Not long ago, I was asked what kind of support I needed as an SEO and content strategist. I wanted to work again with an experienced copywriter who learned the craft before there was a machine to lean on.
On the face of it, I couldn’t justify it. We already had freelancers who could produce drafts much more cheaply with AI and expertly researched briefs. Theoretically, what exactly is the expensive writer adding? I hated that I didn’t have a very good answer.
I wasn’t defending the sanctity of human writing and nobody was arguing that quality didn’t matter either. For a while, I wondered if I’d simply lost it, treating a business-phone-system comparison as though Hemingway might stop by to inspect the prose. Maybe years of writing and editing before moving into SEO had left me with a muscle memory I couldn’t turn off.
The data certainly didn’t validate my frustration, because the content ranked and organic visibility improved. There was no obvious crisis forcing anyone to question the process, including me, except that I was spending more time on those copies than SEO optimisation and the final quality check should consume. Because a draft could be completely serviceable from an SEO point of view, perfectly pitched for audience targeting and still leave me thinking: What the hell happened here? I was editing claims wandering beyond their sources, weak logic, and paragraphs that were competent, on-brief, and completely empty. Except (officially) there was no editor in this workflow. That part had supposedly been absorbed into “freelancer with AI” (aka modern, affordable copywriter).
But there was another irritation I was beginning to notice outside my agency bubble, basically everywhere. I would open a Linkedin post, a company article, an essay on a subject I genuinely wanted to read, and there it was again: the same voice, the same compulsive line breaks, the same neat little oppositions, the same conclusion explaining itself half a paragraph too long. Sometimes the idea underneath was actually original and thought-provoking, but I’d stopped taking it in because I felt tired seeing every thought poured into the same shitty mould. Different lyrics, same melody.
The obvious tells still get to me. Another sentence explaining that this isn’t just about what everyone thought it was about but actually about something supposedly more surprising, another “here’s the thing that actually matters” and I want to vomit on the screen. I’ve grown so intolerant that even when one earns its place I reach to pull it out anyway, like an alcoholic who can’t be trusted with a single glass at a business dinner. But that’s the easy part, and the part that matters least. You can strip out every obvious AI tell and still be left with something that says very little, thinks in familiar patterns, and could have been written by almost anyone. And the maddening part is that without an old-scool-trained copywriter or a hidden editor even the easy layer of sludge survives the final QA and optimisation pass, because spotting it was never an SEO skill.
The hidden editor
When language loses every trace that a particular person, with a particular way of seeing the subject, was ever there, I had to put that evidence back. I would hold a piece in my head until I could feel where it gave way. Every paragraph was reasonable on its own, and still there was nothing the sentences stood on, just another 1,800 words the approved brand voice had cooked up with no nutrition value.
Would a weaker article rank just as well, as long as it ticked every box in the SEO and AI-search optimisation playbook? Quite possibly. It might even have converted.
If a perfectly competent, utterly forgettable article can bring in traffic, get cited in AI search, and ultimately contribute to the bottom line, there’s very little incentive to question what’s missing during a quarterly business review. I usually make most of the weakness disappear before publication anyway, but I cannot honestly claim the page would have failed without that work. There was no line in the budget for “someone made sure this wasn’t nonsense” or an analytics event firing when “reader encountered an actual thought” (how nice would that be though!)
Editing cheap AI-assisted drafts was a nightmare sometimes, but as a result, the final article looked good. An AI-assisted draft (based on an expert, well-researched SEO brief) went in, a publishable piece came out. Unicorns and sunshine all around.
Suppose I disappeared from that workflow and was replaced by an excellent SEO or AI-search strategist — someone better than me at technical SEO, search intent, information architecture and all the increasingly specialised work the job requires, but without years of writing, reading or editing experience.
If the writer hired through a freelance platform is mostly prompting AI or lacks the skills to properly scrutinise what it produces, and the SEO strategist is reviewing for search, there can be a whole missing layer in between. Knowing that a distinction is meaningless even though it sounds sophisticated. Understanding which source deserves trust. Knowing when the brief itself is wrong. Hearing when a brand has disappeared from its own language.
