Everyone needs to know when something is written by AI in case we accidentally publicly “like” something we enjoyed that is later to be revealed as invoking AI in some remote way.

When Anthropic announced it was watermarking AI content, my first thought was the same one as when I read about publishers and unreliable AI detectors: why do they even need to? Aren’t these guys on LinkedIn?

It seems to me that every other post on LinkedIn nowadays is by an AI detection expert announcing, with the swagger of a customs officer who’s just found contraband, that they can spot AI writing from thirty paces. They never need to check. They just know.

Anything they don’t like: AI slop. Anything they disagree with: AI slop. Anything written by someone they don’t like and want to disagree with: AI slop. It is their sixth, seventh and eighth senses rolled into one, sharpened and primed and able to tell an AI slop post by the very first letter of the post. Some don’t even need that!

I searched online for any MA courses, or at least a measly BSc, in identifying AI slop by the space between two words, but it seems the University of AI Slop Detection isn’t advertising.

So seeing as I have sweet f.a. else to do once the lights go out of an evening in our continuing Gambian power crisis, I set about creating the definitive guide to AI slop detection. Because obviously everyone needs to know when something is written by AI in case we accidentally publicly “like” something we enjoyed that is later to be revealed as invoking AI in some remote way.

And so, TNPS proudly presents The Complete Idiots Guide To Spotting AI Content, seeing as the experts regard anyone taken in by a stray em-dash or the word ‘genuinely’ used in the right context as a complete idiot.

So sit with what follows and let this genuinely land properly, since you obviously were misled by your English teachers when you were at school and were clearly taught AI slop English by Sam Altman when he time travelled back to your era.

The Complete Idiots Guide To Spotting AI Content

  1. It uses an em dash. Nobody even conceived of such a literary device before November 2022. Occasional appearances in classic literature were just ink smudges due to a quashed fly on the ink-roller.
  2. It uses the word “genuinely.” A neologism, apparently, dating to roughly the release of GPT-4, and unknown to English speakers for the several preceding centuries in which they used it constantly but thought they were writing gentlemenly.
  3. It varies its sentence length. Unheard of. Every sentence, from Chaucer onward, was required by law to be exactly the same size. Don’t take my word for it! Dig out that dusty copy of The Canterbury Tales you last read in High School amd take a look.
  4. It’s spelled correctly and punctuated properly. A historic first. Humanity built the entire written record without ever managing this, right up until a certain chatbot showed us how on that fateful day in November 2022.
  5. It starts with a capital letter. Dead giveaway. Humans have always started sentences in lowercase, like we were too cool for caps lock. Only ChatGPT respects the Queen’s grammar.
  6. It uses paragraph breaks. Suspicious. A real human would write one unbroken 400-word wall of text about “humbled to announce…” with no breathing room whatsoever.Structuring your thoughts into readable units is, as everyone knows, an entirely synthetic behaviour.
  7. It uses bullet points or numbered lists. No human in the history of LinkedIn has EVER organised thoughts clearly. We prefer to bury our insights in a meandering anecdote about a delayed flight to the Caribbean.
  8. It has a beginning, a middle, and an end that relate to one another. Deeply suspicious. Real writing simply stops when the author remembers they left something on the hob.
  9. It uses the word delve. Because before AI, no one had ever delved into anything. We just vaguely looked at stuff.
  10. Using letters all the way from A to Z. A dead give way. There are 26 letters, and humans only have ten fingers. Even is we use our toes we can’t hit all those keys!
  11. Proper verb agreement. A lack of “your/you’re” confusion? This is unforgivable. Humans have been communicating in grunts and interpretive dance since the dawn of time. If the grammar is too good, it’s AI.
  12. Using an adjective. Artificial.
  13. Using two adjectives. Definitely artificial.
  14. Using an adjective followed by a noun. The Turing test has been failed.
  15. Quotation marks. Definitive proof that the author has access to artificial intelligence rather than, say, a keyboard.
  16. It uses the word “however.” The H-word. As if humans never held two competing thoughts in their minds simultaneously before Sam Altman bestowed that lexical gift upon us. Dickens wrote “It was the best of times, it was the worst of times” and then stared at the page for six hours because he didn’t have the transition vocabulary to reconcile them. Unheard of until 2022.
  17. Using Synonyms to Avoid Repetition. A real person would write: “The synergy was synergistic, which really synergized our core synergies.” But if a post uses “crucial,” “vital,” and “important” in the same paragraph to avoid repeating themselves? That’s a neural network flexing its terabytes of training data.
  18. Semantic near‑duplicates across posts. If the post you are reading sounds like 3,000 other posts published this morning, clearly an LLM at work; humans never copy each other, follow trends, or recycle talking points.
  19. It groups things in threes. Fast, cheap, reliable. Blood, sweat, tears. Life, liberty, pursuit of happiness. Yep, Thomas Jefferson was clearly a large language model user.
  20. And, the ultimate tell: it’s coherent. Legible. Doesn’t trail off mid-sen—

And yeah, that’s a genuine em-dash. Rare here at TNPS – I hate the bloody things! – but unavoidable, as I’ll explain later.

