Let me start by saying this: LLMs are fantastic writing tools. They work by predicting what word should come next, and they are exceptionally good at it. Objectively seen, they can write gooder than most humans, and in a fraction of the time.
That said: I usually don't read for the prose. Beautiful prose is a nice bonus, but I actually read to get an insight into someone's head. Crucially - I want to read the author's thoughts in their own voice.
Text as a vehicle for feeling
When you build a cool project, I want to hear you rave about it. I want to see your passion ooze out from between the words of your blog post. Using an LLM to promote your new app bleaches your post of your real, authentic, human-felt enthusiasm. It replaces genuine biological excitement, joy, frustration, nervousness - with a facsimile of human emotion, produced by an unthinking machine. A machine which can emulate excitement, but has no idea what it actually feels like.
I talk to Claude every day for coding projects. I talk to Gemini every day for research. But I also talk to real people every day. Why? Because when I share a feeling with them, they can relate to it. When I text a friend to tell them I ran a half-marathon, I know that their reading of that message will evoke some kind of feeling in them. I know that their reply will have been written while feeling something, and that feeling will have decided the words they sent.
If I told Gemini I ran a half-marathon, it might congratulate me on the achievement. It might say all the same sorts of things a good friend would say. But it would be a facsimilie of enthusiasm. Gemini will have constructed this response based on a predictive model of what I expect to hear back. It could even reply with exactly the same message as my brother, and yet I would feel nothing while reading it. Because I know the machine will have felt nothing while writing it.
The Research
I recently read this article.
It explains that most humans prefer to read novels written by other humans - not surprising. But only when they know the book was written by a human. Give the same people a mix of unlabeled human- and LLM- written novels, and most people prefer the AI prose. What does this mean?
It means objectively the AI writes better. But that we give less value to AI text. I'd hazard a guess that part of this value is related to what I feel here: that most people don't want to read stories about fictional people created by mashing up and remixing existing novels. They want to read about fictional people serving as a veneer for real people, with events based on actual lived experience from other humans.
Writing for joy
I enjoy writing. I used to enjoy programming. I can no longer find any joy in programming knowing the Claude is both better and quicker than me. With code, it doesnt matter whether the compiler can feel the passion I have for such and such bug fix oozing out from between the monospaced words. The compiler is obviously incapable of feeling it, and it also rudely strips out any hand-written comments I might have lovingly crafted. In short - I see no reason to write my own code anymore, and doing so leaves me feeling hollow and useless, as if Im rubbing my face in my own redundancy.
But writing for humans is another story entirely. An LLM could probably have written this blog post more concisely while conveying more information. If that's important to you, feel free to stuff it into your AI of choice. But that's not why I write. I write in the hope that you might enjoy reading it, but to be honest, I mainly write for myself, and I do find joy it in. I used to fall very often while coding into a flow state - a state of mind where your attention is focused entirely on the task at hand. With agentic coding that's now very firmly a thing of the past - my attention is often divided between multiple Claude sessions, across multiple apps. Deep focused work is a thing of the past, and Claude has made it irrelevant anyway.
But I can still find flow state when writing. The work feels important, knowing that it cannot be achieved by AI - at least not without stripping away the scent of the human. When I write I organise my own thoughts, while hopefully also presenting them in a way that is clear and interesting to you the reader.
Disrespect
There's a lot going on here. It makes me sad to see people devalue their actual lived human experience. I find it tragic to see people en-masse so willingly hand over agency to machines. I think its a shame people are missing the chance to share their real, felt emotions with the world. But I want to go further. When people expect me to read an AI generated text, I find it disrespectful.
We all have limited time. We hate the feeling of our time being wasted, and I think this has become even more prevalent in the age of AI - so much to do, so many agents to check up on.
This is exactly why people use AI to write texts for them.
But the reader of your AI text also has limited time. And forcing me to read a 50 line Slack message that Gemini wrote for you based on a one line prompt is disrespectful to my time.
When I see a clearly AI written blog post, I skip it. It doesn't bother me - I don't need to read your blog, and you don't need my readership. No one loses. And when you give a presentation full of love, even when it's obvious you had help from AI, I'm still unbothered. But when I am forced to sit in a meeting at work and listen to someone read off from AI written slides - slides they are reading for the first time - that is disrespectful.
I think the core of this issue is that in the past, in writing a text the author would have to have to spend at least as much time as the reader would have to spend in reading - in fact usually many times longer.
With LLMs, this relationship has flipped. It is now possible for me to produce a 800 page novel in 30 seconds, and publish it online for you to spend 3 months reading. It used to be that, before I read anything, I knew that some real human effort had been expended in writing it. I knew that if a book was 800 pages long, it means someone spent years of their life writing because they were passionate about what they were writing. Again, it could have been 800 pages of drivel, but at least I had the guarantee that someone cared enough to stick it out.
No such guarantees anymore - in fact, when I come across a piece of text in the age of AI, there's a real possibility that I will be the first human being to ever read it. It's possible that not even the "author" has read it!
This is the core of the disrespect - as an author, you are offloading the cognitive load of processing the text you have produced to your reader. If you cannot read and understand the output of your LLM yourself, you have no business presenting it to other people to do the work for you.
To err is human
So this piece, and all of my pieces, are 100% human-written. I don't even ask AI to check my writing and correct spelling mistakes anymore. Because Im not reading for the prose and the perfect grammar, and I expect you aren't either. Im reading for the real human impulse, hidden inside the words, that inspired the author. And in fact, in the age of AI, I think the little errors even add something.
When I see text clearly written by AI, I cringe. I sometimes even feel a little bit sick. But its not always so obvious. LLMs can be prompted to write "sloppily, like a human", and with enough prompting it can be hard to tell the difference.
I already feel the need to differentiate my writing from AI-prose. I want to make it clear to you, the reader, that I, a fellow human, wrote this. I find myself deliberately missing punctuation to leave a scent of human in the text. To err is human, as Seneca said. Where a few years ago I might have spent hours proof-reading a text to make it perfect, I now force myself to publish without too much checking - I hope the unpolished nature of the text convinces you of my humanity, and makes it easier for you to read without being overcome with AI-revulsion every few lines.
The future
Despite everything I've said, I don't consider myself a luddite. I fully expect to see AI-prose proliferate further in the coming years. I already hear stories of people using LLMs to write wedding vows and eulogies, conquering even the last vestiges of human creativity.
But I also expect we'll see growing pushback. I wouldn't be surprised to see the emergence of human-specific online language in backlash against AI - a way for us to differentiate our voices from the tide of AI slop. We already have a number of tell-tale markers of AI writing: perhaps we'll see shibboleths developing in the opposite direction also, in the form of words that only humans will ever use.
I expect we will conciously unconciously diverge from the way AI talks in an effort to establish our own culture, much like a teenager rebelling against the mainstream, with AI our uncool parents. We might end up in a cat-and-mouse situation, where AI trains on our new dialect in order to continue to disguise itself as human, forcing ever quicker cycles of linguistic invention and disposal.
I find it personally frustrating that many of the hallmarks of AI writing are simply hallmarks of good writing. LLMs like rules-of-3 so much because we like rules-of-3 so much. Will an effort to differentiate human writing from AI involve the concious adoption of worse writing technique and deliberately unclear style? Or will these changes end up being purely stylistic with no impact on comprehension?
This is all speculation. Perhaps next week Google will announce the first Gemini model that can feel real feelings and I'll be forced to retract this piece. Until then I'll keep on writing the old fashioned way.
Its not AI — it's Luke Needham.
