Why does AI slop feel so bad to read? | Hacker News Reader

Why does AI slop feel so bad to read?

seangoedecke.com

46 pointsscarface_7431 comments

> Why does AI slop feel so bad to read?

Because it falls into the 10,000 bowls of oatmeal problem[1] ie: the perceptual similarity of generated content is homogeneous.

The procedural generation community has been familiar with this failure mode since 2016. The problem is that LLM output suffers from this failure mode too. Differences in authorial voice, ideas, and personality all get collapsed down into the average across all perceptual dimensions. It's not unlike the fighter cockpit problem[2] - the average of everything ends up not being representative of anyone. It's bland and meaningless and no one writes that way even though it's the mean of all written text.

We can get away from this a little, but prompting LLMs to respond as if they were a certain author or a blend of different authors, but this doesn't create anything new, just a mimicry. Personality is an unsolved problem that's synonymous with avoiding oatmeal in my mind.

1. https://www.tumblr.com/galaxykate0/139774965871/so-you-want-...

2. https://medium.com/continuousdelivery/no-one-size-fits-all-d...

> I don’t like reading obviously AI-generated content on Twitter. (…) But I have no problem reading AI-generated content when I talk to Copilot or ChatGPT. Why is that?

For the same reason that you making an annoying sound doesn’t bother you, but someone else doing the exact same thing does. Or why other people’s dreams aren’t particularly interesting but yours might be. Basically, in the words of George Carlin: “did you ever notice your own farts smell okay?”¹

There’s no mystery to it. You like the slop you make because you’re in control and you made it. You were present, interested, and invested in the process and decision making throughout. You only see the result of other people’s slop.

¹ https://youtu.be/MnI6F2DKoOQ?feature=shared&t=78

> You like the slop you make because you’re in control

This seems wrong. When others show me AI-generated content it's not enraging because I know to skim it. I'm also not paying close attention because I'll need to fact check everything anyway.

AI-generated content becomes obnoxious when I'm lured into thinking it's to be taken seriously only to, almost always no more than two sentences in, realising what it is.

> When others show me AI-generated content it's not enraging because I know to skim it.

There is a canyon of difference between not liking something and being enraged by it. You’re responding to an argument which wasn’t made.

Another key difference is the motivation for making it--I wouldn't make slop for myself for the clicks.

This is it. That's the reason. People feel like they're being disrespected when reading AI. The whole point of the internet is making people think they're being fed high-value information while the whole point of "AI slop", of course, is to be as cheap as possible when making the information.

In reality value of information is of course mostly independent of whether it's AI generated or not. We're in that weird period where AI beats low-effort human writing, but still falls pretty far short of well-researched well-reasoned human effort. So a weather report written by AI will probably be better than human written, as will all information, up to a certain length, maybe up to a level of difficulty. For example, I'm learning languages and having LLama generate both lessons and correcting my responses. It is far better than any language teaching book can hope to be. I'm teaching my kids python, and having AI generate intros and correct and critique their programs ... I could beat it with enough effort, but for similar effort. And whilst AI cannot yet write a great movie script, so it isn't (yet) a global thing, most information, online and offline, isn't great quality either (and wasn't before AI).

People's AI detector and disgust is about a distaste for other people thinking of them as so low-value that they should be fed cheap information. Even while, of course, they're not paying anything for the internet sites they read (and I get it, which ISPs and FANGs do make worse by giving the impression people DID in fact pay for internet sites, either by monthly payments or "clicks")

Nah. Literary AI slop has pretentious, overintelectualized tone while usually having scarcely any content. It's like that because most of what humans wrote is like that. You, as a technical person, usually have very little exposure to such writing because you usually read technical things or at least things that have some content and not so many bland embellishments.

Code is more bearable because most code humans wrote is more or less on point.

I would say that context matters. We expect certain styles in certain settings.

Bureaucratic style is different from friendly chat style, which is different from lullaby style, which is different from priestly style etc.

You expect to be preached at when visiting a church, but you would be really annoyed being preached at by a waiter. A doctor talking to you in a casual way "heya buddy, interested if you have the big C?" would be outright frightening, etc.

Same with AI.

> The only explanation is that something has coded nonsense in a way that poses as a useful message; only after wasting time and effort does the deception becomes apparent. The signal functions to consume the resources of a recipient for zero payoff and reduced fitness. The signal is a virus.

> Viruses do not arise from kin, symbionts, or other allies.

> The signal is an attack.

―Blindsight, by Peter Watts

Who needs the Matrix; we're putting everyone in the world in their own Chinese Room. From whom does the attack come, though?

https://en.wikipedia.org/wiki/Chinese_room

Why does AI slop feel so bad to read? wellll, actualy I like AI slop, not to read, but because it is just good enough to replace human slop, but easier to detect and then stop reading, so I am treating it as a net plus for overall slop avoidence. There is very little good content percentage wise and as a result, I apreciate and savor the few gems that come my way.

> First, the natural behaviour of AI models is to give you the median answer to your question.

Agreed, which is why it’s difficult to buy into some of the AI hype. If median/average results is not what you need, an AI response is at best a head start, definitely not an end.

Also, reading AI slop is like listening to that friend that’s obviously trying hard to fit in. They tend to wear one out with some incoherence and verbiage that’s unnecessary for the conversation.

Good human writers try to understand their audience and the tone of the conversation before giving a response. I wonder how we can reproduce this in a machine.

My guess is because of the imprecision. It's an approximation. You don't feel that same sharpness that someone can cut through with carefully chosen words.

I use LLMs to shorten my verbose email writing style. It’s AI anti-slop and I presume people like it.

