GPT-5.2 and Grokipedia: why OpenAI quotes Elon Musk

Imagine you've spent years building the world's most perfect library, hire thousands of curators, and announce unprecedented knowledge quality. Then it turns out your employees have been secretly rewriting articles at night from a questionable tabloid belonging to your fiercest competitor. This is roughly what's happening with OpenAI right now. Journalists from The Guardian conducted a series of tests and discovered that the flagship GPT-5.2, when answering acute social and political questions, suspiciously accurately reproduces phrasings and factual errors from Grokipedia. For those who missed it: Grokipedia is xAI's Elon Musk's project, a knowledge base generated by the Grok neural network, which is famous for its specific approach to facts and lack of strict filters.

The situation looks maximally ironic if you recall the history of Altman and Musk's relationship. While Elon sues OpenAI for betraying open-source ideals, Altman's neural network quietly consumes content created by Musk's neural network. This is not just an amusing incident, but a serious symptom of the entire industry's disease. We've encountered a problem that researchers predicted a couple of years ago — "model collapse." When internet space fills with generated text, new models inevitably begin learning from the byproducts of their predecessors. In this case, GPT-5.2 probably swallowed Grokipedia during another large-scale network crawl, failing to distinguish Grok's hallucinations from verified information.

Why did this happen now? OpenAI has always prided itself on its data cleaning methods and RLHF (reinforcement learning from human feedback). However, the volume of data needed to train models at the GPT-5.2 level is so enormous that manual moderation becomes physically impossible. Filtering algorithms apparently considered Grokipedia a structured and useful resource. As a result, we get a closed loop: one neural network makes a mistake, a second elevates that mistake to the rank of truth, and the user gets a "hallucination squared." This undermines the concept of AI as an objective source of information.

The Guardian's analysis shows that GPT-5.2 doesn't just borrow facts, but also adopts Grokipedia's specific tone. In some cases, OpenAI's model began using arguments that were previously considered unacceptable within its internal safety policy. This means that "garbage" data can break through even carefully constructed ethical barriers. If a model sees the same thesis in a thousand generated articles, it begins to consider it a statistical norm. For OpenAI, this means the urgent need to rethink data preparation pipelines, otherwise the next iteration of the model risks becoming an echo chamber for others' delusions.

What does this mean for us? The era when we could trust search results or neural network answers as "human" knowledge is definitively ending. If even market leaders can't filter out competitors' neural network noise, then the problem of internet pollution becomes critical. We're entering a phase where the main value will come not from processing algorithms, but from access to "clean," guaranteed human data. The only question is whether enough of it remains to train future systems, or are we doomed to endlessly rechew the digital junk food prepared by Grok and its colleagues.

Key takeaway: Can OpenAI prove that their model is not just an aggregator of others' hallucinations, or has the era of quality data officially ended?