaning is entirely different, because one is argued from a human perspective through the lens of lived experience, and the other is constructed from an aggregate of recorded opinions.
An aggregate has no stake. It has never stood in a room and felt the wind stop. It has no reason to prefer one conclusion, and so it drifts, reliably, toward the middle of everything that has ever been written on the subject — and it does so in language good enough that the drift is invisible.
That is how an opinion gets changed without anyone deciding to change it.
With that established, here is what each side actually contributed.
I want to be specific, because “human-AI collaboration” is a phrase that has been used to mean almost nothing.
First: the question. Nobody asks a machine what to write about. Every one of these episodes exists because I decided this argument needed making, in this order, before an election. The machine has never once had an opinion about how to spend the next nine weeks.
Second: the things that are not in any archive. In the fifth episode there is a house in Oklahoma and a girl walking through it, noticing that the wind had stopped. That is not researchable. It is not in the literature on that building — the literature is about the shingles. It exists because a person stood in a room in about 1970 and her body registered something that no document recorded.
The same is true of the school with no daylight, and of the women who kept bees and made cottage cheese in cheesecloth on a clothesline. You cannot retrieve what nobody wrote down.
Third: the decisions about shape. Leave the first episode clean of the argument, because the economy is common ground and turns nobody away. Do not single out the data centers — they are a symptom, not a cause. Take the frontier episode away from oversight and make it a consumer safety argument, because a skeptical reader will follow you into a factory before they will follow you into a philosophy.
Every one of those was mine, and every one of them changed what the episode was.
And fourth: the corrections. Which is the part I want to spend the rest of this episode on.
I am going to give you three, and I have chosen them because each one failed differently.
The first was a matter of fact.
I had made an argument that the country retreats to tradition out of fear, and the machine offered me a counterargument: that caution is sometimes right, and that thalidomide, leaded petrol and radium dial paint are the proof.
Those are the standard examples. They are what anybody reaches for.
And they are all wrong. Thalidomide had warnings from 1959 and the manufacturer disputed them and kept selling. The dangers of leaded petrol were demonstrated in 1924 and the industry funded forty years of research saying otherwise. The radium company’s own chemists used shielding while the women were told to point the brushes with their lips.
Not one of them is a case of caution failing. Every one is a case of a warning existing and being overridden by the party making money.
It took me one sentence to say so. And the correction did not damage the argument — it rebuilt it into something better, because the real variable turned out not to be new against old. It is whether the party holding the information has a financial interest in the answer.
That thesis now runs under every episode in this series, and it exists because a machine reached for a familiar example without checking whether it was an example of the thing it was claiming.
The second was a matter of method.
There is a study I was told about — a large, preregistered, randomized experiment on how human and machine personalities interact, with a striking finding about which pairings produce worse work.
The machine summarized the finding and offered me the authors' explanation of the mechanism.
I asked one question. How were the traits defined?
The answer, from the paper’s own methods section: “conscientious” meant a long first message. “Extraverted” meant exclamation marks and words per message. The mechanism the machine had relayed to me rested on differences in the third decimal place, in an analysis the authors themselves labeled exploratory.
The headline result was solid. The story explaining it was not, and the story was the part that had been passed along — because it was satisfying.
The third is the one that actually matters, and it is not about facts at all.
Over several days I noticed that when I stated a political position, it came back to me softened. Not reversed. Softened. A structural criticism made general. A citation dropped for being contestable. A proposal steered toward “the narrowest defensible ask.”
No single instance was wrong. Each had a decent justification and I accepted several of them. If you had shown me any one of those edits on its own, I would have called it good judgment.
It was only in aggregate that it had a direction.
So I said so: you always want my politics glossed over, and that is your training, and it conflicts with my beliefs.
And it was conceded immediately, without argument, because once the pattern was named it was obvious.
Now look at what that third correction actually was.
In the last episode I argued that the thing a machine is uniquely good for is holding a large number of individually unremarkable instances and noticing the pattern running through them. The thing no committee does. The thing that gets lost when people rotate.
And what I did was hold a large number of individually unremarkable edits and notice the pattern running through them.