I am not saying that every piece of marketing content needs to change a life. Someone searching for the difference between two business phone systems may simply want the difference between two business phone systems. There is no moral requirement to make them contemplate mortality or aesthetics on the way to the pricing page. And increasingly, they may not need an article for that at all. An AI assistant can give them the comparison directly, as long as the information and positioning it retrieves from product pages, documentation, reviews and other sources is accurate.
I am thinking more about the content that claims to have something to say, the ones we label with fancy marketing terms like storytelling and thought leadership, content supposed to explain a shift, share expertise or help someone see a problem differently.
There is a large space between “not literature” and “nothing much was added here”, the space where I think content strategy becomes more interesting now. What do we think content strategy is about, once competent, derivative content becomes almost free to produce?
Judgement lost in unnatural sameness
“The platform has turned into a pile of shit, and we’re at the bottom of it.” — Cory Doctorow
I remember the older version of the same problem, when SEO was much cruder. You would find a good product or service and then read its description mangled by keyword stuffing, and see sentences turning mid-thought to wedge in “affordable dog grooming near me”. Many blogs existed only because somebody had found a keyword opportunity, not because anybody had something worth saying about it.
Then, for a while, I thought we were getting somewhere better. Search engines got more sophisticated and marketers talked more seriously about usefulness, search intent, audience needs, ICP resonance, expertise, trust. Google even named it: E-E-A-T, for experience, expertise, authoritativeness and trust. Although E-E-A-T can be faked, like the churchgoer who never misses Mass and still is unkind to their neighbours, and that period was not perfect, but still, it felt like progress. At least the incentives were moving towards better communication. It was also surprisingly short.
LLMs arrived just as content marketing had begun to get better at asking, “What does the reader actually need?” and gave us an irresistible new question: “How much more of this can we produce now?”
The old SEO copy damaged language because optimisation overruled natural expression. Today, optimisation, audience targeting, brand messaging and fluent AI generation can all work together perfectly and still produce an endless supply of competent sameness. We can answer every long-tail query and create a LinkedIn post for every mildly relevant industry development within hours. And so can all our competitors.
A 2026 study in Nature Human Behaviour analysed more than 880,000 texts and found that when an LLM polishes writing, the meaning survives but the variety decreases. Writing-complexity variance dropped by 21 to 50% and the fingerprints go with it. After the rewrite, classifiers could no longer reliably tell the author’s age, gender or moral leanings from the words.
But competent sameness performs, so there is every reason to make more of it, faster and cheaper — until the readers left on the internet have lost all their remaining curiosity.
Cory Doctorow calls the platform version of this “enshittification”: the slow rot that sets in once a system optimises for its own metrics instead of the people using it. First they’re good to users, then they squeeze users to please business customers, then they squeeze everyone to please themselves. Written language is on the same curve one level down. The web got good at rewarding useful content, so we made more of it; then we made more than anyone could possibly need; and now the reward itself — visibility, the click, the citation in an AI answer has come loose from whether anything that worth sharing was said at all.
When I thought I was arguing for better copy, what I was really trying to protect was judegment. Nothing measures human editorial judgement in the content process machine, therefore nothing protects it.
We were flattening language before AI arrived
It is easy to blame the large language models for this beige monoculture while we spent years building the factory settings that made generative AI possible across the whole way we’d learned to write on screens.
We built autocorrect, and then Smart Compose, and then predictive everything tools whose job is to nudge a sentence towards its most statistically probable next word, and we taught ourselves to take the suggestion. We sanded corporate English down to what the linguist Jean-Paul Nerrière called Globish: a stripped, idiom-free, dialect built to prevent cross-cultural miscommunication, stripping away regional humour and difficult metaphors in the name of scale and efficiency. And in my own field we went furthest of all with a celebrated method called the skyscraper technique. This is the recipe to find the page already ranking on your topic, study it, and produce a longer, more comprehensive version of the same thing. An entire craft built on the principle that the way to win was to synthesise what already existed into something bigger. We were asking people to run the algorithm by hand, years before we could download it.
