Please Stop the Slop: What AI Slop Is and How to Spot It

Please stop the slop. AI slop is cheap stuff a machine churns out in bulk with no human value in it: songs, videos, books, posts, and the comments under them. Last weekend it showed up under my own post. So this is my verdict on it: what it is, how to spot it, and why being human still wins. It’s also a petition you can sign at the end.

“Does this in any way invalidate my question?”

On 3 October I posted on LinkedIn about Kolibri, the new open German model from Aleph Alpha. The top comment opened with “Great write-up, especially the tokenizer test”, gave me a number that sounded specific, said “One thing I’m missing” and closed on a neat question. Another one opened with “The 44% vs 11% ‘I don’t know’ rate is the number I’d test first”, which read like the same assistant had written both on the same afternoon.

So I asked the first one “answer honestly did you generate this reply with ai” and got another question back: “does this in any way invalidate my question or argument?”

It does. So this is what I wrote back:

So you did eh? It does invalidate your question or argument because I’m not talking to a human at this point but some machine. I prefer to keep my communication with machines to machines, not machines in disguise.

A third person told me I was “playing AI slop card to avoid responding to legitimate questions”. My answer was “there is no ai slop card to play. I speak to humans.”

Then I asked on X: “vibe check am i being an asshole” and most of the replies said no. Rizèl Scarlett has the same problem I do:

Phil Zona went straight to the point: “if it was a question asked out of true curiosity, they wouldn’t have needed AI to write those 3 sentences. they’re not actually interested in the answer.”

Slop is most of the internet now

I wish this were just my comment section. Pangram, a company that detects AI writing, ran over a million posts through its checker this summer and The Register reported what it found. 41% of LinkedIn posts over 250 words were written entirely by AI. On X it was 25%, with another 23.2% written with AI’s help.

LinkedIn, posts over 250 words

Entirely AI

41%

With AI's help

4.3%

People

55.2%

X, all posts

Entirely AI

25%

With AI's help

23.2%

People

52.7%

Bots now make more of the web’s traffic than people do. Imperva, which has measured bot traffic since 2013, found in its 2025 Bad Bot Report that automated traffic passed human traffic in 2024 for the first time in a decade, at 51%. In its 2026 report it was 53% (the earlier years below are from its 2022, 2023 and 2024 reports).

40%50%60%20212022202320242025Bots 53%People 47%

Year	Bots	People
2021	42.3%	57.7%
2022	47.4%	52.6%
2023	49.6%	50.4%
2024	51%	49%
2025	53%	47%

LinkedIn’s chief product officer told Fortune that LinkedIn blocks “hundreds of thousands of automated slop comment attempts every day”. Every day! Whatever they’re doing, I guess they need to do it better because the slop keeps coming. In July LinkedIn added a “Seems like AI slop” button to every post and comment. Merriam-Webster even made “slop” its word of the year for 2025.

What is AI slop?

AI slop is cheap content a machine manufactures in bulk, with no human care or value in it. Merriam-Webster’s definition is “digital content of low quality that is produced usually in quantity by means of artificial intelligence.” Comments are the slop I see most, but it’s in everything:

• Music. Deezer gets nearly 90,000 fully AI-generated tracks a day, which was more than half of all new music uploaded at its peak in June 2026.
• Video. Kapwing went through the first 500 videos a brand new account gets shown and found that 59% were AI slop on TikTok and 21% on YouTube Shorts.
• Movies. YouTube shut down Screen Culture and KH Studio, two channels that spliced official footage with AI images into fake trailers for big franchise films, after they’d pulled in over a billion views.

Hardly anyone wants it though. On Deezer, all that AI music was 1 to 3% of what people actually streamed:

Of new uploads

over 50%

Of what people stream

1 to 3%

I’m an AI engineer so I’m not against machines making things. My problem is what’s missing: nobody cared about the thing so there’s no human value in it. In my comments it comes with a disguise on top, a machine talking to me through a person who acts like it’s them. When I reply to that comment, I’m spending my time on a conversation with nobody in it.

It makes me cynical (and I won’t let it)

The worst part of slop for me is what it does to me: it makes me negative and cynical. Every polished comment now gets a little suspicion from me before it gets any warmth. I hate that.

Slop doesn’t get to change me though. My humanity needs to stay constant whatever the bots do.

I wrote in What Does the Bible Say About AI? that every human is made in the image of God. Nothing we build is. So the bot doesn’t get my warmth, but the person who sent it still does, like every human who writes to me in their own words, typos and all.

How to spot AI slop

You don’t need a detector for most of it, you just need to know what to look for.

