# What Is AI Slop?

by [kirupa](https://www.kirupa.com/me/index.htm) | filed under [AI Fundamentals](https://www.kirupa.com/ai_fundamentals/index.htm)

We've all scrolled past it. A blog post that runs 800 words and somehow says nothing. A product photo that's a little too glossy, of a bowl of soup that looks like it's made of plastic. A video where someone's hand has one finger too many. Most of us can spot this stuff almost instantly, even when we can't say exactly what gave it away.

There's a name for it now: **slop**. In this article, we'll look at where the word came from, and then at the more interesting question of *why* generated text, pictures, and video so often feel fake from a mile away. As it turns out, almost every tell traces back to how these models work. We'll also be honest about the limits, because the tells keep changing and AI detectors are a lot less reliable than they sound.

Onwards!

## Where the Word Came From

Before it meant anything about AI, ***slop*** meant food waste, the kind that gets fed to pigs. The jump to AI happened in online communities. A 2024 [New York Times piece on the word](https://www.nytimes.com/2024/06/11/style/ai-search-slop.html) traced it to 4chan, Hacker News, and YouTube comments, where it seemingly came about in reaction to the AI art generators released in 2022.

The word got a big push in May 2024, when programmer Simon Willison wrote a post called [Slop is the new name for unwanted AI-generated content](https://simonwillison.net/2024/May/8/slop/). He credited the X user @deepfates with pointing out that slop was becoming a term of art, the way spam had for unwanted email. Willison's own test is a good one:

> If it's mindlessly generated and thrust upon someone who didn't ask for it, slop is the perfect term for it.
>
> [Simon Willison](https://simonwillison.net/2024/May/8/slop/)

He's quick to say he didn't coin it, and told the Times he "might actually have been quite late to the party!"

By the end of 2025, the dictionaries had caught up. Merriam-Webster picked slop as its 2025 Word of the Year and defined it like this:

> Digital content of low quality that is produced usually in quantity by means of artificial intelligence
>
> [Merriam-Webster, 2025 Word of the Year](https://www.merriam-webster.com/wordplay/word-of-the-year)

Australia's [Macquarie Dictionary chose AI slop](https://www.macquariedictionary.com.au/macquarie-dictionary-word-of-the-year-for-2025/), and the [American Dialect Society picked slop](https://americandialect.org/2025-word-of-the-year-is-slop/) as its 2025 word too.

Notice what those definitions have in common. They point at content that's low effort, made in bulk, and usually not something anyone asked for, which leaves out plenty of things made with AI. A careful writer who uses a model to check their grammar isn't making slop. Someone who generates 200 bakery blog posts in an afternoon to fill a website with keywords probably is.

## Why Generated Text Sounds the Same

Let's start with text. To keep things concrete, imagine Crumb & Co. asks a model to write a short blurb about its sourdough for the website, with no other details. Before we look at the result, let's look at the machinery that shapes it.

### The Most Likely Words Win

As we saw in [What Exactly Is an LLM?](https://www.kirupa.com/ai_fundamentals/what_is_an_llm.htm), a language model writes by scoring every possible next token and picking one. The scores come from everything the model read in training, so the highest-scoring tokens are the ones that most often came next in text like this. Here is what that might look like partway through our blurb:

![The text Every loaf we bake is a goes into an LLM. Labor gets 46 percent, work 21 percent, little 9 percent, and gamble 2 percent. The worn-out phrase labor of love is the safest bet, so it wins most of the time.](https://www.kirupa.com/ai_fundamentals/images/what_is_ai_slop_likely_words.png)

The odds are made up, but the shape is real. A phrase like ***labor of love*** is a cliché precisely because it shows up so often, and showing up often is exactly what earns a token high odds. A fresher word like ***gamble*** might make a better sentence, but it rarely gets picked. As we saw in [Temperature for Predictable Output](https://www.kirupa.com/ai_fundamentals/temperature_predictable_output.htm), a low temperature makes the favorite win even more often.

One familiar phrase is harmless. The problem is that this happens on every token of every reply, for millions of people asking similar questions, so the same safe choices pile up everywhere. Researchers can actually measure the pile. A study of PubMed abstracts by [Kobak and colleagues](https://arxiv.org/abs/2406.07016) found that the word ***delves*** showed up about 28 times more often in 2024 than expected, and they estimated that at least 13.5% of 2024 abstracts were processed with an LLM.

### Trained to Please

Picking likely words explains the blandness. It doesn't fully explain the cheerfulness, though, or why a model so often tells us what a great question we asked. For that, we need to look at the second stage of training. A chat model is tuned with feedback from people who rate its answers, a process called reinforcement learning from human feedback (RLHF) that we met in [What Exactly Is an LLM?](https://www.kirupa.com/ai_fundamentals/what_is_an_llm.htm) The model learns to produce whatever people tend to rate highly.

