Where did the AI gremlins come from

It was May 2024 when a user on X called gabe put the words 'LLM slop' out there. LLM is shorthand for large language model, the kind of AI that churns out text. A day later, deepfates at Tumblr was already forecasting the term would be as commonplace as spam. By the end of the week, you had Simon Willison, the programmer and co-creator of Django, making his case in public for 'slop' to be the go-to term for any AI output you don't want. He put it in the same camp as spam. Before 1993 spam didn't have a name; once it did, the whole conversation shifted and it became something with laws and filters attached to it rather than an irritation you could just put up with. Willison's point was that AI slop is due for the same.

And it stuck because it is an exact word. Slop has nothing to do with quality per se but with intent, or lack thereof. It is content put together to pad a feed or get some eyeballs on it from someone who doesn't actually care if you read it.

The Shrimp Jesus effect

You could say the Shrimp Jesus phenomenon put a finer point on things. The slop problem was becoming impossible to look past. Facebook feeds were awash with AI renderings of a certain crustacean-bodied Christ figure known as Shrimp Jesus. Tens of thousands of accounts were liking these photorealistic oddities put out by pages looking to farm engagement. It was silly stuff, but it worked. You could see then that Facebook's algorithms were no better at telling the difference between what was of interest and what was simply exploiting the feed.

Then Google rolled out its AI Overviews and we saw some unanticipated trouble. The system would serve up top-of-search results like advice to put glue on your pizza so the cheese doesn't slide off, or to eat rocks for your digestion. These weren't isolated hiccups; they were born of an Onion piece or a satirical post on Reddit that the algorithm mistook for sound counsel.

Or take Coca-Cola's Christmas ad in November of that year. They did a generative video version of their old holiday spot and people were having none of it. The ad looked 'wrong', to be sure, but there was a feeling that you shouldn't be using AI to make something that is meant to be human and emotionally resonant.

In the end, the dictionaries made it official. In eighteen months time, Merriam-Webster and Macquarie both got around to it. For 2025, Merriam-Webster declared 'slop' its word of the year, while Australia's Macquarie went with 'AI slop'.

The vocabulary fingerprint

You could call it a vocabulary fingerprint. But the academic marker was of a different sort. A 2025 study in Science Advances from the University of Tubingen put numbers to it. The researchers took 15 million biomedical abstracts on file with PubMed and looked at the frequency of certain words in the period before and after ChatGPT's debut.

The results were telling. 'Delves' was up 28 times on its baseline, while 'showcasing' saw a ninefold rise. Then there were 'underscores', 'intricate', 'pivotal' and 'commendable'. All of them made statistically impossible leaps that followed the curve of ChatGPT adoption right down the line.

By their reckoning, some 13.5 per cent of all 2024 biomedical abstracts had run through a large language model; in a few sub-fields you would find the number as high as 40%. And we are not talking about undergraduate work here but published, peer-reviewed papers in the world's premier medical database.

There is a popular theory as to why the model makes these word choices, one that has to do with OpenAI's workforce. In training the model via RLHF, humans are put to work rating AI output to shape its behaviour. According to a Time piece from January 2023, much of this data labelling was done by people in Kenya and elsewhere. The idea is that the kind of formal English you get in Nigeria or Kenya will have you using 'delve' or 'commence' or 'utilise', and so the raters taught the model to view such terms as marks of quality.

The Tubingen authors don't go so far as to say that themselves. They put the vocabulary shifts on record without tying them to where the RLHF staff come from. It is a plausible link and often cited, but in the end it is an inference, not something the study can claim as a fact.

The content ratio

You can trace the 'dead internet theory' back to a 4chan conspiracy from 2021, one that held the line that bots were churning out most of what you see online and putting humans in the minority. These days, in 2025, it is less of a conspiracy and more of a matter of measurement.

Take the Ahrefs study from April this year: they looked at 900,000 new web pages and turned up AI-produced content in 74% of them, even if only 2.5% were pure AI with no human hand in the editing. Then there is Europol's 2022 projection that 90 percent of online material will be synthetic by 2026. Graphite has a more measured take, their analysis of 65,000 URLs suggesting that as of late 2024 AI and human articles are on an even keel.

The numbers vary depending on how you measure it and what you call AI-generated, but everyone agrees on the trend. With every iteration, the models are fed more of the output from the last model and less of the human original. In short, what we write is fast becoming a minority input into systems trained on the internet.