The phrases. Claudisms is a public banlist of over 120 phrases and sentence shapes that Claude leans on: “sit with”, “load-bearing”, “the tell”, “Here’s where it gets interesting”, “the one that surprised me most”. Wikipedia’s editors keep their own field guide, Signs of AI writing, which lists the rule of three, “It’s not X, it’s Y” and words like “delve”, “tapestry” and “testament”. Graphite measured how much more often Claude Opus 5.5 uses some of these than people do, against human writing from before ChatGPT:

"this matters"

116×

"is more than a _ it"

98×

"why _ matters"

92×

"looking ahead the"

40×

"rather than simply"

32×

116 times! People do say “this matters” (I’ve said it on stage) so one of these proves nothing on its own.

The pattern. On LinkedIn, slop comes in a stack: a compliment about one specific detail (“especially the tokenizer test”), a number that sounds researched, “One thing I’m missing”, and a neat question to finish. Or it picks one detail and crowns it: “is the number I’d test first”, “the part that actually matters”.

The swap test. Imagine the same comment under somebody else’s post on the same topic. Slop reads just as well there.

The shapes. tropes.fyi keeps a catalog of the tropes the newest models fall into. A few of them are all over slop replies. The reply opens by handing you your own question back (ask which database to use for a side project and it starts “For a side project like yours,”). It tells you an answer exists before it gives you one (“has one big thing going for it:”). It staples an afterthought onto a sentence that was already done (“, though.” or “, beyond the basics”). And its last line points back at the one before it and crowns it (“That last part is where the two differ.”).

The vetter. tropes.fyi also has an AI Vetter: paste a link and it scores the page from “Human” to “Pure Slop”, trope by trope. I wanted to know how it decides so I fed it test pages until I could rebuild its scoring. The tropes multiply: one on a page scores about 7, 2 of them score 23 and 4 score 71. So a short comment with 2 small tells already reads as “AI-assisted”. It isn’t proof though! My post on why I don’t like speaking at remote conferences is from 2020, years before ChatGPT. It scores 42 (“Suspicious”), mostly for starting 4 sentences in a row with “I want”. That’s just me!!

Why human writing wins

Slop costs trust. Oliver Schilke and Martin Reimann ran 13 experiments and found that people who disclosed their AI use were trusted less than people who didn’t (and being exposed by someone else cost even more). So the commenter who gets caught loses twice: once in the trust a human reply would have kept and again when somebody points it out.

The machines need us. Shumailov and colleagues showed in Nature that models trained on their own output lose the rare things first and drift into nonsense within a few generations. Their conclusion is that “access to the original data source must be preserved and further data not generated by LLMs must remain available.” Every human sentence on the internet is part of that original source. So slop eats the very thing the models learned from.

Only you were there. A model can write a perfect paragraph about a tokenizer, but it wasn’t there the afternoon you ran one and can’t tell anyone what surprised you or why you cared. Good taste stays human for the same reason: it’s judgment you build from your own experience so nobody can generate yours for you.

Is all AI slop?

No. AI output becomes slop when nobody cared enough to make the choices or check what came back. With somebody doing both, the same tool helps make craft. Most of my talks and the design of this website are made with AI’s help these days. I use it to research, cross reference and draw connections exactly the way I would without it, just faster.

I said this on stage last month, opening The Critical Advantage With AI at Agent Conf in Warsaw: “everybody now has a hammer” but “we’re not all building awesome houses”. And “we all have cameras in our phones, but you still hire a wedding photographer, or at least you maybe should, right?” I told the room there’s a very clear line between slop and craft. The rest of the talk was about taste and how the research says you build it.

Simon Willison, whose post from May 2024 helped the word catch on, put the line in the same place: “Not all promotional content is spam, and not all AI-generated content is slop. But if it’s mindlessly generated and thrust upon someone who didn’t ask for it, slop is the perfect term for it.” He uses AI for all sorts of things and still wrote “I attach my name and stake my credibility on the things that I publish.” By the end of that year he was describing slop as AI content that is “both unrequested and unreviewed”.

AI slop vs. craft

Richard Sennett defines craftsmanship in his book The Craftsman as “an enduring, basic human impulse, the desire to do a job well for its own sake.” He even counts Linux as craft: “technological craftsmanship, the intimate, fluid join between problem solving and problem finding.” That’s literally what building software feels like on a good day!

Ted Chiang makes the case against AI art in Why A.I. Isn’t Going to Make Art. His argument is that “art is something that results from making a lot of choices.” A 100 word prompt makes about 100 of them. So a model writing a 10,000 word story from it “has to fill in for all of the choices that you are not making” (and an average of what everyone else wrote is “the least interesting choices possible”). I agree with him!! He leaves one door open, for a program nobody has built yet: one with “extremely fine-grained control” where “a person could use such a program and still deserve to be called an artist.” His one example is the film director Bennett Miller, who generated “more than a hundred thousand images to arrive at the twenty images in the exhibit.”

