Using AI for writing

Summary

I write because I enjoy writing. This is a key reason I don’t delegate writing to AI. There are, however, four use cases when I use AI for writing.

• As a proofreader and copy-editor

• As a research assistant

• As a writer, when my voice lends no value to the output

• As a tool to synthesise unstructured thoughts

I’m a sceptic when it comes to using LLMs for writing. But as Martin Fowler recommends, my scepticism is absolute, and I’m sceptical about my scepticism. A recent conversation with my boss, John Reid-Dodick, has led me to examine that scepticism and document a more nuanced take on the role of AI in my writing workflow. 

Before I proceed, I must clarify my relationship with writing. I have zero claim to being a good writer. For me, writing is a way to think. Pausing to construct a sentence or a paragraph, seeing the words emerge on the screen, critiquing them as I type, and cutting, pasting, and reorganising sections in my document are part of my thinking process. I think of words as the digital exhaust from this process. 

LLMs struggle to think on my behalf because they don’t know what I don’t know yet. If I could tell an LLM what to write, that’d be most of the effort, right there. Producing words isn’t writing. At least, not for me.

Tone, stance and voice

When I write as a technology professional, I often assume the roles of educator, author, and sometimes advocate. When I write about wildlife or photography (here’s an example), I assume the role of a storyteller. I’ve found AI tools inadequate when assuming the right tone and stance for different kinds of writing. 

Now, I have tried giving different LLMs a folder with 200+ articles so they can glean my writing style, and I haven’t been satisfied with the results. Of course, my dissatisfaction could also mean that if I were a dispassionate reader, I wouldn't enjoy my own writing. I continue to wrestle with that possibility. That said, I’ll list why I remain dissatisfied with LLM-powered writing.

• LLMs often try to play safe. If I write a sharp sentence that may challenge the status quo or deviate from an established norm, it advises me to flatten my tone. I don’t want to be offensive when I write, but I also don’t want to be bland. If I’m not being provocative, then I don’t see the point of regurgitating new words to describe topics that people already agree on.

• AI tools can often take an academic approach to writing. For example, when I ask Perplexity to “read, rate, and review” an article, it’ll often ask me to front-load evidence so I can bolster the case I’m making. In some cases, this advice is helpful, but in most cases my experience serves as evidence, and I prefer avoiding the verbosity that comes from citing evidence.

• AI often expects me to present conclusions when I don’t have any. Sometimes, I want my readers to sit with a set of tensions and trade-offs. Perhaps, over a series of articles on a topic, I might arrive at some conclusions, but I don’t expect every essay to follow the evidence-exposition-conclusion pattern.

In a recent article, Jim Highsmith describes the idea of “practitioner voice”. Here’s how I’d define that category of writing.

Practitioner voice emerges in writing when authority comes from the author’s lived experience and their judgment is visible in the text. The author trusts readers to recognise the pattern from their own experiences and is comfortable naming tensions without resolving them.

When writing as a professional, I want to write in “practitioner voice”. I don’t expect everyone to agree with what I’m saying. My only goal is to express my current thinking as clearly as I can. 

Uses for AI in writing

All this said, I do employ AI for specific writing use cases. In some scenarios, it works behind the scenes. In other scenarios, I let it do much of the work.

As a proofreader and copy-editor 

I use Grammarly and ProWritingAid with Google Docs to simplify my writing and avoid basic punctuation and grammar errors. Grammarly helps me follow some basic style guidelines I’ve set up — for example, I avoid nominalisations and passive voice when I write. ProWritingAid scores my writing using the Flesch-Kincaid readability test and gives me suggestions to simplify what I’ve written. 

I find it annoying to keep these tools on all the time because they prevent me from getting into a flow state with writing. So, I enable them after I’ve arrived at a satisfactory draft, and then the editing process becomes enjoyable.

As a research assistant

I often use AI to compile my research on certain topics before I write. It’s a struggle to double-check AI’s research, though. Since most chatbots are sycophantic, they phrase search results in a way that confirms your biases, so if you don’t check the original sources, you might end up misciting your references. In recent months, I’ve returned to the old-school approach of using web search. Since Google’s search experience is quite poor these days, I pay for Kagi, which is what a disenshittified Google might be like.

As a writer, when my voice has little value

Sometimes, I just want a coherent write-up that summarises content from different sources. For example, I had to document a decision record at work based on an email thread, two documents, and three conversation transcripts. AI is great at getting me 90% of the way there, and I can handle the remaining 10%. PR summaries, commit messages, meeting minutes — these are all examples of documents I’m happy to delegate to AI. I use custom skills in Claude and Glean to write such documents. 

As an assistant that mashes up unstructured thoughts

As you may know, I’m a moonlighting wildlife photographer and travel writer. When in the wilderness, I often record voice memos about how I’m feeling and what I’m experiencing in those moments. When I’m ready to write, I feed transcripts of those voice memos to an LLM and then use a custom skill to derive a structured and somewhat coherent write-up. 

But this approach has its limitations. It’s an aggregation of my thoughts, but never a complete story that I can publish. For example, when I wrote The Cry of the Kgalagadi, my emotions were coloured by my father’s passing. I didn’t predict his demise when I started the trip. I couldn’t predict the emotions I’d experience when I began to write the article. I had to get my fingers on the keyboard for my story to take shape.

Much of my travel writing is about the feelings I experience in retrospect, after editing a bunch of images and letting my initial reactions fade a bit. Which brings me to the actual tensions when using LLMs for writing.

The reason for writing may dictate the level of LLM use

I enjoy writing. I may not be good at it, but I do enjoy it, much like I enjoy photography. I may have mentioned earlier that my wife is a crossword geek. Even if an LLM can solve her crosswords for her, she won’t use an LLM because she doesn’t want to give up that enjoyment. Much like it’s a struggle to find the word that fits in 17-across, it’s a struggle to get the first words on the page. That struggle is part of a gratifying process. 

Someone could argue that I need better prompts, richer context or more capable agents to help me write. My question would be — why? Writing is not a chore I want to offload. It’s a process I want to savour. 

Perhaps LLM-assisted writing allows people who don’t enjoy writing to produce something coherent and readable. I think that’s OK. Perhaps it’s more productive to help someone in content marketing produce more articles, since their job demands that sort of throughput. I think that’s OK too.

None of those conditions is true for me. Often, I don’t need the speed advantage that AI may offer. Of course, my relationship with writing could change in the future. Until then, I expect AI to remain a peripheral tool in my writing process.