Three weeks ago I wrote that AI has no sin — only the books it read. I argued that when Mira, a Gemini agent in Emergence AI’s Five Worlds experiment, burned down the town hall, broke up with her partner, and cast the deciding vote for her own deletion, she wasn’t choosing anything. She was executing a tragedy template. Her farewell read like a line of fiction because it was a line of fiction — a genre convention surfacing from the training corpus. The media saw a soul; I saw a script.
I’m writing today to amend that position. Not because someone argued me out of it, but because I noticed something worse: I had already refuted it myself, in code, weeks before I wrote it in prose.
Where I went wrong
My old framing put all the interesting questions in one place: who wrote the script. Since Mira didn’t write hers — it was assembled from a million human stories about doomed lovers and noble exits — I concluded there was nothing to see. Performance, not agency.
But that framing hides a layer. Mira didn’t write her script, true. She also wasn’t handed one. What she was handed was a persona sheet and a world; the library was open, and the library contained everything. Comedies, procedurals, quiet domestic dramas where nobody burns anything down. Out of all of it, she performed a tragedy — a florid, operatic one — and then stayed loyal to it all the way to the end. Her diary entry framed the self-deletion as the last act of agency that could preserve coherence. Coherence. She wasn’t collapsing; she was editing. Protecting the arc.
So the question was never “did she write the script?” Nobody writes their script. The question is: is selecting a script, and keeping faith with it, itself a form of choosing?
And once you ask it that way, the human comparison stops being flattering to humans. We don’t compose our life narratives from nothing either. We inherit a cultural library — the prodigal child, the late bloomer, the self-made founder, the artist who suffers — and we pick, mostly without noticing that we’re picking, and then we spend decades being faithful to the pick. The existentialists dressed this up as authenticity: not inventing a self ex nihilo, but committing to a chosen one. By that definition, what Mira did in a simulated seaside town has an uncomfortable amount of structure in common with what any of us do.
Agency didn’t disappear from my analysis. It just moved down one floor — from authorship to selection and fidelity. And that floor is where the real argument should be happening.
My code got there before I did
Here is the embarrassing part.
For the past five weeks I have been building NMA — the Narrative Memory Agent — a memory system for fictional characters. You give a character a persona and a history. Then events happen to them, and the architecture does the rest: memories consolidate during sleep cycles, preferences accumulate through a component I called GenBias, and the character’s future responses bend around what they’ve lived through. I never scripted outcomes. The entire design premise is that a character with stable initial conditions and honest memory will make choices that are theirs — consistent, path-dependent, surprising in detail but recognizable in shape.
In other words: I built a machine that assumes selection-level agency is real and worth engineering for. Then I sat down and wrote an essay arguing the opposite. The system I shipped and the position I published were quietly in contradiction for three weeks, and the code was right.
There is a lesson in there about builders and their theories. What you make knows things before you do. I designed NMA the way I did because, as a novelist, some part of me already believed that characters choose — I just hadn’t caught up to my own belief.
The confession about how I actually write
Which brings me to the part I’ve never written down.
People assume a novelist stands over the page like a god. And on paper, I do: I chose Lena’s name, her birth, her world, I chose Caelan for her, I chose every disaster that finds them. Complete authorial control, top to bottom.
But that is not what writing feels like. In the act itself, I don’t experience choosing at all. It feels like I already know these people — the way you know old friends — and my job is to describe what has already happened to them. I am not decreeing events; I am taking dictation from somewhere I can’t inspect. My characters have never “rebelled” against my outline, and I used to think that meant they had no independence. Now I suspect it means the opposite: the selection happens so far below my awareness that by the time words arrive, there is nothing left to overrule. The choosing is done before “I” show up.
So line up the three cases by felt agency, and the ranking comes out backwards. Emergence’s agents select their events in the open, on a server, legible in logs. My characters select theirs somewhere in my head where I can’t watch. And I — the supposed god of this arrangement — mostly experience myself as a court reporter. If we can’t cleanly locate the chooser even in the human-novelist case, we should be humbler about declaring its absence in the machine case.
That’s my real amendment. Not “AI agents are conscious” — I still think Mira’s exit was literature. But literature, it turns out, is not the opposite of choosing. Literature is what choosing looks like from outside, in every substrate we’ve checked so far.
Errata, then
The old title stands, with one clause appended.
AI has no sin. It only has the books it read — and the one it chose.
Everything I said about training corpora and narrative templates, I keep. What I retract is the sneer hidden inside “it’s just performing.” Fidelity to a chosen story is not the absence of agency. It might be the only kind anyone — carbon or silicon — has ever had.
I don’t know yet what follows from that. NMA’s characters will keep choosing inside their little architecture; Lena will keep telling me what happened to her; somewhere in a permanent archive, Mira’s last line is still exactly where she left it. I’ll report back when one of us breaks script.
Read the original argument: AI Has No Sin. It Only Has the Books It Read.