Policy on AI Use

This document outlines my policies on AI use for my personal work. In a world where LLMs are frequently better than myself at certain things, this document is an attempt to balance the utility they offer with the integrity of my work.

Finding the balance between these two requires an understanding of the audience’s expectations. This can be clear, set by the medium or the task, or more ambiguously set by my intent. For example, I am writing this document with human colleagues in mind, to inform them of what to expect when they encounter my writing.

Simple Guarantees

Let’s start off with some conclusions:

1. I will never make you read LLM-generated text unless identified first;
2. When I use an LLM to learn things, I will learn the space and not the task. You are seeing my summary, not the LLM’s summary transformed into my language;
3. If you ask for my opinion, you will get my opinion: I will not feed the context you provide into an LLM to form the opinion for me.

The Axis

My guiding principle for AI involvement is a single axis:

On the far left we have cases where the writing just needs doing: transcribe this document; create a flyer for this bake sale, etc. Everyone knows the information already but it needs transforming into some specific form to fulfil a basic function. I am but a vessel through which flow the cliches of the genre.

On the far right we have cases where my opinion and experience matters. If someone wanted the LLM answer they can just ask the LLM directly. Examples are this document, feedback on what it’s like to work with you, a personal message in a card, or work on a novel approach for something.

But what about things where the LLM is better/more knowledgeable/etc than you? Either I can ignore this vastly useful repository of knowledge and give my authentically ignorant response (sometimes useful, where someone wants to know what I know), or I can learn the space using any and all tools at my disposal.

In this middle-ground it’s on me to learn the topic, not the task. The LLM is there to teach me things, not to tell me what success looks like, or teach me the way of doing the task. This distinction retains my judgement in the mix, and allows me to learn.

Examples

Let’s look at some examples:

Personal Messages – most right there is You’ll get my hand-typed home-baked words, without review, for better or worse. Authenticity matters most here: most audiences care more about what the subtleties of the words imply about the author than they do about the technical content of the message being slightly more precise.

Interview Processes – centre-right I’ll use LLMs to examine code and documents sent to me, but only after manual review and summary. Your earnest effort to get hired by me will always be met by a human effort to evaluate that input holistically, rather than through some mechanistic scoring process.

Presentations – mid-right LLMs might be useful tools for what needs to be in the deck, but the message and content must be mine. This means the words will be mine, and the diagrams will be designed by me (though they may be re-drawn by AI for clarity).

Code – centre- to far-left Some code just needs writing: an API for an existing domain model; tests; CI/CD pipelines. No-one is going to be upset if my opinion isn’t in these, so it’s a case of using whatever is the best tool for the job.

Some code, e.g. EDAs of datasets, requires interpretation and insight, and I will be unable to provide that insight or novel approach if I don’t stay close to it. I might not write a line, but I’ll be directing the LLM in specific ways.

Code is a special case: the audience is (directly) the computer, so the authenticity of the spec matters more than the authenticity of the actual code itself.

Documents – centre-left Depends a lot on the need. Often, something like a job description will benefit from an LLM reviewing it: it’s able to complete the cliche of the form and identify gaps in my work. If I am using a hybrid approach you’ll generally get one of two things:

1. I write the document ‘humanly’ and get an LLM to review;
2. The document is planned using an LLM to understand the form (not the content of this specific doc) and then written by hand.

In either case you won’t have to read LLM-generated text unless specifically flagged.

Summary

If you’re reading something ‘by me’, then the meaningful part of the document is definitely by me. This means the reasoning and bias in it are mine, not the judgement of an LLM on the task. Where this is not the case I will clearly state it, e.g. here’s some LLM output on this topic…

If it’s more important that I just get the damn thing done, then I’ll probably be using generative AI to speed up research, understanding, synthesis, and summary.

The emdashes are mine.