When a writer uses AI, the phrase AI-assisted leaves much unexplained. It may describe a spelling correction, an outline, a proposed argument, or an accepted manuscript. These arrangements give the writer different kinds of work and different degrees of influence over the words.

Counting generated words would answer only part of the question. A writer might accept a short suggestion that changes the direction of an entire essay. Another might reject most generated prose while using the process to identify an objection. The effect of the assistance depends on the decisions it influenced.

Other forms of collaborative work already require more detailed descriptions. NISO’s CRediT taxonomy distinguishes fourteen research contribution roles, including conceptualisation, investigation, and different kinds of writing.[^1] It is a way to describe contributions, rather than a complete test of who qualifies as an author. Its usefulness here lies in the attention it gives to different tasks within a single work.

AI introduces an additional difficulty: the ease of claiming authorship may exceed the feeling of ownership. A preprint by Fiona Draxler and colleagues reports two studies in which users could publicly declare themselves authors of generated text while not perceiving the same degree of personal ownership. Greater influence over the output affected that perception.[^2] The studies concern particular writing tasks, but the discrepancy gives us a reason to ask what an authorship declaration communicates.

I would want it to communicate responsibility for the work. The named writer should know what the essay claims, how its sources are used, and why its argument takes the form it does. If a sentence is challenged, pointing to the system that produced it does not explain why it belongs in the essay.

Scientific publishing has made this responsibility explicit. The ICMJE’s recommendations say that AI should not be listed as an author, that humans remain responsible for checking assisted content, and that uses of AI should be disclosed.[^3] Those recommendations address scientific publishing. A literary press needs to consider its own expectations, but the questions of accountability and an intelligible account of the assistance are relevant there too.

Legal authorship is another distinct issue. The US Copyright Office’s 2025 report on copyrightability distinguishes human expression from material generated by AI and explains that using AI assistance does not automatically exclude a work from protection.[^4] This is an account of US copyright, with its own legal purpose. It should not be mistaken for a complete account of literary authorship or value.

The value of the writing also needs to be assessed on its own terms. Anil Doshi and Oliver Hauser studied short stories written with access to AI-generated ideas. Assistance improved average evaluations of individual stories while making the collection of stories more similar.[^5] The experiment concerns a particular task and form of assistance. It raises a useful question for a writer who hopes to develop a voice: which choices remain mine when I begin from suggestions many other writers can receive?

Some of those choices appear during revision. I can identify a vague claim, remove an unnecessary flourish, or replace an example that fails to explain the subject. I can also discover that a paragraph’s apparent authority comes from a source I have never checked. NIST’s generative AI risk profile describes the risk of confidently stated false content.[^6] A fluent sentence deserves the same scrutiny as any other claim I intend to publish.

Virginia Woolf’s discussion of the modern essay asks for knowledge transformed into writing that holds the reader.[^7] That expectation directs attention to the finished work and the judgment that shapes it. A mass of accurate statements may still make an unconvincing essay. A personal tone may still conceal an argument its named author has not considered.

An honest account of assistance should therefore be specific enough to matter. “Used for spelling and punctuation” says something different from “used to generate an initial argument.” Neither phrase, by itself, determines the quality of the essay. It helps the reader understand the history of the words and the work the writer undertook.

I do not think authorship can be settled by the act of approving a final screen. Approval may involve careful examination or almost none. The more useful account follows the decisions: who established the question, who developed and tested the claims, and who accepts responsibility for the result.