I’ve written fairly regularly about how I use GenAI in my writing process, including why I don’t think AI is very good for a first draft, and how I believe these technologies have flattened the strata of production and publication for digital texts. But the way I write with and without AI continues to shift, and as someone using artificial intelligence daily, I think it’s important to keep updating these posts and exploring what the technology means for writers.
My Process
For a long while, I’ve been using a fairly consistent process to write articles for this blog and various kinds of writing, like draft sections of journal articles, social media posts, emails, and other communication. I write two thirds or more of that content verbally. Because I’m a creature of habit, I’m using more or less the same applications for this as I was 12 months ago, but some of the processes have changed slightly.
If I’m out for a run or a walk, I’ll use the iPhone Voice Memos app with my AirPods. If I’m driving, I’ll do the same, but I might clip on a DJI mic which cuts out the road noise. I might run for 45 minutes to an hour, and if I’m driving from home to Melbourne or Adelaide, four or five hours on the road gives me ample time to record one or more articles.
When I get back to my desk (we don’t have any mobile or Wi-Fi connection out on the farm), I share the recording directly to the Otter transcription app. I could use any number of apps for this, but the iPhone’s built-in transcription is utterly hopeless, even since apparently improving their AI features. Otter uses OpenAI’s Whisper transcription model under the hood, which is the most reliable model I’ve found.
Of course, transcription is nothing new, and tools like Dragon Dictate and even Microsoft Word have had transcription as features since long before generative AI was a thing. But I need transcription that can handle background noise, long pauses, the sound of me gasping for air as I struggle through a run, and occasional car horns as I idle at a traffic light. Those older transcription models can’t do any of that, and for the most part, I also can’t use them when I don’t have any internet access. So Voice Memos to Otter it is.
The transcription that comes out of Otter offers summaries, again using a GPT model, but I found it is not capable of interpreting and retaining my original language without ChatGPTifying it. So the next step is to go to a more capable language model, usually Claude. I’ve been using basically the same prompt for two years, and I’ve shared it before, but here it is again:
I will provide an audio transcript. You will retranscribe it with appropriate formatting such as paragraphs, lists, headings and subheadings. You will correct obvious transcription errors (mis transcribed words, repetitions) but otherwise you will NOT change the wording. you will not attempt to edit or improve the wording, or reword or rephrase outside of the errors. At times, I give instructions directly to you, the large language model, for example instructions on how to create a table or do certain formatting. You’ll know it when you see it. Follow those instructions. Do you understand?
That gives me an output which looks something like this, formatted with headings, lists and paragraphing and ready to copy-paste into my blog.
Recently, I’ve added another step. If I have specific articles in mind to include in the blog post, or if I’d like to flesh out some of the ideas with recent articles, I now take the cleaned up transcript to ChatGPT’s “reasoning” model and use a prompt like this:
Find appropriate places to include these articles in this blog post, perhaps identifying suitable pull quotes from the articles. Following that, search online for recent articles, Substacks and Medium posts written by authors on the topic of using AI in their writing process. Provide me with a list of links to follow up on.
As you can see here, the links that I provided this time were posts from my own blog, which I’d already referred to. That’s because when I’m out for a run, I obviously don’t have my computer in front of me, but I do have a pretty decent working memory of the articles that I’ve written over the past few years, and so I know which ones I might want to refer to. I can’t remember exactly what I said in those articles, but if I’m pretty sure that there’ll be quotes that I can pull back out of them, I might as well send ChatGPT off to do that for me.
For the additional research, I found that OpenAI’s o3 does a generally better job than any of the other models of finding interesting and relevant articles. It sometimes misses the mark (in this instance, it brought up a bunch of junk posts along the lines of “improve SEO with AI!” and “How I write 50,000 articles per day with ChatGPT for fun and profit!”), but it also often surfaces articles I never would have found or read otherwise, like this Wired story exploring how many Substacks might be written by AI.
Those subsequent articles might feature in the one I’m writing, or I might bank them for another day and another article.
