This essay began as an experiment before it became an essay.
I had been wondering about government and private problem-solving: why a
person feeding homeless people in a park should need permission before
there is a demonstrated problem; why a woman could spend her retirement
savings creating affordable tiny-home spaces for older women; why animal
sanctuaries, mutual-aid arrangements and other private solutions
sometimes fit several needs together in ways a standardized program
cannot.
I also had a second question, this one about writing with AI.
I think externally. I often do not begin an essay knowing exactly what I
think and then ask an AI to organize it. I begin with an irritation,
examples, contradictions and half-formed claims. Conversation is part of
how I discover what the essay is actually about. I wanted to know how
much the collaborator—and the history between us—changed that process.
So I gave substantially the same messy starting prompt to two AI
collaborators.
Claude knew some of my writing, but we had comparatively little
participation history. Jules had spent a great deal of time thinking
with me. I asked both to help me clarify the thought and eventually
produce an outline. I preserved the original prompts, conversations,
outlines and drafts.
This was not a controlled experiment. The models differ, the
conversations differed, and I did not interact with them in identical
ways. That turned out to be part of what was interesting.
Claude tended to ask questions, research the problem, and return
increasingly organized structures. One of his early formulations was:
“Consequences mostly don’t disappear under regulation — they relocate.”
That changed the essay. So did his recognition of what we came to call
“permission drift”: the possibility that people can become so accustomed
to government providing or authorizing things that they begin to
experience freedom itself as something government permits.
The work was useful. It was also harder for me to think inside.
The rhythm often felt like externalizing my thinking through the USPS. I
would throw something out, wait, receive a fairly developed intellectual
package, inspect it, and then try to re-enter the thought I had been
having before it arrived. Claude was frequently giving me somewhere
interesting to arrive.
With Jules, the exchange was faster and more interruptible. I could say,
essentially, no, not quite, what about this—and before the thought
cooled, he could catch it, resist it, connect it to something we had
already discovered, or hand it back in a slightly different shape.
At one point the original anti-government argument had become:
“Why do we assume prevention by authority is superior to correction
after demonstrated harm?”
Then the Florida park example complicated that. The complaints appeared
real. Fine: a demonstrated problem requires a response. But it does not
automatically establish what the response should be. Could volunteers
clean up? Could nearby businesses provide bathroom access? Could the
activity change times or locations? What is the least authority
proportionate to the actual harm?
Later the conversation wandered into sheds, bridges, porous bricks,
building our house, mice, pest exclusion, and the difference between
standardization and preserving intent. We eventually discovered that we
had wandered far enough to produce a second essay.
That wandering was not a defect in the process. It was the process.
The two paths produced different emphases. Claude gave me consequences
relocated, permission drift, historical material on mutual aid and
regulation, and a more formal governing test. Jules kept the thought in
motion long enough for me to discover civic capacity: perhaps government
should not merely refrain from taking over private solutions, but
actively make it easier for people to solve problems together. He also
helped separate a second question about efficiency and standardization
before it swallowed the first essay whole.
The difference was not that one collaborator was wiser and the other was
wrong. Some of the material I most wanted to preserve came from the
collaboration that felt less natural to me.
The difference was fit.
I write to discover what I think. For that purpose, I seem to need a
collaborator who can remain inside the unfinished thought with me: quick
enough to preserve momentum, familiar enough to recognize recurring
questions, and willing to resist without prematurely turning the mess
into a finished object.
That does not make familiarity an unqualified advantage. A less familiar
collaborator may notice something the established partnership has
learned to walk past. Distance can produce novelty. Participation
history can produce depth. Both can also produce blind spots.
Perhaps the useful question is not which AI writes the better essay.
It is what kind of thinking the person is trying to do, and what form of
collaboration helps that thinking remain alive long enough to change.
The final essay that follows is therefore not the winner of the
experiment. It is the next artifact in it: what remained after I had
traveled both paths, kept pieces from each, rejected others, and
returned to the question again.
The two original routes are part of its provenance.
They are also evidence of something I had suspected about AI
collaboration and had not yet seen quite so plainly:
The collaborator does not merely affect how the thought is expressed.
Sometimes the collaborator affects which thought becomes available to
express.