Jean-Marc Watecamps
nixCraft 🐧the irony of this is that Perplexity is posting this one 🤣
NexusSfan
Caden
Xylight@tarix29 @nixCraft I thought this too, but then I learned that a lot of modern LLMs are trained on intentionally synthesized/AI generated data.
the only one I know for a fact was is gpt-oss, but I'm sure the other big ones do too.
so unfortunately model collapse wont really affect these guys that much
Fahri Reza 🍉@Xylight @tarix29 @nixCraft I think that is just how they create generative AI model
they have one model that generates random samples and one model that judges whether the samples look like somethings that human creates
the closer that the first model in compliance with the second model you have a generative AI model
Ivan Tsenov 🇧🇬 🇺🇦@nixCraft I bet even the post is AI generated.
technikhil@nixCraft Don't you think the folks in the AI business would be aware of this and trying to find new algorithms and models to mitigate this ? 🤔🤔
Jim Jones@technikhil there is no reliable way to fingerprint AI generated content as of yet. There are a number of papers on this in computer science. It is called the ouroboros problem and it leads to model collapse. Some developers are reporting that most coding agents they are using are creating the same strange code and can't figure out why. The ouroboros problem may already be here.
@GreatBigTable Thanks for clarifying will try to read more about model collapse and the ouroboros problem🙏🏾
Tobias Knöppler@technikhil @nixCraft That post above literally *is* the folks in the AI business warning about this.
Reason being: There is no lasting solution to this. Every time you discover a tool that can recognize AI generated content you created a trainer to remove that flaw.
However, as of now, we just don't have tools like these (not with sufficient reliability).
Although, one way around it would be actually paying humans to produce content for them. But unfortunately that's not a reality I see coming.
Stumpy The Mutt@nixCraft When does the feedback degenerate into a continuous scream like it does with audio systems?
@AlexBlock has moved@StumpyTheMutt @nixCraft Great analogy!
Jmvars
GrumpyDad 🇺🇦🇵🇸@nixCraft Even AI agrees that AI is bad...
Nico 
@nixCraft Libraries will have a strong comeback in the next few years as the internet will become an AI advertising place about AI ads
Jack Yan (甄爵恩)@Illdisposed @nixCraft So glad I kept my 1983 encyclopædia set.
tschenkel@jackyan @Illdisposed @nixCraft
I pulled a local copy of Wikipedia in 2022, just a few months after ChatGPT release.
@iveyline
Tom 
Original post if anyone cares: https://x.com/AskPerplexity/status/1978587411803156942
Ask Perplexity (@AskPerplexity) on X
Oxford researchers just confirmed what we feared: The internet as we knew it is dying. AI content went from ~5% in 2020 to 48% by May 2025. Projections say 90%+ by next year. Why? AI articles cost <$0.01. Human writers cost $10-100. But the real crisis is model collapse. When
Orlaya
dissident
Peter H@dissident @thomas Me too, so I pasted the post into perplexity and asked it to list the sources.
It came up with these:
https://www.linkedin.com/posts/sachaghi_oxford-researchers-just-confirmed-our-worst-activity-7384488192781156352-L9SC/ (exactly the same post)
https://www.forbes.com.au/news/innovation/is-ai-quietly-killing-itself-and-the-internet/
https://futurism.com/artificial-intelligence/over-50-percent-internet-ai-slop
The last one shows the numers used in the graph and names this report as the source: https://graphite.io/five-percent/more-articles-are-now-created-by-ai-than-humans
The LinkedIn post is interesting: is it a source for Perplexity, or just a copy? 🤔
Oxford researchers just confirmed our worst fears: The internet as we once knew it is dying. In 2020, only about 5% of online content was AI-generated. By May 2025, that number reached 48%, and… | Sacha Ghiglione
Oxford researchers just confirmed our worst fears: The internet as we once knew it is dying. In 2020, only about 5% of online content was AI-generated. By May 2025, that number reached 48%, and projections show it could exceed 90% next year. The reason is simple: AI-generated articles cost less than a cent, while human-written ones cost $10 to $100. But the real danger isn’t just cheap content. It’s model collapse. When AI systems learn from AI-generated material, quality deteriorates. It’s like making a photocopy of a photocopy. The details fade, originality disappears, and everything starts to sound the same. It’s a recursive spiral. Today’s AI sludge becomes tomorrow’s training data, producing even lower-quality output, which then feeds the next generation.
ECHAEA
Arapalla
LAUREN
Coq@nixCraft
Studies have shown that only 3 to 5 iterations of AI out –> AI in content lead to litteraly bullshit content, with nonsense in every LLM "production". In other words, we are at maximum 4 iterations from content collapse.
Studies have shown that only 3 to 5 iterations of AI out –> AI in content lead to litteraly bullshit content, with nonsense in every LLM "production". In other words, we are at maximum 4 iterations from content collapse.
Em 
Droid Boy 
Patrick 
Light Owl ☸️@nixCraft I think the wide-open, globally-social parts of the internet are suffering a "tragedy of the commons". It's sort like one super-huge communal mailbox, getting stuffed to the brim with junkmail flyers, now that they're almost free to create with #AI .
