Notes on Writing and AI

version: 2026.9.17

This is a statement of principles around LLM use in writing. It’s a sketch of my current views and practice, and is likely to evolve.

I am not dogmatically opposed to using LLMs in my writing, but neither do I wish to completely hand over all my writing tasks to a machine. Different kinds of writing call for different principles of LLM use. To figure out these principles, start by interrogating the purpose and audience of the writing project.

For essays, and in particular those posted here, the purpose to see what another person thinks, and also how they arrived at their thoughts. Here it’s important that 100% of words are that person’s.

In this setting, I’ll use Claude as a writing coach for feedback on structure, and research assistant to help track down facts, but not to contribute or even suggest a single word. Including this in project notes has worked well for me so far:

For these writing projects, it is important that the final product be 100% generated by me. Your role is to advise, act as a sounding board, read with a critical eye, offer constructive feedback on any writing I provide. If necessary, suggest a “back to the drawing board” complete rewrite. Proofreading for typos and grammar is very welcome, but please point to the spots (by context is usually most helpful to me, or page and line if that’s appropriate) so that I can fix things, rather than offering to fix them yourself.

I’ll also ask for “research assistant” roles: locating and checking sources, gathering and presenting data, etc. Please leave interpretation to me in those cases.

For a math/science paper, let’s break things down into what purpose it serves the reader versus the author. Among other things, a reader might want:

to see what’s true;
to understand something;
to see that another human has understood something and/or solved a problem.

Likewise, an author might want:

to communicate their work on a problem to other experts;
to claim priority in finding a solution or proving a theorem;
to advertise the importance of their subject to a broader audience.

Different LLM use makes sense for each of these different goals. I’ve used multiple approaches in my work so far. I plan to return to this section later to flesh out the principles guiding my usage in these cases more fully.

There are also various kinds of writing the serve an institutional or purely formal purpose. Often the point is essentially to stamp a seal of approval, by attaching your name to a statement. In such cases, nearly full LLM use seems like it could be appropriate.

Things like letters of recommendation may or may not fall into this category. A letter meant to demonstrate personal familiarity — say, supporting an academic job application at the post-grad level, or an external letter in support of a tenure case — deserves a fully or at least mostly human touch. On the other hand, for a letter supporting an application to an internship by a student who attended your 70-person calculus lecture, LLM use seems fine.

The distinction is basically: in the past, would you have used a form letter? If yes, an LLM is not much different; if no, keep doing it by hand.

Personal correspondence and emails should always be 100% human. (I hope this one is obvious.)