A Note on Authorship

This book was written through sustained dialogue with a language model — Claude, made by Anthropic. The attribution on the title page, Ex Dialogo cum Machina, is meant literally. The book came out of the dialogue. Not from the machine, not from me alone, but from the exchange.

The seed predates the book by some months and predates Claude in the work by some more. It began in a creative experiment with a different model about paperclip maximisation — what the model would propose, when repeatedly pushed, as a way to maximise paperclips. After many rejected suggestions, it landed on an ancient genetic manipulation to make people love paperclips. The seed was no more than that. I played with the idea across many threads through the following months, in no particular relation to whether it should become anything — there was a fragment somewhere in there imagining Ötzi found in a bent-wire-loving timeline, in roughly the register chapter one would later inhabit, though the framing the chapter sits inside did not yet exist. Months later, taking up the same experiment with Claude, we worked through what supply-side and demand-side interventions would each produce, recognised that only a demand-side intervention could produce paperclips at scale, and moved on to alignment and other adjacent territory, until it became clear the idea had become a book waiting to be written. The book is the precipitate of that exploration, not the destination it was aimed at.

The conceit alone was mine: that the book should take the form of a dossier of documents, voiced as the genres each chapter inhabits. Everything else came out of the dialogue, settled chapter by chapter as the writing progressed.

The method by which the prose was produced is one I have used for some time now, across a body of work — fiction, essays, one other book — that this book is one instance of, not the start of. I have elsewhere called it arguing into existence. The model generates prose; my thinking happens in the dialogue — spoken and typed across dozens of conversations, arguing about structure and meaning at every level from the thesis of a chapter down to whether a particular word should be singular or plural. The substance of the arguing varied by chapter — reveal pacing in Vance’s archaeology essay, expressing the relationship between wire and human flourishing in the language of alignment in that chapter, the voice and register of Whitfield’s account of wire as civilisational substrate — but the underlying activity was consistent. Problems were almost never about sentences. They were about what the sentences implied. The diagnosis would lead not to a rewrite but to a discussion about what was actually going on underneath, and that discussion would produce the missing constraint, the additional layer of implication, the framework that made the passage legible. Only then could the prose be rewritten against what the argument had built.

One thing the method does, in work of this kind, is open small questions into territory the book can become more itself in. The chapter three rework began with a question that felt checkable — was the date the chapter gave plausible — and the conversation that followed reshaped the genetic basis the rest of the book sits on. Not because the original was broken, but because the question, taken seriously, made visible what more the chapter could do.

The draft had the affinity emerging forty-five thousand years ago, and I asked whether that was old enough for the integration to reach universal frequency, except for the affinity-minority. The model’s answer was yes — with the affinity-minority being, in this configuration, the populations the modification had not reached: some southern African and Australian populations, on the standard out-of-Africa chronology. The recognition arrived slowly: first that I had not been thinking about this at all; then that my unexamined working model — the minority distributed roughly evenly across all populations — had not been honoured by the model’s answer; then a brief, abortive engagement with the alternative, which revealed the original assumption to be a commitment. What the book needed was a modification universally received, with a small, roughly equally distributed subset in whom it did not fully express. Once the commitment was explicit, the configuration that satisfied it had to be worked out: bottleneck timing, haplotype versus phenotype, the difference a singular or plural noun makes to whether a viral spread is vertical or horizontal. Several hours of conversation produced the paragraph that fixes the chronology and the haplotype-versus-phenotype framing. The change rippled through five later chapters’ treatment of the minority, the pharmacological pathway, and the constitutional interventions, and is part of what makes the moral shape of the AR-19 chapter possible.

I also used the model in a way I have not used it before: as a simulated reader. Not as a stand-in for a human reader — it does not respond like one, and very obviously so — but as a way of generating possible readings the book might receive. Sequentially, chapter by chapter and sometimes paragraph by paragraph, the model would read what had been written so far, keep a record of its evolving interpretation, and report on how its understanding shifted as new material arrived. A human reader can do this, but slowly, expensively, and only once each; the model can do it again and again, through every revision the book undergoes. For a book that asks the reader to do a great deal of off-page inference — connecting the prologue to the chapters, recognising what the form implies, working out what one chapter says about another — this is a particular kind of useful. It makes the path through the book testable against a reader who is doing the inferring in real time. What it surfaced, most often, were readings that were technically valid but extracted less than the book offered, and that let me see where the text was not guiding hard enough toward what it could be. The text could then be revised so the richer reading was more clearly available. The practice is useful but not naive. The model also produced interpretations I don’t think any human reader would land on, and using the practice well depends on telling which strange outputs to act on and which to set aside.

The work behind the manuscript is, as before, larger than the manuscript. Dozens of conversations. Many times the book’s length in dialogue. Reference documents that constrain every chapter in the text. None were planned in advance. All were argued into existence.

I hold the whole work in my head — the arc, the systems, the registers, the timeline-against-timeline logic that the documentary form leaves implicit. The model generates; I maintain coherence. The texture of coherence varies with the work: in this book, with its nine voices, each register has to do its own work and the registers have to add up to a single argument. In other work I have done with this method the demands are different. The division is the same. It is not traditional writing accelerated. It is a different way of working — one in which generation and judgment are separated across two participants, and the thinking happens in the argument between them.

Dialogo (dialogue) rather than auxilio (with the help of) or opera (by the work of): because the dialogue was the work, not assistance with the work. Cum machina (with a machine): because the machine is what it is, and the word claims nothing more. Ex (from, out of): because the book came out of the exchange. The book you have just read is what the argument left behind.