To avoid all this effectively "junk-mail"-like content, people will have to stick to the smaller groups and organizations to which they belong in real-life, being more regional, national, city-level, community-level, special interest, etc. They will get more closed-in and private (not unlike medieval castles with draw-bridges), and will involve more shibboleths to prove humanity (say, passkeys, minisign signatures, etc), in order to exclude #AI.
To avoid all this effectively "junk-mail"-like content, people will have to stick to the smaller groups and organizations to which they belong in real-life, being more regional, national, city-level, community-level, special interest, etc. They will get more closed-in and private (not unlike medieval castles with draw-bridges), and will involve more shibboleths to prove humanity (say, passkeys, minisign signatures, etc), in order to exclude #AI.
Alex@rtnVFRmedia Suffolk UK@sbb @nixCraft a lot of car-related forums already have various "human checks" to avoid spammers/viral marketers which were a scourge even before AI - to join Team BHP (the Indian motoring forum) I had to leverage my own Asian ancestry and even answer a quiz about Tata Motors (which wasn't too difficult as they own Jaguar Land Rover here in UK)
@nixCraft how much of this affect finance?
the market trend creates AI response which generates market trend that the AI will respond
Dizzy
James 🦉 #FBPE 🇪🇺@nixCraft A curious game, the only winning move is not to play.
𝐬𝐭𝐞𝐯𝐄 🌼
beka valentine
Leigh Garland@nixCraft mmm, those numbers look very made up. An AI tells you what you want to hear. Where's the link to the paper? 48% of what content? On the Internet? Seems highly unlikely.
@nixCraft the number of completely credulous replies in this thread doesn't give me a lot of hope.
Folks. This is a screengrab of an X post, by an AI account, with no link, no names, no data, no citations. "Oxford researchers" would be disappointed in you.
AI tells you what you want to hear. Even if it's how awful AI is.
@nixCraft Not to mention it's a _sales pitch_, telling you how little an AI article actually costs to produce.
A lot of people left X to join Madison because of the fearmongering and hate. Don't let it permeate here.
@toychicken someone posted bunch of sources in this thread, but yeah they do use fear as a marketing tool
@nixCraft yeah, and the actual article is a lot more bullish about prospects.
No 'Oxford researchers', but an AI powered SEO firm, based in US
Not 'all content' but % of _new_ content.
oh, and using an AI tool, on a dubious sample set.
I guess the takeaway from this is:
* Don't let the Fedi become like X
* Don't trust a guy with a Mushroom avatar.
DC LUG
DJGummikuh@nixCraft I believe "AI Incest" is more appropriate
Talia Hussain@nixCraft if AI trained on AI output gets worse, why would humans trained on AI output not also get worse?
Mudlark 
@nixCraft it’s not ironic. When ai companies post things like this, it’s to bait journalists and social media into sharing it and keeping their ai companies in the public eye, to keep people talking about them.
It’s a deliberate and effective tactic for these companies to talk about the ‘dangers of ai’ or the ‘possible collapse of the ai industry’ or some other existential crisis. It’s effective because people fall for it over and over 
Henrik Pauli@nixCraft "But the real crisis is model collapse"
Yes! Ha ha ha... yes!
Also deserves a microscopic violin.
Luyin@nixCraft https://en.wikipedia.org/wiki/Dead_Internet_theory
> The dead Internet theory is a conspiracy theory which asserts that since around 2016 the Internet has consisted mainly of bot activity and automatically generated content manipulated by algorithmic curation, as part of a coordinated and intentional effort to control the population and minimize organic human activity.
xs4me2The irony indeed…
The second law of thermodynamics illustrated. Everything turns to shit if not actively organized and fed. AI does not add intelligence, it just digests what is there. It also does not think, it just copies. We have not found the soul yet… which might be a good thing even.
methuselah
Elrick_Winter@nixCraft For those interested: The study is real, but has a lot of caveats. Both in the positive and negative direction.
I'm linking to an article about the original one, because the actual source is behind a paywall:
https://futurism.com/artificial-intelligence/over-50-percent-internet-ai-slop
Audrey Winter has moved!@nixCraft Dead Internet Theory is a self-fulfilling prophecy.
@audreygwinter @nixCraft Every technology has a specific life expectancy. AI simply moved the Internet's date up a bit.
Matt HallI never _wanted_ all the articles written by humans for $10 - $100.
So... The slop wagon was already here for my purposes. The internet isn't a _magazine_. It's a _creative_ network and repository. I'm here for the folks who're sharing ideas. Not the folks who're _paid_ to generate 'content' or the folks who're generating 'content' to sell ads.
AI just _amplified_ this behavior because it's 'cheaper' for the 'content generator'.
I'm holding out for the art makers and thinkers.
Resolviendo la incógnita 🌐@nixCraft Homogeneous, bland and boring content that wastes your time has been a ubiquitous reality for more than 10 years. Clickbait headlines were aimed at them. The problem has always been SEO and dependence on Google. AI made it easier the creation of such content.