Guilty Feet
How AI changed my writing without realising it
The video for this article is here: Guilty Feet
Recently, I was turning some articles that I wrote between 2018 to 2020 into video scripts. Reading those articles again felt like a breath of fresh air, not because they are necessarily brilliant but because I recognise myself in them. The style felt freer. I recognised how I approached and developed an argument, how I would suddenly go off on a tangent and would never fully find my way back to my original argument. Or I would veer into sarcasm or humour and some of the sentences were far too long. Not great writing, but not bad either, and, crucially, all mine. I liked the looseness and the freedom of it.
Then, I contrast this with articles I wrote over the last couple of years. Recent pieces seem straighter, tidier, more segmented and more controlled. The style feels boxed-in, even templated.
Of course, I knew why.
I had always taken a hard line against AI-generated writing. I wrote an article where I called it a ‘sin’. A friend sent me a video message for my birthday this year and she said ‘No AI was used in the making of this message’. This is proof that my position on AI writing was well-known and I hadn’t realised that it had started to become a joke. A funny one at that.
Why do I take such a stance? Because writing is bound up in my professional identity. I work in health communication, specialise in information design, have expertise in readability and language in general. It follows, then, that if you cannot express an idea effectively in writing, then there is something odd about claiming expertise in fields concerned with how ideas are articulated. If you cannot write well, you simply do not have expertise in communication. Believe me, plenty do not. It’s often a shibboleth, that separates the person who can from the person who cannot, i.e. the imposter.
I am definitely not a puritan with regards to technology or a Luddite. I am very comfortable using AI-generated slides for my videos, teaching materials and presentations because I do not claim to be an expert in graphic design. Those slides are only visual aids used for teaching and they are not frivolous or suck-it-and-see generations, where I just wait for random designs to come up, abracadabra: they are deliberately steered by me. It is the same with music. I use generated music for the intros and outros for my videos. I developed a concept of creating a Morse Code message of our strapline ‘Putting patients at the heart of research, globally’ and then prompted into different musical idioms: Al-Andalus, slow jazz, blues, Balkans brass, crazy ideas like merging Portuguese fado with Mussorgsky or Guatemalan marimba on Classical Arabic scales but played through different instruments and so on. It’s fun. Of course, it is not music that would move anyone emotionally, but it us just ‘okay’ — and only some of it is okay and only then in a quirky way– for little youtube video essays that no more than 200 people have ever seen. We have reached the dizzy heights of 56 subscribers.
So, I’m not trying to pass off expertise in graphic design or music. How could I? And I understand why there is the current backlash, the aesthetic disgust against AI slop. I cannot bear seeing the oh-so-obvious AI-generated diarrhoea in LinkedIn posts. It repels me. It has caused me to take a few steps back from that platform. It feels like everything has been washed through the same machine. The Blandscape of Enshittification. Somebody else might see my AI-generated slides or music and have the same visceral reaction. Why use AI-generated slides when I could use a human graphic designer to do them? Why use AI-generated music when you can employ the services of a professional musician?
I get the argument, but I would have not done that anyway. I have used human graphic designers on several occasions. A well-known Jamaican artist designed our company logo back in 2012, which we love. But for presentation slides, the graphic designers I have used since then have been a disaster, very unimaginative and expensive. Just think of the proliferation of look-alike company logos during the 2000s that were simple variations of the Nike swoosh and paying a king’s ransom for the privilege. They were everywhere in the UK at that time and persisted for years and years. I am yet to meet a human graphic designer that understands what I want to do. I had several conversations with marketing companies in 2024, parted with 12000 pounds and didn’t like what I got in return. I scrapped everything and started doing what I needed their help for on my own with no skills and outputs that clearly show a learning curve, but what I do meets my personal needs.
So, what would I be doing instead of the generated slides? We’d be using Shutterstock or Canva stock images, which you can see in earlier videos. These are images and videos that tens of thousands of other companies have also used and they only ever partially illustrate the visual point I want to make. AI can produce something more specific to my thinking than generic stock material. What about taking my own photos? I’m not going to do that. Should I learn piano to compose the music as well? Of course not.