AI-assisted vs. AI-generated

Amazon spells this line out for every book on Kindle Direct Publishing (KDP) in its content guidelines. If an AI tool created the actual text or images, the book counts as “AI-generated,” “even if you applied substantial edits afterwards.” If you “created the content yourself, and used AI-based tools to edit, refine, error-check, or otherwise improve that content”, or used one “to brainstorm and generate ideas, but ultimately created the text or images yourself”, it’s “AI-assisted”. Authors have to tell Amazon about AI-generated content. They don’t have to say anything about AI-assisted content.

Readers draw the same line. For its Generative AI and News Report 2025, the Reuters Institute asked people in 6 countries (Argentina, Denmark, France, Japan, the UK and the US) how comfortable they’d be with news made in different ways:

Entirely by AI

12%

Mostly by AI, with a human in the loop

21%

Led by a human, with some AI help

43%

Entirely by a human

62%

A human checking the machine’s work only raises comfort from 12% to 21%. A human leading with some AI help doubles that to 43%!! People want a person making the choices.

When AI makes your work worse

In Fabrizio Dell’Acqua and colleagues’ experiment with 758 consultants at Boston Consulting Group (BCG), the ones using AI on 18 tasks inside what AI could do finished 12.2% more of them, 25.1% faster and with “significantly improved quality”. On a task picked to sit just outside it, they were “19% less likely to produce correct solutions” than the consultants without AI. I wrote more about this jagged frontier in my post on super intelligence.

It’s easy to fool yourself about it too. In a 2025 study by Model Evaluation and Threat Research (METR), experienced open source developers took 19% longer with AI tools and still believed afterwards that AI had sped them up by 20%. Wild! (METR’s 2026 follow-up estimates a speedup now, but so many developers won’t work without AI any more that METR calls its own new data “an unreliable signal”.)

Even when AI helps, it pulls everyone toward the same place. Anil Doshi and Oliver Hauser gave some of 293 writers story ideas from an AI model. And those stories were rated more creative and came out more like each other. In their words, “writers are individually better off, but collectively a narrower scope of novel content is produced.” That’s the 2 LinkedIn comments from the top of this post!!

At work it has its own name now. Researchers from BetterUp Labs and the Stanford Social Media Lab call it workslop in Harvard Business Review: “AI generated work content that masquerades as good work, but lacks the substance to meaningfully advance a given task.” Of 1,150 full-time employees in the US, 40% had received some in the last month. Each piece took them an average of 1 hour and 56 minutes to deal with. And about half of them saw the colleague who sent it as “less creative, capable, and reliable” than before.

How I try to avoid AI slop

Every talk I give starts with me deciding what it’s about. For The New UX at CityJS London in April, I told Claude what I wanted to highlight (streaming interfaces instead of making people wait for a whole answer, building and using Model Context Protocol (MCP) servers, and ChatGPT apps with their own interface) and asked it for an outline and a sample repo. The next day I was arguing with it about the demo’s API: “look man the api needs to be as lean as possible”.

Then there’s research. Even Google’s guidance on AI content says “Generative AI can be particularly useful when researching a topic”. For my post on how long a keynote should be I wanted to know what science sits behind the 18 minute TED talk. Following the citations back with AI got me to a paper from 1978 about note taking!! It checks me too. On stage I credit my whole definition of taste to a 1993 paper. When I checked it with AI for my post on good taste, the paper’s abstract only backed part of it. So the post says so.

This website went the same way. When I asked for a brand, I wrote down my life and then: “Red has to be a part of it for blood (mine, Christ’s, etc.). Maybe Schwarz/Rot/Gold because I love Germany too?” The AI found free typefaces close to the ones I named, measured the contrast of every color and wrote the code. It took my Schwarz/Rot/Gold and made the red line on /story go missing at age 4 and come back gold. In my own notes, that gold is “because somewhere on the journey, I met God.”

I throw a lot of what it makes away. “its so fugly revert it” is a whole message I sent about a redesign in July. The first version of the woven thread on /story got “the tapestry is bullshit… its just dots and lines. use vgpu”. The shader that’s on that page now came out of that message.

Sign the petition to restore humanity

Signing this petition costs you nothing but a little effort:

1. When you reply to a person, write it yourself (typos and all).
2. If a machine wrote it, say so.
3. When you see slop on LinkedIn, use the “Seems like AI slop” button in the “…” menu. It hides the post for you and helps LinkedIn learn what slop looks like.
4. Reward the humans by replying to their posts, reposting their work and telling them when something they wrote landed.

To sign, share this post with one sentence you wrote yourself.