There are many “how to spot AI writing” and “stop letting AI do this” lists doing the rounds at the moment. Most will tell you that em dashes are no-go. Along with tidy triads, rhetorical questions and contrastive aphorisms.
Some of these lists are really useful: they shine a light on our own written tics, as well as those of AI agents. And the best of them do a good job of cleansing, of preventing content from sounding like the worst kind of AI slop.
But there are risks here too, partly because they tend to over-simplify, partly because people are using them to bundle together three different jobs, and they only really do the easiest one.
Having wiped the machine’s crumbs off the floor and its fingerprints off the surfaces you may find there’s little left underneath, bar clean, competent, anonymous prose.
Three jobs, not one
The three jobs are these: not sounding machine-made; creating great quality content; and sounding unmistakably like yourself, or your organisation. They are not the same task, and they get harder, and more valuable, in that order.
1. Hygiene
Cleaning your content to sound less like AI is almost entirely a matter of removal: take out the em dashes, the “in today’s landscape”, the “it’s not just X, it’s Y”. A checklist handles it nicely, which is precisely why the lists stop there. But stripping out every tell may only get you out of the uncanny valley. If that’s all that happens, you may land somewhere worse: a place where content is competent, inoffensive but still faceless. Less-like-AI doesn’t necessarily get you that far down the road towards better.
2. Craft
Creating good content is craft, and craft is mostly judgement. Clarity, structure, rhythm, cutting the sentence that merely repeats the one before it, actually having something to say. Some of content craft can be taught as deterministic rules; a lot of it goes beyond find-and-replace and pattern-matching.
This is what the AI-detection lists tend to sweep under the carpet: content can have em/en dashes and be brilliant, or not have them and be gobbledegook. Content (whoever or whatever wrote it) can sound profound, be free of AI-tells and mean very little to anyone.
3. Identity
Sounding like yourself, or your organisation, is the third job. It’s your framing, your voice, the turns of phrase nobody else would use. Even, arguably, your content tics. It’s brand implementation. It’s sounding not-like-AI but also not-like-anyone-else. It’s your content fingerprint.
Identity is arguably the most valuable of the three exactly because it can’t easily be reduced to a universal set of rules, which is also why it gets dropped from the listicles: it’s not tidy enough to make a satisfying post. You can’t subtract your way to a voice.
Beyond forensic graphology
People are understandably concerned about a world where human creativity is subsumed into the robotic. Where the whole starting point of a piece is that one particular human sat down and wrote it – a personal essay, a testimonial, a founder’s note – then passing off fully automated, machine-written text as that person’s own labour is misleading and inauthentic.
But most web content is not that. Most organisational content was never the work of one human alone at a keyboard, and nobody imagined it was. Organisations’ content systems are already a machine of sorts, making many different people sound like a single consistent one. Briefs, templates, style guides, the sign-off chain, the ghostwriter who has always existed: organisational content has always been manufactured, and no one was deceived, because the thing standing behind it was the organisation, not a person.
And “a human wrote it” guarantees nothing in any case. Plenty of content on the web is stale, bloated, unreadable, years out of date and entirely without a pulse. We have been saying this for a long time: one of the first things we ever published at Contentious, nearly 10 years ago, was a piece called Create less crud, written well before “enshittification” had a name. Poor content is not a new disease that arrived with large language models. The web has been drowning in it for an age.
AI didn’t invent slop, or crud, or ROT (redundant, outdated, trivial). AI industrialised one particular flavour of it, and provided everyone with a convenient focus for their frustrations.
“Human or machine” is almost always the wrong question. Better ones, kept apart, are: is the provenance honestly represented, where provenance is the point? And is it any good? Is it true, current, readable and recognisably yours?
AI as a quality ally
Working with AI the way a writer works with an editor, or with a subject-matter expert, sits comfortably inside this model. The human supplies the direction, the judgement, the point of view, the thing they will put their name to; the machine drafts and tidies. That is a partnership, and an old one.
It’s a different animal entirely from “go away and write me a hundred blog posts”, which is not partnership but abdication, and abdication is how you get crap at scale.
Exactly what good looks like for you goes beyond the deterministic: it depends on your strategy, your brand, your audience, your context. With the right scaffolding and the right instructions, AI can help you check content against all these things, but they’re bespoke and not universal.
The em dash is not the enemy
Take the em dash, since it has become the pantomime villain of every list. The dash is not the problem. We rather like dashes (though on this side of the Atlantic we use en dashes not the em-sized ones). The problem is the ubiquity and the misuse: the page so thick with them that it reads as machine-made even when every mark is a correct one. But the lists tend to over-correct into “never use a dash”, which is just as bad because it signs you up to a different kind of monotone.
The tools that catch this sort of issue, Slopless and its kind, are deliberately per-pattern: they flag what’s bad in roughly any register, and they stop there. Whether a particular dash is a genuine enhancement or a lazy AI tell is a per-brand judgement; it depends on whose writing it is. You can’t add up a stack of universal bans and arrive at a voice. The hygiene layer is mechanical and shared; the identity layer is taste, and it’s yours alone.
The machine can help you sound human
The backlash against AI misses the best part: once you stop asking “is this AI?” and start asking “is this any good, and does it sound like my organisation?”, the tool stops being the enemy. The same models that flood the web with faceless prose can, pointed the other way, help you rise above the flood.
Used well, AI earns its place in the toolbox. It can read every page you have and flag the stale, the bloated and the years-out-of-date, show you where the writing has drifted from your voice, and help a small team keep thousands of pages healthy in a way that was always an uphill struggle before. Monitoring and lifting content quality at that scale is precisely what we built Content Health Check to do.
AI works best in partnership where people supply the judgement, the taste and the point of view. Left to mark its own work, AI can fall in love with its own cadences. Done right, it can help shape your direction, spark ideas and push you somewhere new. It’s more than “just another tool”. But it’s best used as a collaborator, not a replacement.
AI didn’t invent robotic content; organisations were struggling to sound less corporate before Claude was ever conceived. And, oddly enough, the machine can help you sound less like one. It can help you cut the corporate throat-clearing, find the plainer word, say the thing a real person would say. Pointed at the right target, AI does not make you more robotic; it can help you sound more human, and connect with the people you are writing for.
Clean floors and delicious food
None of this makes the detection lists worthless. They are useful the way clean floors, and clean hands, are useful in a kitchen: necessary, but nowhere near sufficient for great cooking.
Wiping the machine’s fingerprints off your writing is worth doing. But for us, the interesting work is in the layers above it, the one no simple don’t-do-this checklist can reach: craft and identity. Your words, your judgement, the things only you would say and mean, put together to brilliant effect.
That combination was always the prize, and now that AI makes the basics cheap and fast, rising above the baseline matters more than ever.
The obsession with “not sounding like AI” will be transient. The system will get better at fixing itself, and as the web fills with prose that’s competent, tell-free and anonymous, the craft and identity layers will become the scarce thing. With a little effort, anyone can now use AI to produce writing with no AI fingerprints on it. The rarer and more valuable trick is writing that is unmistakably good, and unmistakably yours.