Detection and the arms race

In the months following ChatGPT's debut, a host of AI detection tools made their way to market: GPTZero, Originality.AI, and the like from Turnitin. What they do is put the statistical properties of your text under a microscope. They look at distribution patterns and things like burstiness, or the ebb and flow of sentence length and complexity, as well as perplexity, the metric for gauging how adept a language model is at predicting what comes next.

For all their utility, these are imperfect instruments. You still have to contend with false positives, which is an issue for non-native English speakers whose style can be mistaken for something a machine has put together.

If you want a more dependable read, try a human who is in the habit of using AI. A 2025 preprint by Russell, Karpinska and Iyyer (arxiv 2501.15654) makes the case for this. Their data shows that those who make regular use of large language models in their own work will spot AI output where it counts; in most test cases they were right on the money. In fact, when you put five such users to a vote, only one article in 300 was misclassified. That is a better track record than you will get from most of the commercial offerings, even if someone has gone to the trouble of paraphrasing to throw them off.

So don't underestimate your audience. If they are AI users themselves, they are far better at telling when you have used it than you might believe.

The position this site takes

I use AI tools every day. I run businesses that use AI for research, drafting, data management, CRM operations, content structuring and more. I am not writing from a position of opposition. I am writing from the position of someone who uses these tools commercially and is fed up watching people publish AI default patterns as if they were their own words. The models are better than they have ever been. The problem is the humans. They take the output, see that it is fluent, and hit publish. They do not push back on the em dashes, the 'not X but Y' constructions, the significance language that means nothing. They should know better by now.

AI, used well, accelerates the parts of knowledge work that benefit from speed: research synthesis, first-draft generation, pattern spotting across big datasets, structural editing, formatting. It is an accelerant for human capability when the human stays in the loop and does the thinking.

AI, used lazily, produces slop. And the problem with slop is not that it exists. Spam existed before AI. Bad writing existed before AI. The problem is that slop passes. It passes editorial review because the reviewer is skimming. It passes peer review because the reviewer is overloaded. It passes the author's own quality check because the author has started to believe that fluent text is the same as good text. The gremlins win when you stop noticing them.

The position this site takes is simple. Use AI and use it aggressively. But do not lie to yourself about what it is producing. If you cannot distinguish your own thinking from the model's statistical defaults, you have lost something vitally important, and your readers will notice before you do.

I went down this rabbit hole because the gremlins kept winning. I was writing content with AI tools, running it through my own writing protocol, editing line by line, and the output was still flagging as AI-generated. So I tried to fix it. I tried harder prompts. I tried banned word lists. I tried rewriting sentence by sentence until the detection tools stopped complaining.

What I found is that none of the tools agree with each other, and none of them are reliable. I have an article that is 99% grammatically correct, part AI written, part human. Originality.AI flags it as AI. So I run it through a humaniser like gpthuman.ai. Originality passes it. Then Grammarly grades the humanised version as poorly written, because the humaniser has degraded the grammar to fool the detector. So I fix the grammar. And Originality flags it again. I have also had pieces that pass Originality completely, then I change a few commas or swap a couple of words and the same tool scores the same piece as 100% AI produced. The tools are not consistent enough to be useful, and the only way to satisfy them is to publish worse writing than I started with.

The only way I could reliably pass the human writing test was to make the writing worse. Shorter sentences. Simpler vocabulary. Grammatical roughness. Deliberate imperfection. The tools are not measuring whether a human wrote it. They are measuring whether it looks polished enough to be suspicious. That is a problem, because the whole point of using AI is to produce better output, not to degrade it until a detector stops objecting.

This is where I ended up and I am not sure I am coming back. The detection tools do not work well enough to be trusted. The AI tools do not produce clean enough output to be left alone. The gap between the two is where every honest person using AI to write is stuck right now.

The articles

This site publishes a series of articles examining the specific patterns, mechanisms and consequences of AI-produced content. Each is researched, sourced and written with AI tools while trying to keep the gremlins out. The irony is deliberate.

• The gremlins living in AI

  The vocabulary fingerprints, structural habits and punctuation tics that AI cannot suppress, no matter what you put in the prompt. The statistical patterns that trained readers already recognise.

• The gremlins are eating each other

  What happens when AI models train on AI-produced content. Model collapse, knowledge collapse, the farmed salmon problem, and why mixing in human data is not a complete fix.

  Coming

More articles will follow as the research base develops and the gremlins evolve. They always do.