Another additional step is purely administrative. If you have a blog on WordPress, you’ll know a few tricks for inserting links (cmd+k is your friend), and you’ll also know how important it is that your blog contains plenty of links to your own articles and external sources. Sticking with GPTo3, I might now ask for the entire article verbatim to be presented back to me as HTML code, incorporating the links to the various articles that I’ve chosen in the code mode.
This is a very lazy way of copy-pasting the entire blog into WordPress without having to manually copy and paste each individual link. It probably saves me 10 to 15 minutes for each article that I write. But if like me, you’re writing two or three articles a week, that adds up.
Once the article goes onto the blog, I convert it back into blocks – the visual editing format of WordPress – and do the final edits myself. There will be errors, though far fewer than using this exact same process a year or two ago, as these models continue to improve both in the quality of audio transcription and in the ability to interpret the transcribed recording.
Following these edits, I’ll add images, any extra links that I may have forgotten about, contact details and so on. And for one final minor AI use, I’ll usually use WordPress’s built-in AI features to generate the slug or excerpt, frankly, because I can’t be bothered. The only things reading these are search engines anyway.
The Missing Parts of the Process
But I realise, whilst speaking this article aloud, that there are important parts of the writing process I’ve missed, and these parts relate almost entirely to what John Warner is talking about in his book More Than Words. Because everything I’ve written about so far is the process of getting the ideas from my brain to the screen, but I haven’t talked about how I formed those ideas in the first place.
ChatGPT cannot write. Generating syntax is not the same thing as writing. Writing is an embodied act of thinking and feeling.
Nine times out of ten, that happens with pen and paper. I’m an obsessive journaler, which is a fancy way of saying I have a diary. Quite a few, in fact. I think I’m up to number 22. As well as my little A6 pocket notebooks (which are, of course, various kinds of limited edition Moleskine, because I am, if nothing else, a total wanker) I own notebooks of every size, colour and texture. I use these variously for PhD notes, business and financial planning, and scribbling down ideas for blogs.
I’ve almost always got at least three notebooks and four pens on the go, just in case. If I tip out my bag right now, here’s what I’ve got:
- An A5 square-ruled Rollbahn notebook with PhD notes
- A B5 Leuchtturm composition notebook for business stuff and ideas for blog posts and professional learning (I will die on this hill: B5 is the best paper size for Serious Work)
- Two A6 Field Notes: one for taking notes during calls, and the other for social media post ideas
- A Moleskine Year of the Rabbit A6 pocket diary (wanker)
- A Pilot Custom 823 loaded up with Iroshizuku asa-gao blue ink
- A TWSBI Eco with the De Atramentis Black Document ink (it’s waterproof!)
- Another TWSBI Eco with Diamine Red Dragon ink
- A Lamy 2000 in green with, of course, green ink – actually a local Robert Oster Peppermint from Mount Gambier
Why the hell am I giving you so much information on the pens and paper I carry around with me? Well, because it’s a particular interest of mine, and I don’t care if you don’t care, but also because when I get swept up in the process of how I’m using artificial intelligence in my writing, I often forget that the vast majority of that writing starts here.
The conversations I have with schools, and the PD sessions I run online and in person, often end up with scribbled ideas in one of the Field Notes notebooks. If I speak to two or three schools and the same kinds of questions come up again and again, I might scribble down a few thoughts in the social media notebook or something more thoughtful in the Leuchtturm notebook.
As I’m working through journal articles for my PhD literature review, or scribbling down ideas whilst analysing the data, those ideas might migrate from the red notebook to the black one. The different pens are just for fun; they have almost no impact on what I’m writing, except for the fact that the Pilot frequently leaks and destroys entire pages of writing, but I love it anyway and refuse to stop using it.
More broadly than these notebooks, though, and the scribbled half-formed ideas they contain, are the conversations which inform them. Some of those discussions unfold on social media, in the comments of my blog posts and via the contact form, through Zoom calls and online and face-to-face professional development sessions. The whole system feeds itself.
