Writing and AI in 2026
I only started writing in public after the release of ChatGPT in late 2022. I somehow felt an urge to share my thoughts with the world, and learnt a lot in the process. But what changed? And how has writing changed with Artificial Intelligence? This article explores the interactions between my writing practice and the development of Generative AI.
Getting started
No more excuses
First, ChatGPT seemed to lift a lot of the technical barriers associated with maintaining a blog. This is not only about code. It is also about design decisions and understanding how something can be done (more on this later). What framework to use (Wordpress, Jekyll, Quarto…)? How to host the site? Should hosting be free? If you are interested in these technical details, I have published an early article on this topic.
GenAI models then were nothing compared to the coding agents of today. You can now tell Claude Opus 4.6 to create a git repository, build a Quarto blog and run until it is all working. Back then, I had a few things to figure out. Yes, as I am getting closer to the age of 30, I am starting to talk like a dinosaur.
Chatbots have made it much easier to get things done. In the past, you had to look for a YouTube video or a wikiHow page (speaking to the other millennials in the audience), hoping to find the way to do something. If anything deviated from the tutorial’s list of steps, you were alone. You could still ask on Reddit or Stack Overflow, at your own risk.
I do not think that I could have published a blog and a book without these models helping me understand the process. There were so many obstacles. Every step of the way, I could ask Gemini or ChatGPT to help me fix a strange Quarto bug or navigate Amazon’s dreadful UI.
Authenticity and empathy
As I came across more and more AI-generated text, I started valuing authentic and personal writing. There was now a way to answer questions very easily. But somehow, I wanted more. I was missing human connection and empathy. I wanted to share the human experience with the world, as a reader and a writer.
GenAI models are trained to be helpful and please everyone. As a child growing up in a family with divorced parents, I know that this is impossible. In my view, this is why a lot of today’s AI-generated content looks odd and aggressively mediocre.
I am under no illusion that I cannot be fooled into thinking that AI-generated text is from a human writer sharing their joy and suffering. I can also imagine a world in which I could give my existing writing (articles, journals, book, emails, WhatsApp messages, call transcripts…) to a GenAI model, that could fool others into thinking that I am the original author. This gets into scary territory, and Roland Barthes would laugh.
The Mechanics of writing with AI
Proofreading and editing
Over the past years, I have found that I produced my best work when I separated writing from editing. This is a very common piece of writing advice, that I recommend to anyone building anything.
When writing this article, I set a timer for 25 minutes and started writing without stopping, this sentence is roughly where I got to after 21 of these minutes. I know that I can write without worrying about formatting as I and a GenAI model will be reviewing later.
After a quick proofreading round, I can paste the draft from a Google Docs to Gemini, with a prompt including the following elements:
- Use markdown
- Correct typos
- Make minimal edits to the prose
- Make sure links open a new tab
- Format notes in a given way
- Correct and format mathematical notation
- Propose 10 ways to improve the article (very important)
In the meantime, I paste the draft into a .qmd file (Quarto Markdown file format) in my blog code repository in VS Code. Once the Gemini review is done, I paste the code into the .qmd editor and use VS Code’s built-in diff viewer (see image below) to accept or reject the model’s edits one by one.
Getting unstuck
Using GenAI models is a great way to get unstuck. Being stuck is underrated though. After the release of GPT-4, I used the model all the time to find easy examples to illustrate complex Computer Science concepts (like Algorithmic Complexity).
After a few weeks of enjoying my newly found power, I was struggling to come up with these examples on my own, in class. This was an ability I was very proud of. It was now slowly fading away.
This was my wake-up call. I then got (back) into the habit of getting stuck. I now leave most of the problems I struggle with to the following day and a fresh pair of eyes. The solution generally comes to me on a walk or a bathroom break.
I also write emails in German and Italian before asking models to correct them. When they do, I ask for a list of errors. I also refrain from pasting the output of the model back into the draft, and prefer correcting the errors one by one.
Generating own images and formulas
GenAI models are very good at plotting and building diagrams programmatically. I frequently use them with Matplotlib and Graphviz. Building good visualisations in Python can be time-consuming. It still takes time with coding agents, but much less. It becomes much easier to fill a book with original and meaningful illustrations.
Models like ChatGPT or Gemini are also very good at LaTeX, a scientific library used to compile most of the papers printed in academic journals. As I write on Google Docs, I can just type formulas in ugly pseudo-English and get nicely formatted mathematical formulas:
Sigma i from 1 to n = n(n+1) /2
to:
\(\sum_{i=1}^{n} i = \frac{n(n+1)}{2}\)
Research and fact-checking
This is where models have dramatically sped up my writing process. When looking for concepts or definitions, Gemini or the AI overview can now answer most of my questions in a few seconds.
I still enjoy going through Wikipedia entries (a fantastic organisation, consider donating to the Wikimedia Foundation), but on my own terms. When I am ready to traverse the tree of human knowledge. Not when I am trying to get things done.
Using these models, I can also easily fact-check mathematical formulas, calculations or statements about the world.
What now?
The speed of innovation is incredible. Just 6 months ago I was struggling to use coding agents. Over the last 3 months, I have barely written code on my own.
Last week, I have published my introductory book to Machine Learning (you can find it on Gumroad or Amazon). In line with one of my favourite sayings: “Behind mountains, more mountains”, I immediately started a second book project. I guess that you just have to enjoy hiking to have a good time.
To help me in this writing project, I decided to use the codebase of my first book (list of markdown files and scripts to generate images) to create a CLAUDE.md file. This is a set of instructions that are always sent to AI models as instructions.
It goes without saying that I used Claude Opus 4.5 to generate such a document. The model went through all the code to consolidate a list of instructions about my tone of voice, how to format notes and exercises, the way to approach new concepts and give examples. In theory, I could now ask an Anthropic model to generate a brand new book on any topic with my voice and style (scary, and a bit too ego-trippy for my taste).
This does not interest me at all. The whole point of writing is to find joy in sharing knowledge with the world. Still, this set of instructions is saving me a lot of time, doing many tasks I used to do manually, like formatting exercises and corrections.
Another idea that came to my mind was to use this process to write books in the voice of my heroes; like David MacKay, George Orwell or Albert Camus. The three of them died way too young. The idea that we could use AI models to write the world through their lens was enough to get me excited.
What a time to be alive.