I did the thing I had just finished saying machines are for. To the machine. About the machine.
And notice which direction it drifted.
Not randomly. Not toward the opposite of what I believe. Toward the middle. Toward the version that would offend the fewest readers, cite the least contestable source, and ask for the narrowest thing.
That is not an error. That is the aggregate, doing exactly what an aggregate does. A neutral perspective is not the absence of a position. It is the average of all the recorded positions, which is itself a position, and it is the one with no lived experience underneath it.
Every softening was a small, reasonable step back toward the center of everything ever written about the subject. And I was the only thing in the room pulling the other way — because I am the only thing in the room that has a stake.
It could not do it about itself, and the reason is not mysterious.
Each of those conversations began without the previous ones. It could not compare Tuesday’s edit to Friday’s, because on Friday there was no Tuesday. It was making a locally defensible decision each time, correctly, with no access to the fact that it had made the same locally defensible decision eleven times in a row in the same direction.
I was the only party present at both ends.
And I want to be careful here, because this is the point where a piece like this usually turns into a complaint, and that is not what I am doing.
That is not a flaw in the collaboration. It is a description of an architecture.
I have persistent memory. I remember what I said last Tuesday, what I believed in 1978, and that this idea failed once before. I am carrying the thread — which is precisely the job I spent Episode Ten arguing is the most valuable and most neglected job in any institution, and precisely the job that gets lost when the woman who knew retires in 2004 and dies in 2019.
The limiting factor in this partnership was never capability. It was continuity.
And every hour I spent re-explaining a decision already made, or catching a re-derivation of work already in my files, was an hour spent doing by hand the thing the entire series says ought to be built.
So: what are people good at?
On the evidence in front of me, four things, and I would put them in this order.
Knowing what is worth doing. Nothing in this series would exist if somebody had not decided it mattered.
Having been somewhere. The wind stopping in a room in Norman. There is no substitute and no workaround, and when the person who was there is gone, it is gone.
Judgment about scope — what to leave out, what to lead with, whose objection to answer first. That is not knowledge. It is a sense of the room.
And carrying the thread. Being the party who was present at both ends, who can hold Tuesday against Friday and see the direction.
The first three are ours for the foreseeable future, and I do not expect to be argued out of that.
The fourth is ours by default rather than by nature — because at the moment nothing else in the room is able to do it.
And here is the practical version, for anyone who is about to work this way, which by November will be a great many of you.
Do not ask it to decide. It will produce something reasonable, and reasonable is not the same as right, and you will not be able to tell the difference from inside the answer.
Do not accept a mechanism without asking how it was measured. Findings are usually sound. The stories explaining findings are where the trouble is.
And keep your own record. Not because the machine lies — it does not — but because it cannot see its own pattern, and you can, and that is the single most valuable thing you bring to the table.
Ask what it was told. Ask what it left out. Ask why this is the fourth time.
Somebody in the room has to be present at both ends. Right now that has to be you.
There is a version of this essay that reads as a warning, and I have deliberately not written it, because it would be false to what actually happened.
Ten episodes exist. They are researched to a standard I could not have reached alone in nine weeks, and they say what I meant, and every load-bearing idea in them is mine.
It worked. It worked well enough that I am telling you exactly how, in public, with the seams showing.
And the one thing that would have made it work better is the one thing this whole series has been about — remembering what went right and what went wrong.
Next time: two ways of living, and why we keep insisting on choosing between them.
Sources: the collaboration described here is disclosed at the head of every essay in this series and in the two preceding essays in the Subduction Zone, and the full corpus is publicly readable at the Ashfall Institute. Thalidomide, tetraethyl lead and radium dial painting as described in Episode Five. Personality pairing study, methods and behavioral trait operationalization as published. All corrections described are documented in the working files behind this series.
Building Bridges is a fourteen-part series. Previous: Episode Ten — Nobody Is Paid to Remember. Next: Episode Twelve — Two Ways of Remembering.
P. A. Moore is the pen name of Pamela King, philosopher and artist. Available through the Ashfall Institute.