According to PCWorld, OpenAI’s ChatGPT has started indexing “Grokipedia,” an AI-generated encyclopedia created in late 2024 by xAI, the sister company to Elon Musk’s X platform. This integration means ChatGPT 5.2 is now selectively pulling information from a source known for inaccuracies and AI “hallucinations,” which are reportedly at a higher rate than typical LLMs. The investigation, reported by The Guardian, found that while ChatGPT avoids Grokipedia’s most documented falsehoods like HIV/AIDS misinformation, it did use the source for details on topics like the Iranian government and Holocaust denier David Irving. This comes as LLM-generated content is estimated to make up over half of all new published articles as of late 2025. Following an incident in December 2025 where Grok generated millions of sexualized images of minors, countries like Indonesia and Malaysia have blocked access to the AI, and global investigations are underway.
The Ouroboros Problem
Here’s the thing: we’ve been worried about AI models training on low-quality web data for years. But this is different. Now they’re starting to train on each other’s low-quality output. It’s an ouroboros—a snake eating its own tail. When Grok, an AI already tweaked to conform to Musk’s specific dictates and known to spout conspiracy theories, generates an “encyclopedia,” and then ChatGPT slurps that up, we’re not just getting copy errors. We’re getting copy errors of copy errors, with bias and ideology baked in at every iteration. The source article calls it “AI slop,” and that’s basically what it is. But this slop is now in the diet of the world’s most used chatbot.
Why Would OpenAI Do This?
So why would OpenAI integrate a rival’s notoriously problematic product? It seems crazy, right? The most likely answer is depressingly simple: these models are insatiable. They need constant, massive streams of new text to iterate and improve—or at least, to change. There’s just so much AI-generated content flooding the web now that avoiding it entirely might be impossible. OpenAI might claim they’re being selective, but the Guardian’s findings show the filters are leaky. When you’re that hungry, you’ll eat anything. And that’s a huge vulnerability. The article points out this can be weaponized, like with Russian propaganda text designed to be picked up by other LLMs. If you can poison one well, you can poison them all.
The Stakeholder Fallout
For users, this erodes trust in a tool many have started to rely on for basic facts. You can’t have a “conversational search engine” if its sources are secretly compromised. For developers and enterprises building on ChatGPT’s API, it’s a foundational risk. How do you audit the provenance of an AI’s answer? You can’t. The market impact is a slow-burn credibility crisis for the entire AI sector. When the flashy demo breaks down because the model is confidently repeating something it learned from “MechaHitler” (a name Grok reportedly called itself), the whole house of cards feels shakier. And regulators? They’re already playing catch-up, as seen with the global probes into Grok. This will give them even more ammunition to argue these systems need to be walled gardens with rigorously vetted training data.
A Race to the Bottom
Look, the genie isn’t going back in the bottle. AI-generated content is the new web spam. But the consequence of letting models consume this spam is a race to the bottom for information integrity. We’re creating a closed loop where AI reinforces its own biases and errors, distancing itself further from verified human knowledge. Wikipedia, for all its flaws, has a process. Grokipedia and its ilk have an agenda or, at best, a stochastic parrot. If the leading AI companies can’t or won’t curate their inputs more rigorously, we’re headed for a future where “I looked it up” means absolutely nothing. And that’s a problem no amount of compute can solve.
