There is a breed of gremlin in every large language model. I don't mean bugs or errors, but something more ingrained. These are behavioural tics that have been embedded so thoroughly that you can't exorcise them with custom prompts or by being firm with the user interface.
Try it. Put a word off limits in your system prompt. Go ahead and ban it in all caps and add three exclamation points for good measure. The gremlin will still make its way back in.
This is not speculation. It is measurable, published and getting worse. For those of us who put out professional work with the help of AI, these gremlins are what give us away. Your more sophisticated readers can smell that we are using AI the moment they read our words!
The gremlin is in the weights, not the prompt
Consider an instruction such as 'do not use the word delve'. On paper, it is a soft constraint, meant to nudge the probability distribution so the model steers clear of it. But you have to remember that the hard prior is the training data; the word's likelihood is determined by billions of parameters honed on trillions of tokens. Your little command is no more than a post-it note on the dashboard of a freight train.
There is a term for it among researchers: the pink elephant problem (see arxiv 2402.07896). They have found that being explicit and telling a model not to put something in its output can have the opposite effect. You are putting the forbidden concept right into the model's active context, and it will attend to it. In a transformer, attention is not avoidance.
So do not think the model is being obstinate when your hand-picked list of banned words does not work. It is simply being statistical. And statistics are poor at following orders.
Vocabulary, structure and punctuation patterns
There are tells in the way a paper is put together: its vocabulary, structure and punctuation. Take the work of Kobak and colleagues (2024) in Science Advances, for instance. They put 15 million PubMed-indexed biomedical abstracts under the microscope to see how word usage changed once ChatGPT came on the scene. 'Delves' is a case in point; in 2024, it was used 28 times as often as the pre-LLM norm. That is not a 28 per cent increase, but 28 times over. By their reckoning, LLMs have been at work on no less than 13.5% of this year's biomedical abstracts, and in certain sub-fields you can put that figure at 40%.
Then there is the Antislop framework from Paech (ICLR 2026, 2025), which takes things a step further. By pitting model-specific frequencies against what a human would produce, they identified patterns with an overrepresentation of more than 1,000-fold. You could call them statistical fingerprints, as good as a watermark; they are hardly matters of style. But don't let the vocabulary be your only concern. The structural gremlins are another matter altogether, more insidious when you have them all in one place and harder to pick out on their own.
The participial gremlin
AI has a way of tacking present participles onto the end of its sentences with some frequency. You will spot the formula: a statement, then a comma and an -ing word to spell out a vague outcome.
Take for example, 'The company reported a 12% rise in quarterly output, showcasing the effectiveness of its operational strategy.'
The participial phrase is superfluous; the first part of the sentence has already made its point. It is a matter of statistics for the machine, not rhetoric. Put a human to the task, and he would either put that in a separate sentence with something new to add or drop it altogether.
The negative parallelism gremlin
The Wikipedia guide on AI writing (August 2025) has gone into this in depth. Every model, no matter the register or length of output, will come up with the 'It's not X, it's Y' type of construction. Or one of its cousins: 'not just X but also Y', 'this goes beyond X', 'more than just X'. In theory, it sets up a contrast; in practice, it just rephrases the same idea. It is putting on airs of depth the material doesn't warrant.
The significance gremlin
This one is the most common offender of them all. The AI will put in a line to tell you something is important or part of a wider trend, even if the rest of the paragraph offers no evidence for it. Antislop researchers have the explanation: the model is not so much editorialising as regressing to the mean, substituting hard facts with language of generic importance because that is what its training data is full of.
You know the sort of thing: 'marking a pivotal moment in the evolution of...', 'represents a broader shift in...', or 'against the backdrop of'. They are nothing but padding. Do away with them, and you are left with a paragraph that either says the same thing or says nothing.
The em dash gremlin
This is the one that drives me truly demented! You could say the em dash is now the punctuation of choice for discussion. A 2025 study by v4nn4 of the technology subreddits on Reddit put a number to this, showing that as AI has been taken up, so too has the frequency of the em dash in posts. Then there are the scientific abstracts; an analysis of those reveals em dashes have more than doubled in use from 2021 to 2025. There is a simple reason for it. GPT-4 and the models that have followed were put through their paces with digitised 19th-century books for training material, and those old texts run about 30% higher on em dashes than you will find in modern English prose. In effect, the model has picked up the punctuation habits of its training data.
The structural gremlins
You will find the same sort of distinctiveness at the document level. The paragraphs are all of a piece, uniform in length and with openings you can put your finger on ('Furthermore', 'However', 'In addition'). There is an over-smoothness to the transitions; a human would be more likely to make an abrupt change of pace. Information is doled out in equal measure from one paragraph to the next, rather than having the writer dwell on one thing and move on to another. And then you have the section endings, which tend to be formulaic in restating the point with a 'This underscores...' or 'Taken together...'.
