I don't know if not getting the idea right, but I'm pretty sure people refer to AI outputs as "slop" not due to (only) repetitiveness. According to some sources:
[1] Wikipedia
> AI slop is digital content made with generative artificial intelligence, specifically when perceived to show a lack of effort, quality or deeper meaning, and an overwhelming volume of production.[1][4][5] Coined in the 2020s, the term has a pejorative connotation similar to spam.[4]
[2] Urban dictionary
> Low-quality randomly generated AI content (images, accounts, text, etc) that has been flooding social media sites among other pages.
Yes, I know those may not be the best primary sources, but I'd say the main shared meaning of the word is lack of quality and effort, not repetitiveness itself.
I've been using ChatGPT fairly regularly for about a year. Mostly as an editor/brainstorming-partner/copy-reviewer.
Lots of things have changed in that year, but the things that haven't are:
* So, so many em-dashes. All over the place. (I've tried various ways to get it to stop. None of them have worked long term).
* Random emojis.
* Affirmations at the start of messages. ("That's a great idea!") With a brief pause when 5 launched. But it's back and worse than ever now.
* Weird adjectives it gets stuck on like "deep experience".
* Randomly bolded words.
Honestly, it's kind of helpful because it makes it really easy to recognize content that people have copied and pasted out of ChatGPT. But apart from that, it's wild to me that a $500bn company hasn't managed to fix those persistent challenges over the course of a year.
I honestly can’t always distinguish AI slop from the formulaic corp-speak used in emails and memos and brochure websites and other marketing. I’m guessing that must be a large component of the training matter.
I don't think that's a coincidence. Right now a lot of the business proposition for LLM bots is selling it to corporations as the ultimate corporate yes-man.
There will be an intersection when the techniques and continued refinements in making tall tale signs of AI and new powerful model meets where it becomes very time consuming, expensive and difficult to tell between human generated and AI generated content.
We are already at a point where we can trick large number of the population, it can without a doubt close the gap even further where we question anything and everything.
Beyond forensics, which require large capital investment and operating costs, to be able to detect AI vs human content will be limited in terms of access. It will be so that its not that we can't detect AI content anymore its that most people cannot afford the service to detect it and thus they lose interest.
This has side effect of making live performances by humans scarce and in valuable.
That’s not what “slop” means. Slop is output produced by generative AI without regards to its quality, not the telltale tics that current models tend to exhibit.
Instead of "surgically adjusting" logits within an existing model, couldn't you just build the slop detector into the loss function during the initial training stage?
ScholarlyArticle: "Antislop: A Comprehensive Framework for Identifying and Eliminating Repetitive Patterns in Language Models" (2025) https://arxiv.org/abs/2510.15061 :
> Abstract: [...] Our approach combines three innovations: (1) The Antislop Sampler, which uses backtracking to suppress unwanted strings at inference time without destroying vocabulary; (2) An automated pipeline that profiles model-specific slop against human baselines and generates training data; (3) Final Token Preference Optimization (FTPO), a novel fine-tuning method that operates on individual tokens, surgically adjusting logits wherever a banned pattern has appeared in an inference trace.
This seems to be fundamentally based on n-grams and manually built regexes. "Slop", or more narrowly annoying -isms and model stereotypes, is not just repetitive n-gram sequences, mode collapse manifests itself semantically. Sometimes repetition/stereotyping is desirable (you need semantics to understand if it's the case), and sometimes undesirable repetition is undetectable by n-grams and regexes, especially in languages that rely on word formation. Fixing the mode collapse probably needs a sufficiently powerful reference model of semantic diversity, which doesn't currently exist.
This is the epitome of patching symptoms rather than treating the disease. Even if you suppress the obvious syntactic slop like 'it's not X but Y', you have no reason to believe you've fixed mode-collapse on higher more important levels like semantics and creativity. (For example, Claude LLMs have always struck me as mode-collapsed on a semantic level: they don't have the blatant verbal tics of 4o but somehow they still 'go in circles'.) Which will potentially severely hinder the truly high-value applications of LLMs to creative applications like frontier research. To the extent that this succeeds in hiding the brain damage in contemporary LLMs, it arguably is a cure worse than the disease.
