OpenAI's latest language model, GPT-5.2, has sparked discussions regarding the sources it uses for training and knowledge acquisition. A recent report highlights the AI's reliance on Grokipedia, an AI-generated online encyclopedia launched in October 2025 by xAI. This reliance raises concerns due to Grokipedia's accuracy and potential biases.
GPT-5.2, released in early December 2025, is designed for professional knowledge work. OpenAI claims it saves ChatGPT Enterprise users significant time daily. It shows improvements in creating spreadsheets, presentations, writing code, perceiving images, understanding long contexts, and using tools. GPT-5.2 also brings improvements in general intelligence, long-context understanding and vision. It is considered better at executing complex, real-world tasks. There are different versions of GPT-5.2, including GPT-5.2 Instant, Thinking, and Pro. GPT-5.2-Codex is optimized for agentic coding in Codex, with improvements in long-horizon work, performance on large code changes, Windows environments, and cybersecurity capabilities.
Grokipedia is positioned as an alternative to Wikipedia. Elon Musk, the founder of xAI, believes Wikipedia promotes propaganda. Grokipedia's content is generated by Grok, a large language model, and forked from Wikipedia, with alterations or verbatim copies. While visitors can suggest corrections, they cannot directly edit articles.
However, Grokipedia has faced criticism regarding its accuracy and biases. External analyses have revealed that Grokipedia promotes right-wing perspectives and Musk's views. It has also been accused of validating conspiracy theories and ideas against scientific consensus, such as HIV/AIDS denialism and climate change denial. Some studies have pointed out the use of sources with very low credibility, like Twitter conversations and neo-Nazi websites. Grokipedia's reliability is considered "good-but-not-gospel," achieving 86% overall accuracy. While it performs well for quick reference and conceptual orientation, it falls short of formal research-grade rigor.
GPT-5's training data consists of massive open internet datasets, multimodal data, and synthetic data generated by earlier models. While GPT-5.1 refined training methods and strengthened data privacy, it relies on the same data foundation as GPT-5. User data is not included in the training set without consent and undergoes anonymization.
The reliance of GPT-5.2 on Grokipedia raises concerns about the potential for the model to perpetuate inaccuracies and biases present in the encyclopedia. This highlights a broader issue of AI models depending on unverified data sources. As language models play a greater role in knowledge production, there is a risk of prioritizing eloquence over accuracy and replacing transparency with automation. Ensuring provenance, bias detection, and promoting critical evaluation are crucial in the age of synthetic knowledge.
