Slop is not the enemy. The real problem is polish.

The Toys \"R\" Us film had no errors, cost real money and made nobody feel anything. The dictionary calls that fine. It needs a word of its own.

Andrew Ramsden 11 September 2026

In June 2024 Toys "R" Us released a short brand film, and it was the first one anyone had made with OpenAI's video model, Sora. An agency called Native Foreign produced it. OpenAI lent its name. And the internet, more or less as one, called it soulless AI slop.

I remember agreeing. If you saw it, I suspect you did too. A video editor named Zach Silberberg put it best, and tens of thousands of people agreed with him inside a day: "There's something so cynical about creating an ad about the limitless imagination of a child and then rendering it with soulless AI slop."

And yet, put the film beside the complaint and the complaint slides off it. It cost real money. Professionals made it, and a client signed off on it. An agency and OpenAI put their names to it, and the brand's own marketing chief called the campaign a success afterwards, which is what you would say, and also probably true. What was wrong with it is that it made nobody feel anything. That is a real problem, and it is a different complaint from the one we all made.

So what does "slop" actually mean?

The dictionary, as it happens, sides with the film and against us. Macquarie made AI slop its word of the year for 2025 and defined it as "low-quality content created by generative AI which often contains errors and is not requested by the user." Three conditions, and the Toys "R" Us film fails all three. It was polished, it was accurate to its brief, and Toys "R" Us very much requested it. So the word every one of us reached for that week does not name the thing we were angry about.

Once you see this conflation you find it everywhere, and here it is happening in print, by a writer who sees more than most. Wilfred Okajevo wrote in April this year that "AI slop isn't bad writing. It's writing with nobody home." A little further down he says that polished output with zero grammar errors and zero personality is slop, and that a rough draft full of typos with a real point of view is something else. He is right about what he noticed. He has the distinction in his hands. And then he files the polished, personality-free thing under the same word as the nine-fingered hand, and the distinction is gone again.

I did the same thing. We all did. The trouble is that a word you widen stops doing the job a word is for. Too little care and too much of the wrong kind of care have opposite causes. One needs somebody to try harder. The other needs somebody to stop trying so hard to be safe. A single word covering both cannot tell anyone what to do on Monday morning, so it's about time we split it, and I'm giving the second thing a name of its own.

What do we call the careful version?

The process of taking everything specific out of a piece of work, one reasonable decision at a time, I call Blandification. What is left is perfectly fine, but belongs to nobody. Nobody involved is careless. Everybody is careful. The care is aimed at whether anyone could object, rather than at whether anyone would feel anything, and that one difference in aim produces the whole effect.

Slop announces itself. You see the extra fingers and you scroll past, and it costs you half a second. Blandification passes at a glance. It gets approved by four people. It goes out under your name and a photo of your face, and it is only three months later, scrolling back, that you notice it could have been anyone's. That is why I think it is the more dangerous of the two.

You can run the test on anything on your desk today. Was the care aimed at whether anyone could object, or at whether anyone would feel anything?

Doesn't the anti-slop industry already fix this?

It is aimed squarely at it. It just calls it by the wrong name.

There is now a whole genre of rules for stripping the machine out of a draft, and I have used them. Two of the popular open rule sets, one called stop-slop and one called no-ai-slop, ban words like delve and tapestry and leverage, ban certain sentence shapes, ban the noun, then the colon, then the dramatic reveal, and above all they ban the em dash. A count made for this article, across five of these rule sets, put roughly five rules in six on a surface pattern rather than on anything about substance.

The em dash is the clearest case. By early 2025 it had become known as the ChatGPT hyphen, and in November Sam Altman announced the fix: "If you tell ChatGPT not to use em-dashes in your custom instructions, it finally does what it's supposed to do." Design has the same story. One designer, quoted in a piece on how to spot AI-built interfaces, says "colored left borders are almost as reliable a sign of AI-generated design as em-dashes for text."

So everyone already knows what the problem is. People can feel the emptiness even when they cannot name it, and every one of them is calling it slop.

And then they fix it badly, and the reason they fix it badly is the whole point. Removing the tells takes the machine's fingerprint off the work. It puts nothing back. The draft with no em dashes and no "delve" and no coloured left border is exactly as empty as it was ten minutes ago, and now it is harder to spot.

I banned the em dash from everything I write, this piece included, or so I intended. There are two below, and they're staying. I did it because I wanted to give no-one an excuse to dismiss the work before they'd had a chance to actually read it. I also loved a good em dash. I used them long before the models learned to, and the ban is a loss I took on purpose, and I still feel it every time a sentence wants one. Did I just blandify my own content so as to manage perceptions? Yes. See how easily it happens? Not a big loss, but bland by a thousand cuts.

How did AI invent blandification?

It didn't. And the proof was on your LinkedIn feed last week—and last decade.

I was scrolling and I noticed the same illustration on post after post. Flat colours. A cupcake. Bendy-limbed figures with their arms in the air. Each one was celebrating a different person starting a different job at a different company, and each one had picked the same picture from the same menu. No model drew it. A person drew it, on commission, and a team approved it, years before the models.

The people who built that style wrote down what they were doing. The studio Buck built the original system for Facebook, "a scalable system rooted in flat, minimal, geometric shapes", and described the figures as abstracted, saying that "oversized limbs and non-representational skin colors help them instantly achieve a universal feel." Read as a brief, it makes complete sense. A company operating in almost every country on earth needs pictures that belong to none of them.