But here’s the thing (and no, not an AI term – Bill Bryson had elevated “here’s the thing”to an art form back in nineteen bow and arrow), if any of the above are what you experts are what is giving the game away, it isn’t AI you’re detecting. It’s competent, edited prose – and the fact that so many people now find that suspicious says a good deal more about the average unedited LinkedIn post than it does about any large language model.

Full disclosure

A confession, at this point, is due.

None of those twenty example as written were mine. True, a few had independently occurred to me, but the truth is I put a version of this challenge to several AI models, gathered their answers, picked the sharpest lines, and presented them here as one voice so they’d read as a single piece rather than twenty different comedians talking over each other. If you laughed at any of it without being able to tell which bits were AI and which weren’t – and you couldn’t, because that was rather the point – you’ve just failed the test the detectives claim to pass at thirty paces.

That’s not a gotcha aimed at you, the reader. It’s aimed at the premise. The whole LinkedIn cottage industry of “spot the bot” rests on the idea that machine writing carries a stable, visible fingerprint. It doesn’t.

What it carries is the fingerprint of editing – clarity, structure, transitions, a point that goes somewhere – which is a feature of good writing regardless of who or what produced the first draft, and has been for rather longer than since November 2022 got crossed off the calendar.

So here’s the actual argument.

Detectors don’t work, and the failure isn’t neutral. The classifiers can’t reliably separate AI prose from human prose because the models were trained on human prose – the categories overlap by construction.

And the false positives aren’t randomly distributed. Studies on GPT-detection tools have flagged non-native English writing at dramatically higher rates than native writing, because ESL prose tends toward exactly the features the detectors treat as tells: simpler sentence-length variance, safer phrasing, fewer idioms.

A detection regime doesn’t catch AI. It catches unfamiliarity with a particular Anglophone register and mislabels it as machine origin – which means the people it’s most likely to falsely accuse are precisely the Global South publishers, educators and second-language writers this publication spends its time defending.

Watermarking only ever catches the honest. A mandatory scheme constrains compliant tools used by compliant people. It does nothing about open-weight models run locally, nothing about a single paraphrase pass, nothing about the tools that will simply never implement it because they answer to no jurisdiction that mandated it. All the enforcement weight lands on the writer who was going to disclose anyway. None of it touches the person actually trying to deceive. That is, at risk of sounding Claudey, a compliance tax on honesty, not a safety measure.

The entire framework is measuring the wrong variable. “Was this AI-assisted” is a process question. The thing that actually matters – plagiarism, a fabricated citation, a cheated exam, a fake byline – is a provenance claim: did the text lie about where it came from? That’s already governed, by academic integrity codes and editorial standards that don’t hinge on which tool was used, only on what was claimed.

A universal detector doesn’t strengthen that. It gives institutions a badge to point at instead of doing the verification work they were always responsible for.

And “slop” was never about authorship in the first place. Content farms, keyword-stuffed listicles, ghostwritten thought-leadership – none of it started in November 2022. AI didn’t invent low-value content, it just lowered the cost of knocking it out it. Sorting by origin instead of by value lets plenty of hollow human-written posts stroll past the checkpoint waving their birth certificates, while genuinely well-argued AI-assisted writing gets flagged for the heinous crime of being readable.

Publishing has absorbed disruptive authorship technologies before – the photocopier, spellcheck, translation software, ghostwriting itself – without a mandatory universal detector for any of them, because what actually protects readers has always been contextual: editorial trust, disclosed methodology, reputation earned over time. That’s a harder standard to meet than a badge.And a more honest one.

Which is really the point. The LinkedIn detection game isn’t an epistemic exercise (no, AI did not invent the word epistemic, unless Sam Altman time-travelled to ancient Greece). It’s a status performance – I have the discernment you lack – and it’s unfalsifiable by design: a correct guess confirms the detective’s instincts, a wrong one is never checked, and being wrong costs the accuser nothing. A real verification mechanism needs to be falsifiable and carry some cost for a false accusation. This one has neither.

Anthropic doesn’t need a watermark. LinkedIn doesn’t need a slop button. Publishers don’t need an AI detector. What everyone needs is a slightly lower tolerance for confident people mistaking their own certainty for evidence.

Although thinking about it, maybe a LinkedIn “Seems Like A Self-Appointed AI Slop Expert Guessing Again” button might not be such a bad idea,

This post first appeared in the TNPS LinkedIn Analysis Newsletter.