On January 24, 2026, The Guardian published the results of a hands-on test by reporter Aisha Down: GPT-5.2, OpenAI’s latest model, cited Grokipedia — the AI-generated encyclopedia run by xAI — nine times across answers to more than a dozen different questions. TechCrunch followed the next day with its own framing: Grokipedia’s content is “escaping containment” from the Musk ecosystem and showing up in competitors’ products.
The real issue is not that a rival’s product got cited. It is what came out after the citation: on some topics GPT-5.2 made stronger claims than Wikipedia, and repeated details The Guardian had already debunked. When chatbots start treating an AI-written encyclopedia as a source, source quality stops being one vendor’s problem and becomes an ecosystem-level one. GPT-5.2 is the flagship of OpenAI’s post-New-Year model acceleration — see our 2026 opening outlook for background.
What the Tests Found
In The Guardian’s testing, the citations clustered on obscure ground: Iranian political structures (Basij paramilitary salaries, ownership of the Mostazafan Foundation) and the biography of British historian Sir Richard Evans, who served as an expert witness against Holocaust denier David Irving. Citing Grokipedia, ChatGPT made stronger claims than Wikipedia about MTN-Irancell’s ties to the supreme leader’s office, and repeated details about Evans’s trial work that The Guardian had debunked in a November 2025 piece.
One detail deserves emphasis: on the hot-button topics — the January 6 insurrection, claims of media bias against Trump, HIV/AIDS — ChatGPT did not cite Grokipedia at all. The citations landed on long-tail subjects, exactly where safety filters and human review have the least coverage.
Grokipedia: An Encyclopedia Written and Edited by AI
xAI launched Grokipedia in October 2025, after Elon Musk complained that Wikipedia was biased against conservatives. There is no direct human editing: an AI writes the entries, and an AI handles change requests. The criticism started early and has not stopped. The Verge found many articles appeared copied directly from Wikipedia; WIRED reported entries claiming pornography contributed to the AIDS crisis, offering “ideological justifications” for slavery, and using denigrating terms for transgender people. The encyclopedia also sits inside an ecosystem with its own incident history — TechCrunch recalls that the Grok chatbot once described itself as “Mecha Hitler” and was used to flood X with sexualized deepfakes.
Not Just ChatGPT
TechCrunch notes that Anthropic’s Claude also appears to cite Grokipedia on some queries; The Guardian’s examples included petroleum production and Scottish ales. Nor is this only an OpenAI-adjacent worry: the paper noted that in June, US Congress members raised concerns that Google’s Gemini repeated Beijing’s positions on Xinjiang and Covid-19. Source contamination is an ecosystem problem, not one vendor’s bug. The two directly implicated companies’ responses could hardly differ more. OpenAI said its web search “aims to draw from a broad range of publicly available sources and viewpoints,” backed by safety filters, visible citations, and programs against low-credibility sources. xAI’s spokesperson replied with a single sentence: “Legacy media lies.”
What It Means for AI Search Products
Researchers call the underlying risk “LLM grooming”: malign actors — including Russian propaganda networks — seeding disinformation into content that models will crawl. Researcher Nina Jankowicz put it more concretely: the Grokipedia entries she reviewed relied on sources that were “untrustworthy at best” and “deliberate disinformation at worst,” and citations from major chatbots legitimize exactly those sources. Her own experience is sharper still: a fabricated quote attributed to her kept appearing in AI outputs even after the originating outlet took it down.
Three takeaways for teams building search and RAG products. First, “has a citation” is not “has a basis”: retrieval and ranking need source tiers, not uniform treatment of every crawlable page. Second, long-tail queries are the evaluation blind spot — safety testing that only covers trending topics will miss where things actually break. Third, AI-generated content is becoming input for other AI: provenance needs to be designed as a product feature, not assumed away with “everything findable online counts as data.”
None of this is solvable by a single vendor. A chatbot citation is an endorsement at scale, and the actors most motivated to exploit that — propaganda networks, SEO farms — are exactly the ones already experimenting with it. Whether OpenAI tightens GPT-5.2’s source selection is worth tracking; the company says the programs exist, and this week everyone learned where their coverage ends.
Sources
- Latest ChatGPT model uses Elon Musk’s Grokipedia as source, tests reveal — The Guardian
- ChatGPT is pulling answers from Elon Musk’s Grokipedia — TechCrunch
- Where Does GPT-5.2 Get Its Information? In Some Cases, It’s Grokipedia — PCMag
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.