AI slop: what it is and how to spot it before it harms you

AI can write, and that is not a problem in itself. The problem comes when publishing entire pages costs so little that the internet fills with texts which, at first glance, merely seem correct. They stand up to a quick look, but if you stay for two minutes, you notice that they are empty inside: everything looks professional, but nothing quite conveys any judgement.

In 2025, institutions and leading media outlets ended up summarising this with a word that caught on: slop. The term took hold because it names a feeling many people recognise immediately, even if they find it difficult to explain: reading something apparently well written while sensing that there is nobody there.

The important question is not so much who typed each sentence as what ultimately gets published. When a text enters the world without judgement, without a voice, and without anyone willing to answer for the result, we are no longer talking only about technology: we are talking about an editorial problem. At the beginning of 2026, this is neither a curiosity nor a passing fad. It is already an entire layer of the web competing for the same space as original work. For a small brand, the cost is direct: a loss of credibility.

What exactly is AI slop?

If we define it as “AI-generated content”, we have misunderstood the problem. Some uses of AI are perfectly legitimate and, in fact, very useful. AI slop is something else. It is content published because it is cheap to produce, not because it is worth reading. It exists to occupy space, feed a calendar, or capture an impression, not to explain an idea better or genuinely help the reader.

AI slop can, in fact, take many forms. Sometimes it is an absurd generated video designed to hold attention for fifteen seconds. At other times, it is a product page that seems correct but does not answer a single important question. And it is often an entire service page full of apparently SEO-oriented statements which, when you finish reading them, have told you nothing that another fifty businesses could not have said.

The pattern is usually easy to recognise when you look closely. The text sounds reasonable, it is orderly, and it uses the correct vocabulary. What it lacks is something else: real experience, context, and a decision with some weight. When you read it, you feel that there are words but no perspective.

Why it is so dangerous

It is dangerous because it lowers both the average quality of the internet and the threshold of trust. Once a user has seen enough pages that look impeccable but say nothing, they learn to distrust more quickly. They read less and retain less information. They need more evidence to believe you. And, for any business, this makes everything that follows more expensive: it becomes harder to persuade, harder to differentiate yourself, harder to close a sale, and harder to recover authority once you have destroyed it.

The situation is particularly delicate for a local or small brand because its margin of trust is much narrower than that of a large platform. A multinational can survive a great deal of mediocre content thanks to brand inertia, distribution, and budget. A local business cannot. If its website sounds like an inflated template or a collection of predictable phrases, the damage is not aesthetic. It is commercial.

Articles from Fisher Phillips, Brandsafety Institute, UF News, and AirOps agree, with nuances, on a fairly simple idea: the important discussion concerns quality, context, and editorial responsibility. The machine alone is not the problem. The problem comes when human judgement disappears from the process (what we call human in the loop) and only speed remains.

How to identify AI slop in a text

This is where many people look for a shortcut: “Are there detectors?” Yes, there are. Are they useful? Sometimes, as a signal. Are they sufficient? No. And they will probably become less so over time. In an environment where a huge proportion of content is already generated partly or entirely by AI, blindly trusting detection tools is a bad idea. In theory, some products advertise extremely high accuracy; in practice, when they encounter mixed content, rewritten texts, or drafts edited with a little care, their reliability falls and false positives soar.

Ultimately, the judgement that matters remains human. And if you have a reasonably good eye, AI slop is noticeable.

The first warning sign is elegant vagueness. It is text that talks about “innovative solutions”, “digital ecosystems”, “sustainable growth”, or “today's landscape” without ever touching a specific situation. Everything is acceptable. Nothing is alive. There is no sentence that only that company, in that industry, could say to that customer.

The second is overly geometric structure. Paragraphs of a very similar length, an overly regular rhythm, lists that could fit any website, artificial transitions intended to sound professional, and excessive symmetry. When everything glides too smoothly, it is often because nobody has introduced friction into the text. By design—it is a matter of probability and statistics—a language model tends to predict the most probable and safest continuation of a sentence. If nobody forces it to move beyond that, the result is entirely correct and banal at the same time.

The third is the lack of a point of view. A genuinely human text takes a position, even when it qualifies it. It knows how to say yes to this, no to that, there is a limit here, or this seems like a good idea but is not. AI slop tends to shelter in neutrality because LLMs are optimised to minimise risk and maximise probability. They try to be useful to everyone and end up convincing almost no one.

The fourth is the absence of examples with weight. When a text never touches a specific case, a recognisable situation, or a real scene, it is usually because it is not based on any experience. You can talk about a plumber in Sitges who relies on emergencies, reputation, and trust. Or you can talk about “professionals seeking to maximise their digital presence”. Both are sentences. But they do not carry the same weight, and the reader notices.

The fifth is the voice test. This remains the simplest and most effective. Read the text aloud. If you stumble, if you run out of breath, if you would never say it that way in front of a client, or if it sounds as though someone wanted to seem clever instead of wanting to be clear, take care. It may be “correct”, but it is not yet good.

There are also smaller but revealing indicators: In Spanish and Catalan We Do Not Capitalise Every Word. If you find a website with headings like this, run. The same applies to the systematic use of the dialogue dash (-): very common in the English-speaking world, it is fairly revealing evidence that nobody has reviewed this content. Individually, these signs do not always mean anything. When they cluster, they almost always mean something.

The underlying technical problem: how a model actually writes

If you want to understand why all this happens, we need to drop briefly to the technical level, but without turning it into an engineering class. A model such as ChatGPT (or Claude, or Gemini) does not write because it “knows” what is important to say. It writes because it predicts which sequence of words is statistically most likely to follow the previous one. This detail is fundamental because it explains why AI is so good at producing plausible language and so inconsistent at producing judgement.

