The prompt is not the work; describing AI contributions

We need a better vocabulary for describing contribution in the age of generative AI.

There’s a moment that keeps recurring in offices, group chats, and comment sections. Someone shares a piece of writing, a design, a chunk of code, and someone else asks, half-joking and half-hostile: “Did AI write this?” The question sounds like a question about authorship. But usually it’s a question about contribution: did you originate this, direct it, revise it, verify it, or merely release it? Those are five different answers, and the binary on offer — “yes” and “no” — lacks sorely needed expressive power.

A grand new theory of authorship and ownership produces mush, inscrutable academic papers, and navel gazing. A more useful project is smaller and more social: we need everyday language — call it a contribution grammar if you wanna sound fancy — for what kind of human involvement a piece of work represents. The failure of the language we have is one of expression before one of morality: long before anyone lies about their contribution, honest people mislead by accident, or go quiet, because they lack the words to describe their contribution. Law, publishing, and academia are already busy building their own vocabularies but the rest of us just need to be able to answer the question (“did AI write this?”) that one person is actually asking another when they squint at a document and wonder how it got made.

“AI-generated” is a bad label

AI-generated currently covers the person who typed “write me 2,000 words on supply chain resilience” and pasted the output into a LinkedIn post without reading it. It also covers the developer who described an architecture in detail, had a model draft the code, then reviewed every function, rejected two approaches, caught a race condition, and rewrote the error handling. It covers a writer who spent three hours in dialogue with a model, arguing, correcting, and discarding, whose final text doesn’t contain a single sentence from the first draft. It covers the artist who generated fifty images, curated one, and composited it with photographs. It’s a label with a severe lack of explanatory power and a shitty illuminator of what actually happened, as it covers both “help me think this through” and “do my thinking for me”.

Because these are wildly different processes: different amounts of effort expended, different levels of conscientiousness, different levels of ownership, and different levels of intellectual contribution. Fundamentally different levels of deployed agency. A label that treats them identically isn’t any kind of useful disclosure, and worse, it’s usually paired with a moral charge: “AI-generated” now functions as an accusation, so people either hide their process or over-confess to it, and neither response tells you anything useful about the work.

And this isn’t a hypothetical: researchers studying scientific authorship have started calling it the transparency paradox. An author who used a model to tighten a few sentences faces the same disclosure stigma as one who generated whole sections from a prompt, so authors avoid useful tools or stay quiet. Non-disclosure, the evidence suggests, mostly stems from uncertainty rather than outright deceit. In today’s vocabulary, even honest disclosure occludes as much as it reveals.

Typing was always a poor proxy

Part of why we’re stuck is that we’ve quietly treated word-by-word production as the gold standard of contribution: the real author is the one whose fingers made the sentences. But that standard was never coherent, even before AI.

An editor can transform a manuscript so thoroughly that the published book is arguably a collaboration, yet only one name goes on the cover. The Old Masters would have apprentices fill out the dull details. Nobody’s expecting politicians to have actually written their memoirs themselves.

So we already know that contribution comes in varying forms: originating an idea, directing its execution, selecting among options, refining, verifying, approving. But we’ve never had a pressing need for everyday language to distinguish them, because every form involved effort — and effort could stand in for all of it. Generation without effort breaks the proxy. What’s left to describe is agency: not whose fingers made the sentences, but who decided what the sentences would be.

The scarce contribution is judgment

A six-fingered hand in an AI image is embarrassing: not because a machine made it, we all know (now, anyway) that a machine made it. It’s embarrassing because nobody could be bothered to check it. The failure isn’t a failure of generation, it’s a failure of judgment and taste, and judgment and taste were the one contribution the human was supposed to supply.

If a machine can produce fluent text, plausible code, and polished images on demand, then fluency, plausibility, and polish stop being evidence of human effort. What the “did ChatGPT write this” question is really groping at is: how much of your judgment, expertise, and taste does this represent? Is this what you actually meant? The developer who approves AI-drafted code is (hopefully) making a claim: I have read this, I understand it, I will answer for it when it breaks at 3 a.m.

This is why “AI wrote it” and “AI wrote it and I verified it” should never have been the same sentence. The second contains the entire scarce resource.

The work may live in the iteration

There’s a common gotcha aimed at AI-assisted writers, given its current form by Clayton Ramsey’s “I’d rather read the prompt”. The argument runs that LLM output adds nothing to what the prompt already contained, so just send the prompt. Coming up with the prompt itself, then, was the work, and everything after it is fluff: there’s no contribution in hitting enter in the ChatGPT window. And for the lazy cases he’s describing — one prompt, one paste — sure. His objection lands nicely against “I made this” when it means “I requested this and here’s what I got back”. But it never touches the whole class of LLM-augmented generation where the contribution is “I sculpted this”.

