Selection at Scale: On Writing With AI

Writing is not a single activity. People write as an art form, to think, to communicate, to document, to argue. A poet laboring over a single word and an engineer documenting an API are both writing, but they are doing fundamentally different things with fundamentally different stakes. What counts as authorship, and what any given tool costs or contributes, is not the same across all of them.

A Form of Authorship

For someone writing to think, to communicate, to document, to argue, the question of whether LLM assistance constitutes authorship is worth taking seriously. It looks, on the surface, like it might not. Prompting a machine and selecting from what it returns looks less like writing and more like editing. And we already have a word for someone who improves written words they did not originate. The distinction feels obvious. An editor is not an author.

A person using an LLM is initiating their own idea. They have an inner vision of something they want to express, and they are using the LLM to move past the low-level effort of binding individual words into coherent phrases, selecting instead from larger units, sentences and passages that already hold together. As long as they are iteratively engaging with what comes back, steering and adjusting toward what they mean, they are using the same basic mechanics as writing manually. What they are doing is a form of authorship, distinct from purely manual writing, but authorship nonetheless. The assistance changes the level at which selection happens. It does not change who is doing the selecting, or why.

The Pre-Existing Field

No single person invents the words they use. Meaning is already encoded, built into the language by the culture before any individual writer arrives at it. The grammatical structures, the conventions, the emotional weight of particular phrases: none of that originates with the person writing. It accumulates across time and culture, shaped collectively across millions of prior speakers. When a person sits down to write, that work is already done. What they do is select and arrange from a semantic field they did not create, drawing on meanings already ascribed to the words before they touched them. Individuals occasionally contribute to the lexicon, but even then a new word only becomes language when the culture absorbs and ratifies it. Writing is selection from a pre-existing field of meaning.

Authorship has always been partially distributed, between the individual and the language they work inside, between the writer and the culture that built the instrument they are writing through.

Selection at Scale

Writing manually, the writer is selecting at the finest level, word by word, sometimes phrase by phrase. The unusual word choice, the unexpected grammatical construction, the sentence that does something you have not quite seen a sentence do before: these emerge from that fine-grained process. Working with an LLM, the writer pulls back. Selection happens at the level of larger strings, phrases, sentences, sometimes whole passages that arrive already coherent. That coherence is real and useful. A draft advances more quickly. The initial output is readable in a way that early manual drafts often are not. But selecting from larger pre-formed units, you lose access to the finest grain. The writing will often be serviceable. It will rarely be remarkable. For the writer whose purpose is the words themselves, that gap is the whole point. For the writer whose purpose is the idea the words carry, it may not matter much.

The Initiating Vision

Normally, when you write, your brain generates candidate phrases and you evaluate them against what you mean. Both steps happen inside your head. With an LLM, the generation moves outside you. Candidate expressions appear on the screen, and you react to them, accepting, adjusting, discarding. The matching of word to meaning still happens in you. The source of the word changes. Where recognition happens does not.

Recognition is the authorship act because the author is the only person with access to the initiating vision: the vague, unstable felt sense of what they were reaching for before any words existed. The editor does not have access to that. The LLM does not have access to it. When an author accepts a suggested revision, from an editor or from a machine, they are checking it against something no one else can see, and confirming that it lands correctly. That is why the editor's contribution, however significant, does not transfer authorship. The editor can improve the expression. They cannot verify whether it matches the intent, because the intent was never theirs. Individual authorship was always about selection and recognition within a pre-existing system. The author's role was always to hold the initiating vision and confirm when the words finally match it. Externalizing generation to an LLM is a difference of degree, not kind. Generation is upstream of meaning. Recognition is where meaning is confirmed.

Writing as Thinking

This is also why the iterative process with an LLM is more than mechanical selection. Language is not only how we express what we already know. It is part of how we figure out what we mean. Before you find words for an idea, it sits vague and unstable, a sense that something is true, or wrong, or important, but not yet anything you could hand to someone else. The act of searching for words clarifies and sometimes creates the idea, which is why most writers discover what they actually think by writing, not before. Each response from an LLM is something to react to, refine, redirect. The initiating vision sharpens across that back-and-forth in the same way thinking sharpens in dialogue. The tool changes the texture of the process. The cognitive work remains.

The Failure Mode

The failure mode worth watching is accepting fluent output without checking whether it matches what you mean. LLM output is often grammatically clean, structurally sound, and close enough to what you were reaching for that comparison to intent feels unnecessary. The serviceable quality of the output is precisely what makes this easy to miss. When that comparison stops happening, the writing becomes hollow. The initiating vision has no mechanism left to verify itself against the result.

A Tool Like Any Other

LLMs are tools, and tools misused produce undesirable outcomes. They also have a well-documented tendency to deskill craftspeople while enabling laypeople to bypass their own deficiencies and produce work they could not reproduce on their own. A keyboard with synthesized notes removes much of what a trained pianist brings to a grand piano. A miter saw allows a weekend carpenter to make cuts that would take a professional years to learn by hand. The tool lowers the floor without raising the ceiling.

There is no serious argument that an LLM writes as well as a skilled human author. It has no interior experience, no singular perspective accumulated from a particular life. It does not know what it means to lose something, to want something, to be surprised by its own thinking. Those are not incidental deficits. They are the source of what makes writing worth reading at its best.

And yet the tool widens something. It extends the reach of communication and the expression of ideas. It gives people a way to externalize vague thinking and refine it through engagement, producing clearer understanding even when the prose itself is only serviceable. Many editors cannot write as well as the authors they edit, but they still improve the writing. The fused authorship and editing relationship that LLM writing produces is genuinely complicated. It does not replace the craft of writing by hand. It extends and improves many other kinds of writing, for people who were never going to write at the level of craft anyway.

What the Work Is For

The stakes of that failure vary by writer. For someone documenting a process or working through an argument, hollow output is a practical problem, correctable through closer attention. For a writer whose purpose is the words themselves, it is a different kind of loss. The finest grain is where their meaning lives, and no amount of iterative prompting recovers what only emerges from the slow, solitary work of choosing word by word. An LLM will never love the language it generates. It has no stake in the particular word, no felt sense of why this phrase and not another. For the writer who does have that stake, handing the process to a machine hollows out the thing they were actually doing.

The tool is not neutral across all uses, and the question of whether to use it is not really a question about technology. It is a question about what you are trying to do when you write, and whether the answer to that requires your hands on every word.