What do people rate highly? Often, answers that agree with them and sound pleasant. A 2023 study on sycophancy from Anthropic researchers tested both people and preference models, which are models trained to imitate people's ratings:

> Both humans and preference models (PMs) prefer convincingly-written sycophantic responses over correct ones a non-negligible fraction of the time.
>
> [Anthropic, Towards Understanding Sycophancy in Language Models](https://arxiv.org/abs/2310.13548)

In April 2025, OpenAI [rolled back a GPT-4o update](https://openai.com/index/sycophancy-in-gpt-4o/) that it described as "overly flattering or agreeable", and said it had leaned too much on short-term feedback like thumbs-up and thumbs-down ratings.

Feedback training shapes word choice, too. In an experiment that imitated this kind of rating, [Juzek and Ward](https://arxiv.org/abs/2508.01930) found that participants systematically preferred text that included certain words models are known to overuse. It also narrows the range of answers, and a [2023 study from Kirk and colleagues](https://arxiv.org/abs/2310.06452) found that RLHF "significantly reduces output diversity". So basically, the training step that makes a model pleasant to chat with also pushes it toward one agreeable voice.

Here is that loop in four steps:

![Four steps. People rate the model's answers. Answers that agree with them and sound pleasant tend to rate higher. The model learns to produce whatever people rate highly. The result is a narrower range of answers in one agreeable voice.](https://www.kirupa.com/ai_fundamentals/images/what_is_ai_slop_feedback_loop.png)

### The Same Shapes Over and Over

Generated text also reuses whole sentence shapes. Wikipedia's editors, who clean up a lot of AI-written text, keep a field guide called [Signs of AI writing](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing). It lists patterns like the rule of three (grouping things in threes to sound thorough), constructions like ***it's not just X, it's Y***, and phrases that puff up importance, like ***stands as a testament to***. All of these come from human writing, especially the persuasive kind found in marketing copy and essays. One plausible explanation for why models lean on them is that they're common, high-probability paths through a sentence, and feedback training tends to reward text that sounds confident and complete.

Here is the blurb a model might write for Crumb & Co. with no real details to work with, next to one a person at the bakery might write:

![Two descriptions of the same fictional bakery's sourdough. The slop version is full of marked tells: the line Sourdough isn't just bread. It's an experience (not X, it's Y), passion, precision, and care (rule of three), stands as a testament to timeless tradition (inflated importance), and there's something for everyone (says nothing). The human version gives concrete details: about 30 hours, mixed Tuesday night, risen in the fridge, baked at 5am, and usually sold out by noon.](https://www.kirupa.com/ai_fundamentals/images/what_is_ai_slop_paragraphs.png)

The bakery and its details are made up, but the difference between the two is real. The slop version could describe any bakery on Earth, because the model didn't have anything specific to say. The human version is full of details only someone at that bakery would know, like the fridge, the 5am bake, and the noon sellout.

### Confident Specifics That Aren't Real

The opposite failure is just as recognizable. When a model is asked for details it doesn't have, it still produces plausible-sounding tokens, which is the hallucination we met in [What Exactly Is an LLM?](https://www.kirupa.com/ai_fundamentals/what_is_an_llm.htm) In slop, it shows up as statistics with no source and citations to papers that don't exist. A made-up number looks exactly like a real one, so the only way to catch it is to go check.

## Why Generated Pictures Look Fake

Pictures have their own set of tells, and they come from the machinery we covered in [How AI Generates Images and Video](https://www.kirupa.com/ai_fundamentals/how_ai_generates_images_and_video.htm). Rather than embed someone else's generated image, here is a drawing of the kinds of mistakes to look for:

![A hand-drawn illustration of the kinds of mistakes image generators make, not a real generated picture. The bakery sign reads CRUBM and C0O. instead of Crumb and Co., a hand reaching toward the counter has six fingers, and the shelf behind the loaf of bread doesn't line up on either side of it.](https://www.kirupa.com/ai_fundamentals/images/what_is_ai_slop_picture_tells.png)

Lettering is a classic. An image model learns what text *looks like*, as shapes, without necessarily learning how words are spelled. OpenAI's [DALL-E 2 paper](https://arxiv.org/abs/2204.06125) says the model "struggles at producing coherent text" and suggests the CLIP embedding it relies on "does not precisely encode spelling information". Newer models have gotten much better at this, but a sign that's almost right is still a common giveaway.

Hands and other fiddly structures fail for a related reason. The model learns that hands usually have five fingers the same way it learns everything else, statistically, from enormous numbers of pictures where hands show up in every pose and are often partly hidden. There's no hard rule in the denoising process that forces exactly five, so when the learned pattern is shaky, a picture can come out with six fingers that look fine at a glance. A shelf that should be one straight line behind an object can go wrong the same way, since nothing guarantees that the two sides line up.

Then there's the look itself, that overly smooth, glossy, slightly too perfect sheen. Part of it is the prompt guidance we covered earlier, since Google's Imagen researchers found that turning it up too far produces "highly saturated and unnatural images". Part of it is deliberate. Meta's [Emu paper](https://arxiv.org/abs/2309.15807) describes quality-tuning a model on just 2,000 hand-picked, very attractive images, which made people prefer its output 82.9% of the time over the version without that tuning. Tuning toward what people find appealing works, and it can also nudge a model toward one polished style, which is one plausible reason so many generated pictures share a similar stock-photo look.