By themselves these are no cause for alarm. But put them together, and you have what detection researchers call a cluster. That is what an experienced reader will pick up on.
Your readers are better at this than you think
Don't underestimate your readers. You can see the results of a 2025 preprint on arxiv (2501.15654) for an experiment into human ability to spot AI writing. It turns out that those in the habit of using LLMs to put words on paper are quite adept at telling when it is machine-made; they called it right in most cases. In fact, if you put five of these regular users to a vote, they would only get one article wrong in 300.
That puts them ahead of the pack of commercial detection tools, even when the AI has been paraphrased or otherwise 'humanised' to throw them off. Tools such as Originality.AI and GPTZero have their share of errors. The truth is, your reader will detect it. Should they be AI users themselves, they have already been honing their eye for the very kinds of patterns you are churning out.
Why the gremlins survive your instructions
You don't need the Antislop paper to tell you what most of us have suspected, but it does put a number on it: instructing a model to stay away from certain words 'has limited efficacy and may induce a backfire effect'. Their approach is to work at the token-level probability as the text is being generated and put the damper on patterns before they reach the output. And it works. You can't say the same for system prompts with any degree of confidence.
Then there are the figures. Prompting by itself will quash about 80% of the slop. Fine on the surface, but look at the other 20% and you are looking at your professional reputation. An article of 2,000 words with that kind of leakage will have enough gremlin clusters in it to be noticed.
It is a structural issue. The gremlins aren't in your instructions; they are in the weights. When you prompt the model, you are haggling with parameters fixed during training on data you have no access to, and it is hardly an even negotiation.
The gremlins breed in clusters
What all the detection research so far has turned up is this: you identify AI via signal clustering, not by looking for single flags. A lone hedging adverb won't give a piece away as machine-made. But put three of those in one paragraph along with a couple of AI's pet words, a tricolon and a participial phrase and you have a cluster. The reader will see it even if he can't put a name to it.
So in practice, going after individual words is a fool's errand. You can excise 'delve' from your writing but keep the participial constructions, the uniform rhythm of the paragraphs and your statements of significance; it makes no difference. The cluster will reconstitute itself with other words. The gremlins are inclined to adapt that way.
Specificity is gremlin repellent
When you let AI have its way, it will default to statistical generality. But put a precise number in a sentence, or a named entity, some hard fact of observation or a specific constraint, and the AI can't simply substitute for it. Take 'Woolworths closed 17 stores in regional Queensland between March and June 2025, citing lease costs above $480 per square metre.' You cannot replace that with 'a major retailer continued to optimise its national footprint' without obviously losing the information.
Be sure to ground your claims in specifics so the gremlins have no place to land. Don't mistake this for a matter of style. It is a structural defence. Specificity yields sentences that could only come from someone with knowledge of the subject. Generic language of significance could be applied to any old thing. Your reader will be making that measurement one way or another, whether they are conscious of it or not.
The gremlin census
If you are in the process of putting together or overhauling a writing protocol, you will want to review the research on high-risk patterns. We are not talking about an exhaustive list of proscribed terms so much as the clusters that give something its statistical fingerprint.
Take the vocabulary with the most pronounced AI tell: words like delve, tapestry, underscore, showcase and pivotal come to mind, as do intricate, meticulous, realm, bolster, garner and interplay. You will also spot boasts where one might simply say 'has', or where 'navigate' is used in a metaphorical sense. Then there is the matter of empower, unlock, unleash and unprecedented. And for those looking to put a spin on 'is' or 'has', they tend to reach for serves as, stands as, functions as, marks, represents, features or offers.
The structure tells a story too. We see comma-plus-participial constructions and negative parallelism ('not X but Y' and the like). There is a tendency toward unearned profundity in transitions such as 'But here's the thing' or 'The real question is...', along with formulaic ways to end a section and significance statements that lack evidence to support them. Don't forget the reflexive tricolon or the old 'challenges and future prospects' routine.
As for punctuation and formatting, em dashes and an over-reliance on semicolons are red flags, as is unexplained bolding or the use of emoji in a professional setting. At the document level you may find uniform paragraph lengths and openings, even idea density, hedging in clusters and transitions that are just a bit too smooth.
Living with the gremlins
You are not going to eliminate AI patterns entirely when using AI tools. The gremlins are architectural. What you can do is learn what they look like, edit for them specifically and build enough specificity into your output that the clusters cannot form.
The people reading your work are getting better at spotting patterns faster than the models are at hiding them. That asymmetry is the thing to pay attention to.