An LLM is excellent at summarising patterns and generating useful variations on an idea. What it cannot do by itself is simply be human: decide what is worth publishing, what is too generic for a specific brand, which sentence is alive and which is merely correct, or when a page sounds as though it was written from inside the business rather than from within a statistical model.

When you give it a poor prompt, the model does exactly what you ask: it fills the gap with the most probable continuation. And in corporate environments, that continuation tends to be safe, polished, abstract, and lacking much friction. This is where an important part of the problem comes from. AI slop has as much to do with technology as it does with poorly designed processes.

The problem appears when you delegate judgement to it

There is an important distinction here. We are not anti-AI. That would be absurd, as well as dishonest. AI has democratised access to many useful capabilities and can greatly accelerate real processes: organising ideas, detecting repetition, comparing structures, generating intermediate versions, helping turn a rough draft into more workable material, or summarising research so the writer can start with more context.

All of this is legitimate. The problem begins with the mental leap that is damaging part of the web: confusing assistance with replacement. AI can help you write, but it cannot assume responsibility for deciding what is worth saying, how it should be said, what should not be published, and what commitment the brand makes when it publishes it.

How to avoid falling into it

The first rule is very simple: start with an idea that is yours. If there is no thesis, there is no article; there is only filler. Before touching any model, you need to know what you defend, what you criticise, and what you want the reader to take away. If even you do not understand this mental architecture, the model will fill it with clichés.

The second is to provide context before asking for text. The emptier the context, the more generic the result will be. If you use AI, do not simply ask it to “write me an article about SEO” or “write me a services page”. Give it your tone, positioning, ideal customer, proprietary data, legal boundaries, the terminology you do not want to use, and the ideas you do not want to repeat. Quality comes from context, not from an inflated prompt ordering it to act as the best copywriter in the world.

The third is to accept that the first draft is not the product. It is raw material. This is probably the clearest boundary between using AI as a tool and using it as a substitute for judgement. If you publish the first text it produces, you will almost certainly end up publishing something mediocre. The final copy needs a human hand to cut, decide, add examples, recover a voice, and accept that some of what was generated will have to be discarded.

The fourth is to work with a genuine standard of clarity. In many cases, this means lowering the level of abstraction and writing at an accessible reading level without patronising the reader, but also without hiding the idea behind fifty vague words. If a sentence can be expressed in ten clear words, it should not need twenty-five vague ones.

And the fifth, which seems minor but is anything but, is reading aloud. If a text does not survive this test, it is not ready. It is an almost physical validation, which is why it works so well.

The process matters too: copy-first versus design-first

Some slop does not come only from the model. It also comes from the process used to build websites. For a long time, many agencies have worked with a design-first approach: first they create the visual mock-up, often filled with dummy text or provisional statements, and then try to fit the copy into modules that are already fixed.

This method is convenient for quickly presenting an aesthetic, but it has a structural problem. When the design has already been decided before anyone thinks about what needs to be said, words become secondary material. They must fit the box, not the idea. This is where AI fits too well as a patch: it fills the gap, provides sufficiently presentable text, and helps finish the screen. But what is gained in speed is often lost in clarity, persuasion, and coherence.

We clearly prefer the copy-first approach, or at least a process in which the message does not appear at the end as if it were decoration. When the copy is written before or at the same time as the structure, the design stops being a pretty shell and starts organising a decision. The words no longer have to be forced into an arbitrary module; the module adapts to the real story the brand needs to tell.

That is why copywriting is not a superficial layer of a website. It is part of its architecture. How you explain a value proposition influences the visual hierarchy, design decisions, section lengths, the order of the blocks and, ultimately, the page's ability to take someone from curiosity to action.

What a local business should do in 2026

If you run a local business, the temptation to fill your website quickly with generated text is enormous. This is especially so because it seems efficient and because the market has become accustomed to confusing speed with professionalism. But this apparent efficiency is very costly if the result is a website that sounds like all the others, conveys no judgement, inspires no confidence, and does not help you differentiate yourself.

A website designed to attract customers does not need a hundred impersonal pages. It needs a clear voice, a clean structure, and enough substance for whoever arrives to think there is a real person there. Someone who knows their craft, understands the customer's problem, and is not hiding a lack of judgement behind correct language.

For a local business, content does not work alone. It works alongside reputation, how people find you, your Google Business listing, the photographs, the page's speed, and the overall impression you convey. If one of these elements feels false or too mechanical, it contaminates the rest. Unsupervised generated copy has a particular ability to do exactly that: lowering trust without anyone fully realising where the feeling comes from.

The right question goes beyond “whether AI is involved”

The right question is: is there someone responsible behind it? Is there someone who reviews, cuts, decides, and answers for what has been published? Is there someone who understands what an empty sentence is, what an overly generic promise is, and what idea genuinely deserves to be on a website? If the answer is no, you have a brand problem before you have a technology problem.

Very good things can be done with modern tools. We use them ourselves. But we do not sell speed without judgement, automation without perspective, or content published because a space had to be filled. A website must not merely look professional. It must convey that somebody knows what they are doing, and that feeling does not come from a magic formula or a clever prompt. It comes from a process in which technology helps but does not decide alone.

If you want to see how this idea translates into a real website, you can also read why Subur exists or how we approach bespoke web development.

Authorship

Written by Mario Vilar, a mathematician and software engineer, founder of Subur and directly responsible for the architecture of the websites we publish. I write about software, technical judgment, SEO and digital presence for small businesses, based on the practical experience of building them.

Based in Sitges, serving all of Catalunya. No sales team and no outsourced writing.