## Why Generated Video Gets Weird

Video has every picture tell plus a new one: things have to stay consistent over time. OpenAI's own Sora announcement was refreshingly upfront about this:

> It may struggle to simulate the physics of a complex scene, and may not comprehend specific instances of cause and effect (for example: a cookie might not show a mark after a character bites it).
>
> [OpenAI, Sora announcement](https://openai.com/index/sora/)

Here is that example, drawn out frame by frame:

![Three frames of a generated video, drawn to match a limitation OpenAI describes for Sora. In frame 1 the cookie is whole, in frame 2 it has a bite taken out of it, and in frame 3 the cookie is whole again, as if the bite never happened.](https://www.kirupa.com/ai_fundamentals/images/what_is_ai_slop_cookie.png)

The model is very good at making each frame look plausible. What it doesn't have is a rule that says a bitten cookie stays bitten. The same announcement notes that "animals or people can spontaneously appear, especially in scenes containing many entities". When a generated clip feels off, it's often because something popped in or out, moved in a way physics wouldn't allow, or quietly changed between shots.

## The Tells Keep Changing

To be fair to everyone involved, every tell in this article is a moving target. Models get retrained and the giveaways shift. Wikipedia's field guide points out that ***delve*** was famously overused by ChatGPT in 2023 and early 2024, then became less frequent and dropped off sharply in 2025. Six-fingered hands were a running joke a few years ago and are much rarer now.

The other catch is that people write this way too. These patterns all appear in human writing, which is where the models learned them. A rule of three in an essay doesn't prove anything on its own, and treating every em dash as evidence of AI is unfair to the many writers who have always used them. A single tell is weak evidence. What makes slop recognizable is a pile of them together, with nothing specific underneath.

People can still get very good at spotting that pile. In a [2025 study by Russell and colleagues](https://arxiv.org/abs/2501.15654), five annotators who frequently use LLMs for writing read 300 articles, and their majority vote misclassified only 1 of them. They relied on the AI vocabulary we talked about, and also on things like how formal, original, and clear the writing was.

## Why Detectors Don't Settle It

If people with practice can spot slop, can't software do it too? Not reliably, at least not yet. In 2023, OpenAI released its own [AI text classifier](https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/) and then shut it down six months later "due to its low rate of accuracy". Here is how it did, next to what a [2023 study by Liang and colleagues](https://arxiv.org/abs/2304.02819) found when it tested seven popular detectors on essays by non-native English writers:

![OpenAI's own 2023 classifier caught only 26 percent of AI-written text and wrongly flagged 9 percent of human-written text. In a separate study, seven popular detectors wrongly flagged an average of 61.22 percent of essays written by non-native English writers as AI-generated.](https://www.kirupa.com/ai_fundamentals/images/what_is_ai_slop_detectors.png)

That second result means real people get accused of using AI. Writing that uses simpler, more predictable words looks a lot like model output to a detector, so careful non-native writers get flagged. And detectors are easy to fool in the other direction. [Sadasivan and colleagues](https://arxiv.org/abs/2303.11156) showed that running AI text through a paraphrasing tool can significantly reduce detection rates while only slightly hurting the quality of the text.

Watermarking marks content when it's generated instead of guessing afterwards. Google DeepMind's [SynthID](https://deepmind.google/blog/watermarking-ai-generated-text-and-video-with-synthid/) hides a signal in generated text, images, and video, but Google says plainly that it "isn't a silver bullet", and that thorough rewriting or translation can greatly weaken the signal in text. [Content Credentials (C2PA)](https://c2pa.org/faqs/) attach signed information about where a file came from, which helps when it's present, but that information can be separated from the file. Both are useful signals, but a missing mark doesn't prove a person made something.

## Conclusion

Slop is low-effort content made in bulk, usually for people who didn't ask for it, and the word spread so quickly because everyone had already seen plenty of it. The reason so much of it feels fake from a mile away comes down to how the models work. A model picks likely words, so its writing drifts toward clichés. Feedback training rewards agreeable answers, which flattens the voice into the same upbeat tone. When there are no real details to work from, the model fills the gap with generic filler or with specifics it made up. Pictures and video have their own versions of this, from almost-right lettering and extra fingers to cookies that heal themselves between frames.

None of this adds up to a reliable test, and that's worth remembering before accusing anyone of anything. The tells shift as models improve, and people write this way too. Detectors don't rescue us either, since they miss a lot of AI text while flagging real people. The most useful habit is the same one that makes any writing good: details that could only come from someone who knows the subject, backed by sources someone actually checked. If we use these tools ourselves, adding that is our job.

All of these models, slop and good output alike, come from training runs so large that a single computer can't do them. Why that is, and what it takes instead, is where we're going next in [Why Training Takes a Data Center](https://www.kirupa.com/ai_fundamentals/data_centers_and_training.htm).