Interesting work but this strikes me as a somewhat quixotic fight against inevitable tendencies of statistical models. Reinforcement learning has a single goal, an agreeable mean. Reinforcement learning stops when the LLM produces agreeable responses more often than not, the only way you can achieve absolute certainty here is if you tune it for an infinite amount of time. I also don't see how this method couldn't be subsumed by a simpler method like dynamic temperature adjustment. Transformers are fully capable of generating unpredictable yet semantic text based on a single hyperparameter. Maybe it would make more sense to simply experiment with different temperature settings. Usually it's a fixed value.
Parsing a LLM as the measure of a series of qc metrics, which isolate for preference strings, whether lexical weights or parameters. Can this create the rules for formal understanding or correlating libraries of Babel?
Searle's paper calls these questions, script, or a story.
25 comments
[ 2.6 ms ] story [ 42.8 ms ] thread[1] Wikipedia
> AI slop is digital content made with generative artificial intelligence, specifically when perceived to show a lack of effort, quality or deeper meaning, and an overwhelming volume of production.[1][4][5] Coined in the 2020s, the term has a pejorative connotation similar to spam.[4]
[2] Urban dictionary
> Low-quality randomly generated AI content (images, accounts, text, etc) that has been flooding social media sites among other pages.
Yes, I know those may not be the best primary sources, but I'd say the main shared meaning of the word is lack of quality and effort, not repetitiveness itself.
[1] https://en.wikipedia.org/wiki/AI_slop
[2] https://www.urbandictionary.com/define.php?term=AI+slop
Lots of things have changed in that year, but the things that haven't are:
* So, so many em-dashes. All over the place. (I've tried various ways to get it to stop. None of them have worked long term).
* Random emojis.
* Affirmations at the start of messages. ("That's a great idea!") With a brief pause when 5 launched. But it's back and worse than ever now.
* Weird adjectives it gets stuck on like "deep experience".
* Randomly bolded words.
Honestly, it's kind of helpful because it makes it really easy to recognize content that people have copied and pasted out of ChatGPT. But apart from that, it's wild to me that a $500bn company hasn't managed to fix those persistent challenges over the course of a year.
I assume the beginning of the answer is given to a cheaper, faster model, so that the slower, more expensive one can have time to think.
It keeps the conversation lively and natural for most people.
Would be interesting to test if it's true, by disabling it with a system prompt, and measure if the time-to-answer is slower for the first word.
We are already at a point where we can trick large number of the population, it can without a doubt close the gap even further where we question anything and everything.
Beyond forensics, which require large capital investment and operating costs, to be able to detect AI vs human content will be limited in terms of access. It will be so that its not that we can't detect AI content anymore its that most people cannot afford the service to detect it and thus they lose interest.
This has side effect of making live performances by humans scarce and in valuable.
> Abstract: [...] Our approach combines three innovations: (1) The Antislop Sampler, which uses backtracking to suppress unwanted strings at inference time without destroying vocabulary; (2) An automated pipeline that profiles model-specific slop against human baselines and generates training data; (3) Final Token Preference Optimization (FTPO), a novel fine-tuning method that operates on individual tokens, surgically adjusting logits wherever a banned pattern has appeared in an inference trace.
From https://news.ycombinator.com/item?id=45546037#45585680 , an additional potential method:
>> Could build a simple heuristic: if similar memory content gets created/updated N times within short timeframe, flag it as potential loop
Oof—-gotcha here’s how I’d handle that
Clutch choice—-here’s a few refinements
Sweet—-let me just…
Ok, here’s the receipts
I love your passion! Let’s try to keep it civil ok?
(Thinking) the user still appears annoyed
—————————————-
I think this annoys them also and yet they can’t change it? Or are they not dogfooding?
Searle's paper calls these questions, script, or a story.
[1]:https://web.archive.org/web/20071210043312/http://members.ao...