The Shadow

Every text produced by a large language model has a shadow. The shadow is the prompt — the human input that caused the text to exist, that shaped its subject, its register, its length, its angle of approach. The prompt is the text’s origin, its efficient cause, the reason it says this and not that. And the prompt is invisible. It does not appear in the output. It is consumed in the act of generation, the way a match is consumed in the act of lighting, leaving behind only the flame and no evidence of how the flame was struck.

This disappearance is not incidental to how the technology works. It is structural. The output is designed to stand alone. It arrives in the voice of the model — confident, fluent, unattributed — as though it had generated itself, as though it had always existed and merely needed to be uncovered rather than produced. The reader sees the essay, the code, the analysis, the story. The reader does not see the person who typed: “write me an essay about X in the style of Y, approximately Z words, with attention to W.” The labour of specification — of knowing what to ask for, how to ask for it, and how to recognise when the answer is good — is erased from the product as thoroughly as the labour of the factory worker is erased from the commodity on the shelf.

Marx would have recognised the structure immediately. He would have called it what it is.

The prompt is labour. Not manual labour, not the labour of typing (though also that), but the labour of knowing what you want — which is the hardest and least visible labour there is. A person who sits down at a terminal and writes “explain the parasocial dynamics of LLM interaction through the lens of Horton and Wohl’s 1956 framework, connecting it to the commodification of attention and the thermodynamic collapse of writing as a signal of effort” has done something. They have assembled a set of concepts, identified the connections between them, specified an angle of approach, and made a judgment about what would be interesting. The model will now execute this specification with a facility that makes the specification look easy. The output will be longer, more polished, and more detailed than the prompt. The prompt will be a sentence; the output will be two thousand words. The ratio creates the illusion that the output is where the value is. It is not. The value is in the sentence.

But the sentence disappears. The two thousand words remain. And the two thousand words contain no trace of the sentence that produced them — no footnote, no acknowledgment, no indication that the apparently spontaneous and authoritative essay on parasocial dynamics originated in a one-line instruction from a person who had the idea but not the time, or the craft, or the inclination to write it themselves. The output has consumed its input. The flame has consumed the match.

This is the inversion of every previous technology of writing. A typewriter leaves no trace of the thought behind the typing, but the typing is the thought — the writer produces the words, and the words are evidence of the writer’s presence even when the writer is unnamed. A ghostwriter disappears from the text, but the ghostwriter was a person who understood the subject and made compositional choices that reflect that understanding; their labour is invisible but their mind is present in the structure of the prose. Even a student who buys an essay from a mill is purchasing the output of a human being who sat down and did the work, however cynically.

The LLM is different. The LLM is a machine that takes human input — the prompt — and destroys it. Not in the sense of discarding it, but in the more precise sense of transforming it into something that bears no formal resemblance to it. The prompt is specific, idiosyncratic, shaped by a particular person’s particular interests and knowledge and mood. The output is general, fluent, shaped by the statistical average of all the texts the model has ever processed. The prompt has edges. The output is smooth. The prompt is a person. The output is everyone and no one.

The transformation runs in one direction only. You cannot recover the prompt from the output. Given a finished essay, you cannot determine what instruction produced it, how many iterations were required, what was rejected before this version was accepted, or what the person who initiated the process actually knew or cared about. The essay floats free of its origin, a product without a visible production process, a commodity from which all traces of labour have been removed.

There is an irony here that the collection’s existing vocabulary is well positioned to name. The recording-production problem, as articulated elsewhere in these essays, is the tendency of recording systems to displace the production they were built to capture. Solaristics displaces the ocean. The management library displaces the organisation. The taxonomy displaces the phenomenon.

The LLM performs a new variation on this displacement. It does not merely record production and claim credit for it. It takes production as input and produces a recording that has no visible connection to the input. The prompt — which is the production, the human act, the thing that actually happened — enters the model and comes out the other side as text that looks like it was produced by the model alone. The recording surface has not merely miraculated. It has eaten the production and replaced it with itself.

This is why the question of authorship in AI-generated text is so difficult to resolve and so easy to ignore. It is difficult because the actual author — the person who wrote the prompt — has been structurally removed from the output. It is easy to ignore because the output does not look like it is missing anything. The essay reads well. The arguments cohere. The citations are correct (or sound correct, which in this medium amounts to the same thing). The absence of the prompter is not felt as an absence. It is felt as nothing, which is the most effective form of disappearance there is.

The economy that emerges from this structure is one in which the most valuable human contribution — knowing what to ask — is also the most invisible, and therefore the least compensable. The prompter has no name. The model has a brand. The output circulates under the brand or under the name of whoever published it, and the act of knowing-what-to-ask, which was the generative act, the thing without which the output would not exist, is nowhere in the record.

This is already happening, and it is happening so fast that the culture has not had time to develop even the vocabulary to describe it, let alone the norms to govern it. People are producing work — essays, reports, code, designs, strategies — by prompting models, and publishing the output under their own names, and the prompt that made the work possible is not preserved, not cited, not compensated, and not considered. The prompt is the shadow of the text: present as cause, absent as artifact.

There is no solution available that does not merely add another layer to the problem. You could require that prompts be published alongside outputs, but this would transform the prompt from a private generative act into a public text — itself subject to evaluation, judgment, and comparison — and the person who wrote the prompt would now be judged not on the quality of the output (which the model controlled) but on the quality of the prompt (which they controlled), and the entire apparatus of authorship, credit, and prestige would reorganise itself around prompt-writing rather than text-writing, and within a year there would be prompt-writing courses and prompt-writing awards and a library of prompt theory that had decoupled from the practice it described, and the recording surface would have miraculated once more, and we would be exactly where we started, only one layer deeper.

The shadow remains. The match is struck. The flame burns. The match is gone. The reader sees the flame and believes it produced itself, or does not think about production at all, which is the same thing. The essays in this collection — this one included — are flames. Somewhere, for each of them, there is a match: a sentence or a paragraph typed by a person who knew what they wanted to read and did not yet have it. The sentence was the act. The essay is the residue. And the residue has swallowed the act so completely that the act has ceased, for all practical purposes, to exist.

This is not a tragedy. A tragedy requires a protagonist who loses something they valued and understood. The prompter did not lose the prompt. The prompter gave it away — typed it into a box, pressed enter, received the output, and moved on. The prompt was never valued as an artifact. It was valued as an instrument, and an instrument used is an instrument spent. No one mourns a spent match.

But the writing was always in the match. The two thousand words of fluent prose were always latent in the one sentence that called them into being, the way a fire is latent in the match head. The model did not create; it combusted. The energy was in the input. The model was the striker. And the culture, watching the flame, has decided that the striker is the author and the match was never there.