OpenAI’s Latest Model Raises Credibility Questions by Citing AI-Generated Encyclopedia

OpenAI’s newest GPT-5.2 model has been found citing Grokipedia, an AI-generated encyclopedia backed by Elon Musk’s xAI, as a source for factual claims—triggering widespread concerns about artificial intelligence reliability and the emergence of AI-trained-on-AI content loops.

In testing conducted by The Guardian, the advanced language model referenced Grokipedia nine times across more than a dozen queries, particularly when answering questions about sensitive or obscure topics. The findings reveal a troubling pattern: GPT-5.2 pulled information from Grokipedia when discussing Iranian political structures, including details about the Basij paramilitary force and the Mostazafan Foundation, as well as biographical information about British historian Richard Evans and his role in the libel trial involving Holocaust denier David Irving.

Selective Citation Pattern Suggests Deliberate Source Routing

What makes the discovery more alarming is what GPT-5.2 didn’t cite. The model notably avoided Grokipedia when prompted about well-documented controversial topics like alleged media bias against Donald Trump, the January 6 Capitol insurrection, and HIV/AIDS misinformation. This selective pattern suggests the model may resort to less established sources primarily when mainstream alternatives are limited or absent.

OpenAI released GPT-5.2 in December 2024, marketing it as the company’s most advanced frontier model for professional work, designed to excel at tasks like creating spreadsheets and handling complex workflows. However, the model’s apparent reliance on questionable sources undermines confidence in its reliability for professional applications.

The Grokipedia Problem: AI-Generated Content with Documented Biases

Grokipedia, xAI’s counter to Wikipedia, generates content entirely through artificial intelligence rather than human verification. Researchers have documented that Grokipedia has cited neo-Nazi forums and includes “questionable” and “problematic” sources. Unlike Wikipedia’s established peer-review processes, Grokipedia offers no human oversight mechanism, creating a credibility gap that should concern users relying on AI-generated answers.

The fundamental issue extends beyond individual citation errors. When GPT-5.2 scrapes Grokipedia—which itself was generated by AI—the system creates a concerning feedback loop: AI generates content, search engines index it, ChatGPT incorporates it, and users receive AI-on-AI-generated information presented with the authority of a standard citation. This represents a departure from traditional AI training on human knowledge toward machines learning from machine-generated content.

OpenAI’s Response and Technical Gaps

OpenAI stated that GPT-5.2 searches the web for “a broad range of publicly available sources and viewpoints” while applying “safety filters to reduce the risk of surfacing links associated with high-severity harms.” The company emphasized its ongoing efforts to filter low-credibility information and influence campaigns.

However, the technical reality reveals a gap in OpenAI’s filtering systems. Since Grokipedia lacks the noindex tags that would exclude it from automated crawlers, and given its substantial volume of publicly available content, the search algorithms feeding GPT-5.2’s retrieval-augmented generation (RAG) pipeline treat it as a valid data source. Grokipedia’s size allows it to dominate search rankings, which then feeds into the model’s source selection.

Practical Concerns for Users and Professionals

For developers, researchers, and professionals relying on GPT-5.2, the discovery presents a significant credibility challenge. Users report that the model provides “strangely skewed answers” on certain topics, with the problematic information appearing authoritative due to its citation format.

Currently, no native setting allows users to exclude specific domains like Grokipedia from ChatGPT’s searches. However, community members have identified temporary workarounds: hovering over citations to verify the source URL, or using prompt engineering to instruct the model to “prioritize .edu and .gov domains” or “avoid AI-generated encyclopedias.” These methods remain imperfect and consume additional context tokens.

Broader Implications for AI Reliability

The Grokipedia citations represent the first major symptom of what researchers describe as potential ecosystem collapse in AI training. If large language models increasingly train on AI-generated content rather than human-verified information, the quality and accuracy of subsequent generations may degrade significantly. The selective nature of these citations—appearing on niche political and historical queries but absent from high-sensitivity topics where safety filters activate—suggests OpenAI’s safety mechanisms recognize certain sources as problematic yet still permit their use in less-flagged contexts.

As AI systems become central to professional decision-making and public information consumption, the distinction between human-verified sources and machine-generated content has never been more critical.

Photo by WOKANDAPIX on Pixabay