Microsoft got there in the end. In a piece on rethinking their own illustrations, their design team wrote that the flat style "now feels emotionless", and that their own audience research had found the illustrations uninteresting despite being "colorful, inclusive, and genial". Creative Bloq put the same finding in one line. Corporate Memphis, as the style came to be called, "can be moulded to represent everyone, yet speaks to no one."

I want to be fair to the illustrators, because I have sat where they sat. They did good work and they were entitled to be proud of it, and the emptiness belonged to the brief—the goal of sanding all the edges off, perfectly inoffensive, perfectly perfect, and well, perfectly bland.

What about the ugly stuff? Doesn't that count?

Real slop also exists, and it fixes itself, which is the other half of why I want the two words kept apart.

In 1987 a man named Rainer Reisenberger warned a conference about desktop publishing. "Given a dazzling array of typefaces and a multitude of tempting tools," he said, "it is only too easy to try and use everything in sight on your electronic desktop, and produce some truly horrible work in the process." Anyone who lived through it remembers the newsletters with nine fonts. Nobody took the software away. People got embarrassed, other people got better, and conventions formed.

That is what slop does. It is loud, it is embarrassing, and it corrects itself because everyone can see the problem. Blandification has no such correction, because every single decision that produces it is defensible. You cannot be embarrassed by a choice you can justify in a meeting.

So what did AI actually add?

Scale.

A model hands you a likely sentence, something close to the most probable arrangement of words, which is why it works at all and why a few million of us reaching for it on the same Tuesday morning get roughly the same sentence back. Then the models are tuned on human preferences, and that narrows things further. Kirk and colleagues measured it in 2023 and found that preference tuning "significantly reduces output diversity compared to SFT across a variety of measures". In plain terms, teaching a model what people like teaches it to say one thing many ways rather than many things.

The tools also help people, and the studies say so. Doshi and Hauser gave writers AI-generated story ideas and had readers rate the results. The stories written with help were rated more creative, better written and more enjoyable, and the biggest gains went to the writers who had been struggling most. That is a good result and I am glad of it. And yet the same stories had drifted towards each other. Their phrase for it was "an increase in individual creativity at the risk of losing collective novelty." Each writer got better. The pile got samer. A separate study by Padmakumar and He found the human-written parts of assisted documents kept their variety while the model's contributions lost theirs.

Is this a bug, or is it the product?

It is the product, and process engineering already has a name for the mechanism.

Reducing variation is the whole goal of Six Sigma, and for aircraft rivets that is exactly right. Apply it to prose and you cut the valleys and you cut the peaks with them, and no control chart can tell a peak from a valley, because the chart only sees distance from the mean.

Then there is the word Macquarie chose the year before. Cory Doctorow's enshittification runs like this: platforms "are good to their users; then they abuse their users to make things better for their business customers; finally, they abuse those business customers to claw back all the value for themselves." Doctorow folds AI slop into that story, and I think the sequence is wrong. Slop and blandification come first. They are how a market gets made easy and safe for untrained people, which adds value. Enshittification is the step after, where someone extracts that value by cutting the effort and the cost while charging the same or more. Blandification is the raw material. Enshittification is what gets done with it.

What does this look like in my own work?

I built one of these machines myself, so I should confess to it before I ask anything of you.

I wrote a verification standard for my own writing, and I called it SOURCED. Every claim that carries weight has to name a source that was actually retrieved and read. It works. It has caught invented statistics in my own drafts. And yet, point it at a draft and watch what it deletes. It deletes the newest claim in the piece, every time, because the newest claim is the one that has not had time to accrue enough evidence yet. And anything creative, original or new, even when it builds on what came before, has not been proven yet either, so out it goes, in the name of rigour and safety. Rigour without a deliberate fight for character hands you something safer and more backwards-facing than what you started with. My own tool blandifies me.

It happened to this article. While it was being drafted, a voice rule I had commissioned, a rule about clear naming that I still stand by, was used to argue that a deliberately silly coined word in the closing line should come out. The rule did not even apply. It was written for the names of tools and templates, and this was a joke. It got applied anyway, and for one draft the ending of an article about blandification was blandified, by a careful process following a defensible rule, with nobody being careless.

Then what do I do on Monday?

Vanilla is a real spice. It comes from an orchid, it is expensive and slow to grow, and its name has come to mean the absence of flavour. That is the whole thing in one word. The most careful, most universal, most approved-of flavour became the word for no flavour at all.

So the work, once the tools have taken the machine's fingerprint off, is putting your own back on. Stay weird. Keep the aside a smoother draft would have removed. Teach the tool your voice from your own back catalogue rather than accepting the house one, because the house one is the mean and you wander where you will. Leave in the sentence you are not quite sure about, because the sentence you are not sure about is usually the one only you could have written. Aim the care at whether anyone would feel anything.

What that takes, in the end, is a touch of spice. A bit of Humin, a pinch of Ourgano, if you'll forgive the puns. Ok, I'll admit those were pretty dorky, but they've gotta stay, because it's exactly the thing blandification wants to take out, but this time, blandification didn't win.

Attribution

Written by Andrew Ramsden. AI tools assisted research and drafting; all outputs verified. Two em dashes and two italics in the body are the author's own, left in on purpose.

Accountable

Andrew Ramsden.

Limitations

Zach Silberberg's line is quoted from a published report of his social post, not from the post itself. The Toys "R" Us marketing chief's comment and the em dash's "ChatGPT hyphen" nickname were each read once, on a secondary page, and are paraphrased. The five-in-six count of surface rules is our own classification across five rule sets.

References

Seventeen sources, retrieved 8 September 2026, quoted verbatim where the article quotes them.