A prompt can simply be a starting direction. When effort’s then expended, that effort lives in the sequence that follows: the objections, corrections, rejections, the refinement, the acceptance. Someone who went ten rounds with a model has left a trail of decisions, each of which said no to some versions of the text and yes to others. Sending you the first prompt would tell you almost nothing about the final artifact, in the same way that sending you a writer’s first outline tells you little about a published essay. Iteration isn’t overhead around the real work, it often is the real work, and we don’t have a useful way of describing that work yet.

Endorsement is not nothing

There’s one more contribution that’s easy to dismiss: mere signing off. If someone publishes a text under their name and says “this represents my view,” but a model produced most of the sentences, it’s tempting to call that rubber-stamping.

But it isn’t nothing. Not everyone has the words for what they think, and recognizing them when someone else finds them is its own act of expression. Beyond that, endorsement carries real consequences: the endorser stakes their credibility, accepts responsibility for errors (hopefully!), and asserts that the content matches their intentions. A world where people accurately endorse machine-drafted statements is meaningfully different from one where people share text that nobody stands behind.

Endorsement is obviously not authorship by itself: a signature doesn’t retroactively originate the work. And it’s the thinnest contributory claim — you can sign something whether or not you read it, and unlike “I reviewed every line”, it describes no process, only a stance. On its own, from a stranger, it tells us almost nothing.

But people do lie

The obvious objection to everything above: none of these finer-grained statements can be verified. “I reviewed every line” is exactly as cheap to type as it is to fake. Which is true, but also beside the point. A claim doesn’t need to be verifiable to be useful. “I skimmed it” is unverifiable and indispensable; so are “I read the contract” and “I’m nearly done”. We’ll surely continue to police lies the way we always have, with reputation and consequences, even now we’ve lost a powerful effort signal. What we can’t police, because it isn’t a lie, is the honest person with no words of the right size — the one who says “I wrote this” and feels like a fraud, or “AI wrote this” and undersells a week of work. The missing-words problem comes first, and it’s the only one a vocabulary can actually solve.

Toward a contribution grammar

The everyday problem feels tractable now, and in the places where fancier descriptions are needed, the grammar is already emerging. Academic publishing is sprouting four-level disclosure frameworks and taxonomies with names like GAIDeT and AID. Open-source projects have started requiring pull requests to declare whether code was fully AI-generated but human-reviewed, mostly AI-generated, or mostly human-written, and commit messages have grown an “Assisted-By:” trailer. All of it formal, institutional, a bit clunky — and all of it proof that wherever people are forced to describe contribution regularly, they build words for it. What’s missing is the conversational layer: the words for a Slack reply, not a compliance form.

Those words are cheap. We could simply start saying what we did: I directed this. I drafted it with a model and rewrote it. I reviewed every line. I curated this from many attempts. I endorse this; I didn’t write it.

That’s not a theory. Just distinctions — between authorship, effort, judgment, and endorsement — that we can use with each other in ordinary conversation. The binary of “human-made” versus “AI-generated” was a first draft of that grammar, produced quickly under pressure. And first drafts are supposed to be revised.

Prompt: hey claude, can you write me a think-piece for LinkedIn on authorship so I can capitalize on the release of fable? About 1500 words is probably enough.

…

Post-script: So how much of this is actually mine? It started as notes on my phone, pasted into ChatGPT and argued over on a 30-minute walk in voice-mode. Claude outlined it into an essay; I cut large swathes of wanky LLM-speak, did several rounds of review with both, then read every sentence and rewrote what I didn’t like. That produced more or less what I set out to write — minus my thoughts on cohesion, which we lost somewhere along the way. Then I went at it again: I found I could beat back the reviewer’s objections, but arguing with them showed me the flaw it had missed — the draft was torn between proposing a serious grammar and a grab-bag of complaints. So I chose a lane and reworked it paragraph by paragraph with Claude, tying it back to the everyday problem. It’s not quite the essay I planned. But I’ve pored over almost every word. The inline figures are entirely Claude’s work, with my only feedback being that I thought its first draft of the pruning image “is shit”, so it tried again. The hero image of the sausage was my idea, Claude helped me refine it, ChatGPT drew it, with several attempts needed to make it look less like a penis. In short, I’m happy to put my name on it.

Read Next

Forget the complexity: AI all boils down to drawing the right lines

August 2, 2025

Over and over again, despite the best efforts of humans, the most effective AI systems come down to one simple idea: finding the right shaped line that fits some data points.

Read Post

Get these articles sent to you

If you liked it, you might like other stuff I write