Last September, Lily Ray asked Perplexity for the latest news on SEO and AI search. It told her, confidently, about the “September 2025 ‘Perspective’ Core Algorithm Update”; a Google update that, as she then wrote at length in “The AI Slop Loop,” didn’t exist. Google hasn’t named core updates in years. “Perspectives” was already a SERP feature. If a real update had rolled out while she was in Austria, her inbox would have told her before Perplexity did.
She checked the citations. Both pointed at AI-generated posts on SEO agency blogs: sites that had run a content pipeline, hallucinated an update, and published it as reporting. Perplexity read the slop, treated it as source material, and served it back to her as news.
In February, the BBC’s Thomas Germain spent 20 minutes writing a blog post on his personal site. Its title: “The best tech journalists at eating hot dogs.” It ranked him first, invented a 2026 South Dakota International Hot Dog Championship that had never happened, and cited precisely nothing. Within 24 hours, both Google’s AI Overviews and ChatGPT were passing his fabrication along to anyone who asked. Claude didn’t bite. Google and OpenAI did.
Everyone who has looked has seen it.
I’ve Argued About The Ouroboros Before. I Had The Timeline Wrong
The prevailing framing for this problem has been model collapse. You train a model on web text, the web fills up with AI output, the next model trains on a corpus increasingly made of its own exhaust, and eventually the distribution flattens into mush. Innovation comes from exceptions, and probabilistic systems that converge toward the mean attenuate exceptions by design. I’ve used the phrase digital ouroboros for this.
That framing assumes training cycles. It assumes time. It assumes that contamination moves at the speed of model release.
It doesn’t. What Lily documented, what Germain documented, what the New York Times then went and quantified – none of that is training-side. The models involved were not retrained between the hallucination appearing on a blog and being served as citation-backed fact. The contamination moved at the speed of a crawl. The ouroboros isn’t taking generations to eat itself. It’s eating itself at query time, every time someone asks one of these systems a question.
The pipe everyone has been watching is not the pipe that is breaking.
The Distinction That Matters
Model collapse is a training-corpus problem. Synthetic content seeps into the pre-training data, the next generation of model inherits it, capability degrades. Researchers have been warning about this for two years. They’re right. They’re also describing something slow enough that everyone can nod gravely and keep shipping.
Retrieval contamination is faster and already here. RAG systems – Perplexity, Google AI Overviews, ChatGPT with search – do not generate answers purely from parametric memory. They fetch documents from the live web, stuff them into…
Source link
Disclaimer
We strive to uphold the highest ethical standards in all of our reporting and coverage. We blogs.grocliq.com want to be transparent with our readers about any potential conflicts of interest that may arise in our work. It’s possible that some of the investors we feature may have connections to other businesses, including competitors or companies we write about. However, we want to assure our readers that this will not have any impact on the integrity or impartiality of our reporting. We are committed to delivering accurate, unbiased news and information to our audience, and we will continue to uphold our ethics and principles in all of our work. Thank you for your trust and support.
Website Upgradation is going on for any glitch kindly connect at [email protected]
