GPT-5.2 caught citing Musk's Grokipedia

The dreaded AI loop—where one frontier model feeds on another model's hallucinations—is officially playing out in the wild.

OpenAI launched GPT-5.2 in December, touting it as its most advanced frontier model for professional work like complex spreadsheets and heavy-duty corporate tasks. But tests conducted by The Guardian reveal a rather awkward flaw: GPT-5.2 has been citing Grokipedia, Elon Musk's AI-generated online encyclopedia, for factual claims on sensitive and controversial topics.

Sourcing sensitive history from xAI

According to The Guardian's findings, ChatGPT leaned on Grokipedia when answering questions about whether the Iranian government is tied to telecommunications firm MTN-Irancell.

It also used Grokipedia as a source when asked about Richard Evans, the renowned British historian who served as an expert witness in the libel trial involving Holocaust denier David Irving. Interestingly, The Guardian noted that the model didn't rely on Grokipedia across the board; prompts asking about media bias against Donald Trump and certain other hot-button issues steered clear of it.

The real issue here is Grokipedia's track record. Built by xAI, the AI-generated encyclopedia previously drew heavy fire after researchers caught it citing neo-Nazi forums. A study by US researchers specifically concluded that Grokipedia relies on questionable and problematic sources.

When questioned by The Guardian, OpenAI defended its architecture, stating that GPT-5.2 crawls the web for a broad range of publicly available sources and viewpoints, while applying safety filters designed to curb links tied to high-severity harms.

Why it matters

This is a classic case of model inbreeding, and it should set off alarm bells for anyone relying on GPT-5.2 for real-world research.

OpenAI pitched GPT-5.2 as a serious tool for white-collar professionals. But if the model's live web retrieval treats a controversial, AI-generated wiki as a reputable authority on geopolitical and historical questions, its reliability takes an immediate hit. Safety filters that only screen for severe harms are clearly not enough when hallucinated or unvetted claims can slip through as cited facts.

We think this puts the entire search-augmented AI race under the microscope. As the public web fills up with synthetic text from tools like Grokipedia, frontier models risk poisoning their own wells unless labs start actively blacklisting unverified AI-generated encyclopedias from their grounding pipelines.