So when I head out for a run and turn on my Voice Memos app, I’m not starting from scratch. I’m starting from that swirling, generally unintelligible mass of smudged ink. I haven’t abandoned the Old Ways yet.
Advantages of the process
Since I started using these technologies as part of my writing process, the volume of writing I produce has increased, and I think the quality of thinking represented in those articles has improved too. To speak confidently, even if you’re talking to yourself, you need at least half a clue about what you’re talking about. Unless you’re a politician and can hire someone else to write your thoughts, of course.
My writing is much more prolific than it was in 2022. In 2024 from January to December, I wrote at least two and sometimes three or four articles per week, most of them hovering around the 1000 to 2000 word mark, which is long for online articles of any description but especially for blog posts. Alongside that, I was writing academic publications and the early drafts of my thesis.
I never would have written that volume of articles if it hadn’t been for this process. The way that I am wired means I can either focus exclusively on a task for four or five hours, forgetting to eat or drink, or I can’t focus on anything at all for more than 20 seconds. Being able to write whilst I’m out for a run or driving forces me into a space, physically and mentally, where I can concentrate on my ideas. It doesn’t matter that the piled detritus of my desk, the dozens of half-filled notebooks and half-empty pens, are in utter chaos, because when it’s just me and the voice memo, I’ve got no choice but to get the article written.
Being able to use artificial intelligence for some of the administrative stuff, like the coding, has also made my articles more interactive, more flexible, and more appropriate for the form of blogging, allowing me to do things like turn an idea about deep fakes into a little game that became incredibly popular and itself reinforced the original ideas. None of these things would have been easy or perhaps even possible before I started using generative artificial intelligence.
The Downside
The downside, of course, is that I spend much of my time feeling like a total hypocrite, given much of my writing, including a lot of the articles written with this process, is writing against generative artificial intelligence. These power-hungry, biased, privacy and copyright nightmares aren’t easy or comfortable to set aside.
But one conclusion I have reached through a few years of continuous use is that we absolutely shouldn’t be putting the sole burden of responsibility on the individual users for fixing the systemic problems with this technology. And if I make 100 or 1000 or 100,000 educators more aware of these problems, and they use that knowledge to push back against systemic adoption of platforms like ChatGPT in education, I feel like I’ve done my job.
I use image generation a lot less than I did in the early days, because I know through my writing research a lot more about the problems with that form of generative AI in particular. I experiment more and more with local and open source language models, waiting for the day when they pull level with Big Tech, which I think is likely coming at some point in the near future, given that the capabilities of these technologies are likely plateauing.
And of course, because I’m human, I use the technology because it’s effective, because it works, and because I enjoy it.
The future of writing
This is basically an update on those earlier articles about how I use AI, and there have only been relatively small changes in past 12 months, but my hope for the future of writing is that some of this knowledge extends into secondary and tertiary education.
If I were to update my CV right now, I would probably put “writer,” or perhaps the slightly more impressive sounding “author,” as my main job, even though I spend less time tapping away at the keyboard overall than I did three or four years ago. It’s been a long time since writing was a purely text-based affair, and it’s been a long time since writing was the best way to assess students’ knowledge. That might never have been the case, although writing is expedient and scales well for assessment purposes.
Generative artificial intelligence technologies have made written assessments of knowledge all but useless outside of supervised conditions. That doesn’t mean that writing isn’t important. It doesn’t mean that thinking through writing isn’t important, either. But it means that my experience of moving in and out of different platforms, drafting verbally, and even bringing in multimodal elements might become more and more common.
I hope that’s true, because I’m having a great time.
Cover image source: Portrait of Lionel Davidson, FG4206-3-27, British Library. Shelf mark: FG4206-3-27. Available at: https://imagesonline.bl.uk/asset/30360/
Want to learn more about GenAI professional development and advisory services, or just have questions or comments? Get in touch:
Leave a Reply