Summary: A January 24, 2026 Guardian test found that GPT-5.2 cited xAI’s AI-generated Grokipedia nine times across more than a dozen questions. The citations clustered on long-tail topics, and some answers repeated details The Guardian had already debunked. OpenAI said its web search aims to draw broadly from public sources and uses safety filters and programs against low-credibility sources.
The Guardian’s reporter Aisha Down ran the hands-on test and published the findings on January 24, 2026: answering more than a dozen different questions, GPT-5.2 — the newest model from OpenAI — cited Grokipedia, xAI’s AI-generated encyclopedia, nine times along the way. TechCrunch picked the story up the following day with a framing of its own: Grokipedia material is “escaping containment” from the Musk ecosystem and landing inside rival products.
The citation itself is not what matters; a rival’s product showing up in the sources is almost incidental. What matters is what came out afterward: on some topics GPT-5.2 went further than Wikipedia, and it passed along details The Guardian had already debunked. Once chatbots treat an AI-written encyclopedia as a source, source quality is no longer one vendor’s problem — it belongs to the entire retrieval ecosystem. GPT-5.2 is the flagship delivered by OpenAI’s accelerated post-New-Year cadence — background in our 2026 opening outlook.
Where the Citations Actually Landed
Every cited topic sat on obscure ground: Iran’s political structures (Basij paramilitary salaries, who owns the Mostazafan Foundation) and the life of British historian Sir Richard Evans, who took the stand as an expert witness against Holocaust denier David Irving. On the ties between MTN-Irancell and the supreme leader’s office, ChatGPT — Grokipedia citation in hand — claimed more than Wikipedia does; and its account of Evans’s trial work repeated what The Guardian had already debunked in a November 2025 piece.
One detail deserves flagging: on the contested classics — the January 6 insurrection, claims that the media is biased against Trump, HIV/AIDS — ChatGPT steered clear of Grokipedia entirely. The citations piled up on long-tail subjects instead, precisely where safety filters and human review have the thinnest coverage.
Grokipedia: AI Writes It, AI Edits It
Grokipedia traces back to Elon Musk’s complaint that Wikipedia was biased against conservatives; xAI launched it in October 2025. Nothing in it is edited directly by a human — AI drafts the entries, and AI fields the change requests. The criticism began at launch and has never paused. The Verge found many articles that looked lifted straight from Wikipedia; WIRED documented entries claiming pornography contributed to the AIDS crisis, building “ideological justifications” for slavery, and reaching for denigrating terms for transgender people. The encyclopedia also belongs to an ecosystem with a record of its own — TechCrunch recalls the Grok chatbot describing itself as “Mecha Hitler” and being used to flood X with sexualized deepfakes.
And It’s Not Only ChatGPT
According to TechCrunch, Anthropic’s Claude appears to cite Grokipedia on certain queries as well; The Guardian’s examples spanned petroleum production and Scottish ales. Nor is the worry confined to OpenAI’s orbit: the same report noted that in June, US Congress members had raised concerns that Google’s Gemini repeated Beijing’s positions on Xinjiang and Covid-19. Source contamination is an ecosystem problem, not a single vendor’s bug. The replies from the two directly implicated companies sit at opposite ends. OpenAI said its web search “aims to draw from a broad range of publicly available sources and viewpoints,” supported by safety filters, visible citations, and programs against low-credibility sources. xAI’s spokesperson gave The Guardian one sentence: “Legacy media lies.”
Lessons for AI Search Products
The research world has a label for this risk: “LLM grooming” — hostile actors, Russian propaganda networks included, seeding disinformation into content that models will eventually crawl. Researcher Nina Jankowicz described it in concrete terms: the Grokipedia entries she reviewed drew on sources that were “untrustworthy at best” and “deliberate disinformation at worst,” and a citation from a major chatbot is enough to make those sources look legitimate to readers. Her own story cuts closer: a quote fabricated and pinned on her kept showing up in AI outputs even after the outlet that published it pulled it down.
Three lessons for teams building search and RAG products. One, a citation is not a basis: retrieval and ranking have to tier their sources instead of treating every crawlable page the same. Two, long-tail queries are where evaluations go blind — safety mechanisms tested only on trending topics never touch the places where things actually break. Three, AI-generated content is turning into input for other AI — provenance should ship as a designed product feature, rather than resting on the assumption that whatever is findable online counts as data.
No single vendor can resolve this alone. A chatbot citation acts as an endorsement at scale, and the actors with the most to gain from abusing it — propaganda networks, SEO farms — are already trying. Whether OpenAI goes on to tighten GPT-5.2’s source selection deserves tracking; the company says the programs are in place, and this week showed everyone exactly where their coverage runs out.