But let’s get back to my observations about my changing writing style. And here comes an admission. I never imagined George Michael, a gift to humanity though I truly believe he was, would provide the moral vocabulary for this essay, but here we are:
Guilty feet have got no rhythm.
I had been condemning AI prose while some part of me knows that my own feet had already wandered onto that particular dancefloor. AI had infiltrated a capability that I regarded as a core part of who I am.
On Substack, I posted an article on modern working life that I wrote last year called Mr Dobalina, Mr Bob Dobalina (let’s see if you catch the cultural reference). I remember writing the article clearly. It was a Sunday afternoon, in my office, late July last year. I spent around 4 hours writing it. So, I know very clearly how the article came about: I first stared at a blank page and a blinking cursor. Substack had introduced a new AI detector in July this year and I was entertained by all of these people making AI confessions because the new detector was suddenly unmasking everybody who pretending writing expertise of great acclaim. The typical excuse was: ‘Yes, I use AI, but I am still the author of the ideas’, kind of thing. And so I decided to put the Mr Dobalina article through it. It came back as 80% human and 20% AI-assisted. How is this even possible? I wrote the article myself.
Early in 2025, I wrote an article with the title ‘About AI assistance’ which was a follow up to the one I mentioned before ‘About AI generation’. I had already been using AI detectors because we had a guest writer who would occasionally write for us and I was suspecting that their articles were AI-generated. The AI detectors were unanimous: scoring around 99% AI. So, the only human input was their name! They admitted to it but gave the same excuse as the Substack ones: I told it what to write. They are my ideas.
Be that as it may, it is not their writing. I was paying for their blog writing services, not those of a machine. I felt conned.
This case was not in dispute: the guest writer had used AI generation entirely and admitted to it when questioned. But, then I was curious, so I put in some of my own articles, such as the ‘Belated Introduction’ article from our blog. I knew this was entirely written by me but it came back as 70% AI-generated. Simply not possible. I was a fairly late adopter of LLMs and I wrote this article just slightly before starting to use them. I put in a chapter of a book that was published in 2007 and it contained a substantial percentage of AI assistance. Not possible! Chapter 3 of the Book of Genesis came back as 30% AI. And then the biggest reveal: language courses and glossaries that I generated for my own learning purposes that had no human input whatsoever — other than my own prompts — all consistently came back as 100% human-written. They were entirely AI! So, clearly, the detectors can be wrong. Some of these detectors have more than the whiff of scam about them. For example, they also sell a service where their AI tool ‘humanises’ the AI-generated / AI-assisted text. So, automatically, there is a clear conflict of interests. Come on, the Book of Genesis, for goodness sake!
So, returning to the article I put on Substack, I knew there was something accurate about the 80–20 split. And there was a reason for it: AI entered before the writing. I was not asking it to write the article, but I was feeding it my notes, often handwritten, my observations, development of arguments and my examples. Then, I would ask it to organise them into a coherent scaffold that I would later base the writing on. So, the ideas were mine, the prose was mine. AI was simply helping with how those ideas were to be organised. The problem with that is that the structure is no longer entirely mine. And those structures carry a written signature with them.
Here are some examples of the notes. These are the notes from this particular article that I cross out as I progress through the writing:
Here are two examples from previous articles where AI provided the scaffolding. And there you can see how, not only how the article is to be organised, but the words and phrases that survive into the finished article.
Exhibit 1:
Exhibit 2:
Suddenly, I am using words that I wouldn’t have normally used, such as “quietly”. Here is an example of it creeping into a final article, ironically about erosion of human roles due to AI:
Get Mark Gibson’s stories in your inbox
Join Medium for free to get updates from this writer.
“We talk about automation replacing jobs, about AI quietly eating away at once secure roles, take accountants as an example.”