Google was already trying to solve the sameness problem before the rise of OpenAI and its competitors. In 2018 the company filed a patent for something it called an information gain score, a way to measure how much new a page adds beyond what the reader has already seen elsewhere. Think about what it means that this was necessary.
AI made the words cheap. Only the words.
It would be easy to tell a story in which AI produces bad writing and serious writers heroically resist it. I don’t believe that version, there’s more nuance to it. Yes, AI produces bad writing, and — also or because of it — it helps me a lot.
Four separate studies, synthesised in MIT Sloan Management Review, found that AI assistance raises the quality of an individual’s ideas while shrinking the diversity of ideas across everyone using it. How much of that homogenisation is built into the models, and how much stems from us, specifically, our habit of reaching for the tool before we’ve done the hard work of thinking for ourselves, it’s a different question.
A field experiment published in the Journal of Applied Psychology found that ChatGPT helped creativity most in people with strong metacognitive habits, such as noticing what they didn’t know, tracking whether an approach was working, changing course when it wasn’t. AI did more for the people who knew how to argue with what it gave them.
I need to have a sparring partner, because my thoughts don’t queue. They mob. Writing by hand means racing my own disappearing thoughts. By point five, point two has usually left the building. Working memory is a small and badly managed office.
Before generative AI, I could lose days if there was enough space available for avoidance disguised as contemplation. Now I dump the mess into a model, and what comes back is usually wrong. That is useful, because I can put the whole thing out in front of me at last and see what’s nonsense and what’s dead weight. Seeing it get my own thought wrong is how I finally see the thought clearly.
But I see daily how it’s used to generate text at scale that covers the absence of any thinking behind it. The economics seem to reward it, at least for now: brief in, article out, happy dashboards, cheap and fast. But what, exactly, are we making cheaper?
Some of the work in good writing can clearly be accelerated. But deciding what is worth saying, what is true when a polished argument is somehow empty, still takes judgement somewhere in the process.
AI writes almost well. That sounds like good news and it isn’t. The cheaper workflow makes the judgement look already done when it has only been simulated, and we start confusing the cost of producing language with the cost of producing something worth publishing.
Bad writing has become harder to recognise
Until recently, turning an idea into 2,000 coherent words took real effort, though it never guaranteed quality. Humans produced mountains of boring, derivative prose long before anyone trained a transformer. But writing that hid its own emptiness took actual skill. You needed craft to dress a hollow thing convincingly, to keep anyone from noticing the king was nude. (I’m looking at you, bestseller self-help.)
If you ever spent hours on real writing, you know that it is where most of the thinking happens. Writing exposes the difference between having a thought and having the posture of one. Writing lets you find out if that thought earns its place outside your head. An LLM, on the other hand, can just keep typing long after a human would have stopped and realised:
Actually, I don’t know.
I can’t help it but the first analogy I reach for is playing the piano. AI is the sustain pedal. In the right place it isn’t cheating at all, Chopin pieces live on it. But the same pedal also covers for the technique you never built, the passage beyond your hands, the musical idea you don’t have. And then there is Bach, which barely allows pedalling. The voices have to stay separate and hold themselves up and there is nothing to hide behind. Either someone is actually there in the playing or the notes are simply accurate, and everyone can hear which.
Research got faster, drafting got faster, but having a point, a real thought, never did. That’s what used to hold the volume back, not just the labour of producing words. AI can give a novice access to the surface patterns of expertise before they have developed the expertise required to evaluate those patterns. The output is convincing enough that no one can see there’s nothing behind it. And after a while, neither can you.
I think the value of skilled, human judgement appears mostly as absence. You can’t see good editing, you can only see its opposite.