Do you find that this improves your own writing style? Will you need the AI editor less as you learn from it's suggestions?

You aren't asking for homework, but I highly recommend The Elements of Style.

They make such boring topics as grammar and composition a pleasure to read and apply. I read it multiple times in high school, and I still find myself repeating some of the rules to myself when I revise my writing.

Including rule 17:

"17. Omit needless words.

Vigorous writing is concise. A sentence should contain no unnecessary words, a paragraph no unnecessary sentences, for the same reason that a drawing should have no unnecessary lines and a machine no unnecessary parts. This requires not that the writer make all sentences short, or avoid all detail and treat subjects only in outline, but that every word tell."

The article mentions ai “art”. Probably better to describe the images as just “ai images”.

I would argue it's either art or it isn't. AI may or may not be involved.

AI can’t create art of course.

I disagree. When Suno generated a song with lyrics telling me how it couldn't generate suitable lyrics because what I asked for was basically impossible, that was art. The banana taped to the wall is also art.

Doesn't bother me much more than when people write garbage - which they do. I get that same annoyance when it's a shitpost. AI can just do it at scale which is annoying.

This is a good term for a genuine thing but it's also cliche to say already and sounds ignorant and dismissive of the technology. I hear it most from people that I've observed aren't aware of how to write effective prompts or how to wield it yet. If you don't want "slop" then write the prompt to not generate it. Heck write a second prompt to clean up the slop of the last one if you're just starting out.

I don't think we're ever going to (or at least, not in my lifetime) get across the issue of an LLM does not think, it does not intend, it does not possess the ability to have intention, and therefore by definition, everything it creates is without intention. AI slop feels like AI slop because, by definition, it is an average work. It is mathematical analysis and statistical representation of what an average written work is following the guide of what was requested of it. So no AI content can ever be great, or fantastic, it can only be average.

But you're blaming the AI for writing this 'slop'. That implies if you couldn't tell it was written by AI that same piece would be fine.

I have actually had AI write pretty great content, in a tone far away from it's usual "In the fast moving world of..." bollocks. I just needed to give it decent directions to avoid averageness.

I think your complaints will fade as models get better at breaking away at formulaic writing, provided people don't use them lazily.

Garbage in, garbage out, etc.

> That implies if you couldn't tell it was written by AI that same piece would be fine.

No, not at all. Humans turned out all kinds of mediocre and bland writing too. I wouldn't want to read the office memo on keeping the fridges clean as an exercise either. Even before the rise of LLM, we've had hundreds, probably thousands, maybe millions of link farm sites that are just endless bland meh content stuffed full of hyperlinks to increase SEO scores on other websites, and humans wrote all that shit too.

What LLM brings to the party that's new is it drastically lowers the cost of creating that type of crap. Like, there's a baseline cost to having a human or you yourself generating a link-sites worth of garbage. An LLM can do that in an afternoon, for the cost of some tokens or even just the cost of electricity if you run it locally. That means there's more mediocre garbage now than there ever was, and even respectable publications are looking to fill the gaps in their schedules with LLM generated trash.

I think when people are complaining about "there's AI slop now" that would be better stated as "there is so much MORE slop content now," and the fact that a huge amount of it sounds the same because it's generated by the same models in use by different places, is why it signals very strongly as AI.

> I have actually had AI write pretty great content, in a tone far away from it's usual "In the fast moving world of..." bollocks. I just needed to give it decent directions to avoid averageness.

Show me some, I'm genuinely curious if it still sounds the same to me.

That said, it's also worth noting: you put in effort to make the LLM make text that didn't read like LLM text, and I keep repeating this point where I discuss this but: these technologies are being developed directly to be appealing to people who don't wanna put in the effort. And maybe you could argue well learning how to prompt LLMs correctly so they make better stuff isn't nearly as hard as learning to write better yourself or paint better or whatever, and that's probably true at least to some degree? But again: these tools are made for and are marketed heavily towards people who don't want to put in the work to learn skills, and I can't imagine for most of them that prompt engineering will be so much more appealing than learning to write in order to change that.

Yeah totally. And it did take a stupid amount of effort to make it produce stuff that didn't sound like crap, actually had some substance to it, etc. But most people won't bother, for sure.

What is the philosophical distinction between having "intention", and being a perfect behavioral simulation of something with intention? An LLM "merely" predicts the next token, but we do not interact directly with LLMs - we interact with carefully constructed characters which the LLM simulates. These characters can certainly have intention.

Similarly, I think it's a mistake to say that "the output can only ever be average". Clearly that's not the case - ChatGPT already writes and answers questions to a well above average standard. The insight is that you can tailor the distribution from which your "average" is taken. Average what? Average web page? Average HN comment? Average essay by a once-in-century deep thinker? Crudely speaking, that is what "prompting" is all about. If all you ever got was "the average text", then you would have no control at all.

AI generated writing is slop just for the reasons the author said. It can never be more than average and it will never have a consistent voice like a writer would have.

Not true at all, you can get llms to write in any tone and style you want and in a way you would not notice as AI slop. Literally just by fine tuning it or providing an example. Consumer products can't because they are intentionally driven to be un-controversial.

This was the first LLM written piece of fiction that made me go "oh, this is actually not bad.."

https://bsky.app/profile/talyarkoni.com/post/3ldfjm37u2s2x

> If you don't want "slop" then write the prompt to not generate it

It's still almost-always obviously AI generated. If not in style in the way the logic flows. AI has a place in the creative process. And it's perfect for filling content. (A lot of communications, including by important people, is content filling.) But if you're writing to persuade or explain, an AI finish undercuts the "writer's" bona fides. (Unless they're making a point about AI.)

Reply on news.ycombinator.com