“Epistemic” is a word I wouldn’t normally use either. Even though I write a lot about knowledge production, linguistic philosophy, cognitive burden, etc. and the fact that most of my work is within the cognitive sciences, my natural writing lexicon hardly ever includes it. But here it is in an article of mine:
“It was not just a wave of bad ideas, it was a collapse of shared reality, epistemic instability. There was a sense that truth itself had become unknowable.”
Other words that are often cited as tell-tale AI vocabulary are words that I would normally use. Here is a list I compiled of lexicon typical of AI-generated prose and I give an honest view of the frequency that I use these words and phrases in natural writing — before and after LLMs:
So, clearly, not proof of AI at all. In fact, it is unfair to earmark these as clues to AI generation. “Resonate” as well is seen as a clear use of AI. That may be the case, but not in mine because “resonance” is a really important concept in the field of cross-cultural health communication. It is a term we should be talking about all the time. So, that’s a false flag, or a red herring, as we say.
As a side note, this is another thing: I never construct my own tables anymore. I feed them into an AI tool and it does the rest. I never humanise the text in tables. I do the same with .srt files — or often to cut down on words, which I should have done with this article. All of that content would be identified as AI-generated text. So, all the video scripts that have been put into .srt format would appear as AI-generated and the same with the tables, as with text where an LLM has trimmed the word count. And when I have to provide summaries of what a new YouTube video is about, I rarely write those myself. By the time I get to that stage in the article / video production, I cannot bear the idea of summarising. I’m comfortable with AI assistance for all of the above.
Getting back to the scaffolding, I noticed, but in retrospect, that whole rhetorical phrases found their way to my final articles. An example of this is: “Why this matters”. I don’t think I would have used ‘matters’ before LLMs in this way, so frequently.
And then there are perfectly polished metaphors, such as these from one of my articles:
“Let the machines hold the noise, while humans focus on the signal.
Let them store the map, but humans choose the road.” (Yikes…)
And then there are the perfectly alliterative titles. Everybody’s articles have perfect alliterations, even writers whose first language is not English and who I know don’t even know the difference between singular and plural verb endings or who can barely put a coherent sentence together when speaking, suddenly have beautifully crafted article titles, such as:
Signals, Systems and Safety
Words, Worlds and Ways of Knowing
And there is another practice of mine where I was a bit careless for a while. When articles became too long, I would ask an LLM to divide them into two, sometimes three articles. And to create two apparently complete pieces, the LLM generated a short conclusion that signalled the end of one article and a 1–2 sentence introduction for the following. Here are examples of this (‘in a world’ is a dead give-away):
“In a world of globalised studies, concept elaboration is not optional. It is essential.”
And
“In a world that is increasingly global, visual communication is often seen as a universal shortcut.”
Dear Lord…
So, guilty feet and all that. I think we are all indulging in some form or another of hybrid writing. To various degrees we are leaning onto LLMs from minor assistance, like the creation of tables or proof reading, to scaffolding, to preserving AI text in final articles to wholesale text generation.
If we are honest with ourselves, each one of us is somewhere on the following spectrum:
At what stage is hybrid still your writing? And if you are using AI to generate the entire article, then you cannot, you simply cannot, claim any expertise for yourself. This leads to, yes, I will use the adjective in this case, epistemic overconfidence at best, but then you will sink neck-deep into the Dunning-Kruger effect. This is the tendency for people with limited knowledge or ability in a field to overestimate their own competence, partly because they lack the expertise needed to recognise the limits of what they know.
And this brings me back to meeting myself in the pre-AI articles. When I read them again, what I recognise is not the imperfections or some mystical human essence. It’s the freedom of movement that I liked most. My earlier writing could overdevelop one argument, underdevelop another, interrupt itself, change speed, become sarcastic (a perennial problem of mine), go off on a tangent and carry on.
The newer prose had acquired guardrails… or prison bars. It feels that it is in a prison cell.
So, my writing has changed. The old ways, imperfect though they were, seem to have died. Do I mourn them? Do I resuscitate them and bring them back? Or do I embrace the Blandscape of AI assistance forever more?
Thank you for reading,
Mark Gibson, Leeds, United Kingdom, September 2026