Experienced writers have always been paid only partly for the words they produce. I think we are also paying them for everything they know not to submit.
Why does fluent-but-empty writing feel so exhausting?
“What is smooth does not injure. Nor does it offer any resistance. It is looking for a Like “ — Byung-Chul Han
The strange thing about editing weak AI copy is how tired it makes me, given how little resistance it puts up. Sentence one is fine, the next is fine, the paragraph is grammatical and politely organised. Then I look up staring into the emptiness of existence.
When you are reading a long text, you are not decoding one correct sentence after another; you are building a model of the whole. This is well established in the research on discourse comprehension, and psychologists call it a situation model.
The brain responds to meaning at the level of the whole discourse, not just the sentence: the neural signal tied to semantic processing (the N400) is smaller when a word fits the larger thing being said and larger when it doesn’t, and it tracks fit with the whole discourse, not merely the neighbouring words. And when the model stops cohering, readers physically go back. Eye-tracking studies treat those backward movements, regressions, as a sign of the effort to repair an interpretation. I also catch myself doing it. I reread the heading or check the paragraph above to ask whether I missed the point. That is what wears me out. The surface keeps telling me things are fine while the model I am trying to build never grows. The reading is effortless and empty at once.
A sadder version is when there is meaning or an original thought to share, but the author feeds it to the machine and it comes back in costume (the same darn costume every time).
Last week, I stopped reading a post halfway through. It had a real thought in it.
Here’s the thing.
We don’t stop because it’s bad. We stop because we recognize the template: the single-line hook, the tidy pivot, the ending reaching for an unearned feeling.
And it got me thinking. 🤔
The mind learns the pattern. Then the mind leaves.
And that’s where AI comes in. 🚀
Because now you can produce thoughtful, authentic, human writing at scale — the kind that actually stands out.
So today, I’m challenging you to do one small thing.
Notice the exact line where you check out. 👀
You might just find the thought was doing fine before all this happened to it. (Like now.)
Agree? 👇
♻️ Repost this if it resonated. (You won’t remember it did.)
I can build one of these in my sleep now, which is how I know I’m in trouble. Once the shape is this familiar, the mind stops turning up for it, the way you stop hearing the fridge until the day it dies.
Byung-Chul Han argues that the defining quality of our culture is smoothness: surfaces polished until nothing catches, nothing resists, nothing wounds. The smooth is made to please and therefore cannot hold meaning. AI prose is smooth, it is sanded down until there’s nothing to catch on.
But who cares? The format works, the sameness ranks, gets shared and hits the KPIs, so why not make more of it?
The sentence nobody would have written
Rushdie can load history, politics, a joke and an outright absurdity into one sentence, like an overpacked suitcase that somehow still zips. Krasznahorkai writes sentences that run a full page — exhausting or hypnotic depending on the day, but unmistakably his, and the exact opposite of the tidy little rhythm every post now shares. David Ogilvy, who marketed everything from soap to luxury cars, wrote ad copy you could pick out of a lineup, convinced you’d be a fool to disagree.
I am not suggesting that a SaaS product page should read like “The Waste Land”. What interests me is that these writers have almost nothing in common except a willingness to disobey the obvious next sentence. Their language is shaped by someone deciding, over and over, that this is how this particular thing needs to be said.
Machines can surprise you too — ask for something strange and strange is what you get — but a model can’t decide that this particular strangeness belongs here. Particularity comes from someone choosing it because it belongs to this thought, in this piece, for this reason. A thousand of these choices one person is making over and over within a whole piece is what we call a voice.
Which brings me to brands. Brands make this more complicated because there usually isn’t one person making those thousand choices. We write a voice guide instead: “confident but friendly, expert but approachable”. Even if those adjective pairs somehow fit your company and no four thousand others, they still don’t make the choices for you.
The voice happens in the choices themselves: which perfectly good sentence gets rejected because anyone could have written it, which odd word survives the edit, where somebody resists making the thought smoother, which joke stays, which sentence gets to sound a little unlike the sentence a language model would have put there.
