AI slop is a process problem

Two dictionaries settled the vocabulary last year. In November, Australia's Macquarie Dictionary chose AI slop as its word of the year and defined it as low quality content made by generative AI, often with errors, that nobody asked for. In December, the editors of Webster's dictionary in the United States chose plain slop: “digital content of low quality that is produced usually in quantity by means of artificial intelligence.” Their president, Greg Barlow, told the Associated Press: “It's almost a defiant word when it comes to AI.”

The developer Simon Willison offered the simplest test in May 2024: “if it's mindlessly generated and thrust upon someone who didn't ask for it, slop is the perfect term for it.”

For a company, the word names a judgement customers now make in a second. Our reading is that slop is not a property of the tool. It is what comes out of a process with no references, no rules and nobody who says no. That makes it fixable.

What slop looks like at work

Slop is not only the strange video in a feed. Researchers at BetterUp Labs and the Stanford Social Media Lab gave the office version a name in the Harvard Business Review: workslop, work made with AI that looks good but lacks substance. In their survey of 1'150 full time desk workers in the United States, 40 percent had received some in the previous month, and each case took about two hours to resolve.

Marketing has its own version: the post that could belong to any company, the picture with the familiar glossy light, the text that says nothing in three paragraphs. Each costs little to make, and the reader pays.

Why it happens

Four causes come up again and again.

The first is the default. Anthropic's applied AI team explained last November why its own model, left without direction, keeps producing the same safe designs: it samples from the most probable centre of what it has seen. They call it distributional convergence, and they note that users call the result the AI slop aesthetic. A study in the journal Patterns found the same for images. Researchers let an image model and a captioning model feed each other for 100 rounds, 700 times, and every run drifted into a few generic motifs they called visual elevator music. They suspect the missing ingredient is the corrective pressure of human interaction.

The second is the missing reference. A prompt describes, a reference shows. Without real pictures of the product and the light, the model fills the gaps with its average.

The third is the missing rule. The model does not know that your brand never uses gradients, never calls anything a solution and always shows the real product. If those rules are not written down where the machine reads them, for the machine they do not exist.

The fourth is the missing review. Willison put it bluntly: “sharing unreviewed content that has been artificially generated with other people is rude.”

Audiences notice, even when they cannot tell

The strongest evidence this year comes from a test by Ipsos and the Newhouse School at Syracuse University, published in May. They showed 20 ads to 3'000 American consumers: ten made by people for brands such as Febreze, H&M and Chewy, and ten made entirely by AI from the same brief, used exactly as generated. Only 25 percent of those who saw an AI ad were at least somewhat confident it was made by AI. Yet the human ads were 14 percent stronger on Ipsos's measure of short term sales effect and 17 percent stronger on long term brand effect. The authors put it in one line: credible is not the same as compelling.

Stated attitudes point the same way. In a June survey by The Harris Poll with the 4As and Infillion, 73 percent of consumers said they would be less likely to trust an ad they suspected was made with AI, and 63 percent said they would be less likely to buy from a brand that uses AI generated ads.

A global survey of 22'000 consumers in 22 markets, commissioned by the ad verification company DoubleVerify and released in July, adds the nuance that matters. 42 percent said low quality or uncanny AI advertising would hurt their opinion of a brand, while 40 percent viewed polished, professional AI ads positively. Its chief executive, Mark Zagorski, summed it up: “AI itself is not what determines engagement. The quality of the output does.”

Platforms are pushing back

The platforms have moved from tolerance to rules. In his letter for 2026, YouTube's chief executive Neal Mohan gave a section to managing AI slop and said YouTube is building on the systems it uses against spam and clickbait to reduce the spread of low quality, repetitive content. In July, TechCrunch reported, YouTube spelled out what can no longer earn advertising money: generic, repetitive or template based videos, content made to distress, and AI personas discussing sensitive subjects such as health and finance.

Pinterest added settings in October 2025 that let people see fewer AI images in the categories where they pile up: beauty, art, fashion and home decor. Deezer said in July that fully AI generated tracks passed half of all new music delivered to it each day at their June peak, yet made up only 1 to 3 percent of streams. It keeps them out of its algorithmic recommendations and editorial playlists.

The direction is the same everywhere. Made with AI is allowed. Made in bulk, without care, is pushed down.

How we keep it out

We make pictures, films and posts with AI every day, so we have had to answer this in practice, with a step for each cause.

Locked styles. In Reel, every style is built from real references, with light, lens and set chosen the way a photographer would, and tested until the client approves it. Every image and clip follows that recipe, so the hundredth picture looks like the first.

Brand rules in the brain. Studio captures a brand's DNA, voice, positioning and copy rules once and locks them. Twin keeps the company's rules and the way it writes in its AI brain, so every assistant works inside them instead of guessing.

People at the ideas and the picks, and a review before anything leaves. In Reel, ideas, references and styles are agreed before anything is made, every result is checked before the client sees it, and a person picks every image and clip that is used. In Tempo, every idea is tested before it is scheduled, and a person approves every post.

We want AI to make, check and speed up as much of the work as it can. The ideas, the picks and the responsibility for what goes out stay with people, because that is where quality is decided and where a name is on the line.

What a company can do

Start with the four causes. Collect twenty real pictures you are proud of and use them as references. Write down your brand rules where every tool can read them: the colours, the words you use and the ones you never use, what the product must always look like. Name the person who picks, and give them the right to say no to most of what the machine makes. Put a review between the generator and the public, every time.

Slop is cheap to make and expensive to be known for. The fix is not less AI. It is more judgement around it.