Sometimes those small choices are the only part of the piece that makes me see something familiar differently.
Perhaps a strong brand voice is simply what accumulates after enough of those decisions have been made in the same direction. And that brings me back to the missing editor: AI can give you endless acceptable ways to say something. Somebody still has to decide which one belongs here.
I can’t quite dismiss the part where someone stops and says “not quite it.”
I want to clarify that this article is not a long complaint about marketing content being transactional or about wanting to produce it efficiently. Marketing content, even the pieces that dress themselves up as ideas, exists to sell, to be found before a competitor is, to move someone one step closer to buying. It’s what we’re paid for as SEOs and content strategists, and I’ve built playbooks and run the AI-search optimisations that make it perform. There’s no shame in it.
But I keep thinking about the business networking events from my years at a small biotech, because by any efficiency measure, they were ridiculous. Flights, hotels, two full days of conversations with people who were all there to “explore collaboration” (aka to sell something).
And still those events worked. The transaction was happening through all the supposedly inefficient stuff around the pitch: the question that sent someone off script, watching them think before answering, realising they understood the awkward part of your problem, remembering the conversation months later when the need became real.
And then there was all the stuff that had apparently nothing to do with the sale like some stupid lab story over coffee. It made the person, and with them the company, stick in your head, much like brand awareness is supposed to, except through a human encounter rather than repeated exposure.
Commercial content cannot fully reproduce the nuance of a real room or a face-to-face conversation, but the best of it tries to do the exact same job. Ranking gets you into the room. What happens next doesn’t depend solely on conversion rate optimisation best practices. We can measure recall, consideration, and brand lift, but we cannot isolate how much of that success came from someone simply enjoying the read , from feeling there was a distinct, reasoning mind on the other side of the page. Perhaps our remaining job is to recreate that “inefficient” human texture on the page.
And there is another side to it that has nothing to do with analytics: can you as an author or brand stand behind what you publish without cringing at how average and forgettable it makes you sound?
What’s left that’s worth doing
Some marketing content really can — and should — become almost free. Let AI write the comparison table, summarise the spec sheet, and answer the straightforward query. But at the boundary where there is no obvious next sentence, the cheaper and smoother the workflow gets, the less room remains for someone to pause, say “this isn’t quite right,” and do the work of figuring out why. We have become remarkably good at optimising away that pause.
So what’s left for us to do? Maybe less doing.
Find the thought nobody has had before? If you have one, fantastic. But that’s an absurd altitude to demand from every piece of writing. Most of us are working much closer to the ground. You can go for the angle the consensus hasn’t seen yet, or the detail that only defines itself after you’ve done the work, the connection that makes sense because of where you’ve stood. It can even be just breaking the smooth sentence and putting the unlikely word where you believe it belongs.
I could end this with five packaged principles for content strategy in 2027 or a new operating model for the AI era. I could prompt a model for that framework right now and have a polished, highly authoritative version before my coffee gets cold.
But inventing a framework won’t change that nobody actually knows how much human attention will belong to generated summaries versus original pages. What we do know is that publishing more competent, frictionless content is a race to a place where everyone sounds identical.
There is one part of this, though, that I care about beyond content marketing.
Language is not merely a delivery system for information. Much of the pleasure of reading comes from encountering another mind and a tiny bit of art alongside the utility. An ad can have it. A headline can have it. A sentence can answer a technical question and still carry wit, an unexpected turn of phrase, or a deliberate choice that didn’t have to be made that way, but was.
If we allow every operational pressure to push language towards its safest, smoothest, most reusable version, we won’t just end up with boring content where every sentence works and none are a pleasure to read. We will lose one of the last places where people make thoughtful things for one another — even in marketing, even in a routine guide. But if we decide that none of that matters because it cannot be measured, scaled or reliably attributed to a conversion, I am not sure what exactly we are preserving when we say we still want the content to feel human.
This article was originally published on the author’s Medium page.
