Recent research published in Proceedings of the National Academy of Sciences provides evidence that automated encyclopedias differ from human-edited platforms in both structure and political leanings. The study suggests that, rather than uniformly removing bias, these automated systems tend to favor longer, more complex narratives while introducing a rightward shift in certain topic areas. These findings raise questions about how artificial intelligence will shape public knowledge and verification of information sources.
In October 2025, xAI, an American technology company founded by Elon Musk, launched Grokipedia. The platform was introduced as the world’s first encyclopedia written by artificial intelligence. Musk has promised that the platform will correct the left-leaning bias that is said to exist in the widely used online encyclopedia Wikipedia.
Wikipedia content is written and maintained by volunteer editors. Grokipedia uses an extensive language model to generate and review content. This is a type of artificial intelligence trained on vast amounts of text to predict and generate human-like language. Visitors can suggest edits, but automated systems review and implement changes without traditional human editorial oversight.
To assess these claims, researchers from Trinity College Dublin and Dublin Institute of Technology conducted the largest academic analysis of Grokipedia since its inception. Scientists Saeed Mohammadi and Taha Yaseri set out to conduct a large-scale computational comparison to objectively map structural and ideological differences. They wanted to determine whether an automated platform could actually correct the bias of human-edited websites.
The team used computational text analysis and machine learning techniques to analyze articles on the same topic from Wikipedia and Grokipedia. They focused on the 20,000 most edited English Wikipedia pages and made sure to analyze substantive articles by excluding lists and calendar dates. The corresponding 17,790 matching articles were then downloaded from the automated platform.
The authors extracted the main text from each article pair and removed menus, sidebars, and formatting codes. They analyzed each pair for readability, vocabulary usage, and writing style. To measure reading comprehension difficulty, they used a standard formula that estimates the U.S. school grade level required to understand a text.
The team also calculated structural differences across the web pages they collected by counting the exact number of references and hyperlinks present for every 1,000 words. To measure how similar automated articles are to human-edited articles, we combined multiple similarity metrics into one test. This score allowed us to directly compare two versions of the same historical event or public figure.
Researchers have found significant gaps in how automated systems process existing content. Although many of Grokipedia’s articles closely mirror Wikipedia, a significant portion of the articles analyzed have been rewritten more extensively. Approximately 66% of the entries differ significantly in style, source, and political leaning.
These rewritten Grokipedia entries were significantly longer than the Wikipedia versions, and the automated text proved to be very difficult to read. On average, Glokipedia articles required a reading level of 14.5, compared to 10.7 for Wikipedia. The artificial intelligence platform also provides far fewer citations to support its claims, providing an average of 20 references per 1,000 words compared to Wikipedia’s 35.
To assess political bias, the scientists analyzed external websites cited as references within the article. They mapped these hyperlinks to an established dataset that assigns political leanings to news media sources based on social media sharing patterns. Overall, Glokipedia articles exhibit a similar political bent to Wikipedia articles, drawing primarily from left-wing news sources.
However, scientists have discovered notable changes within a series of highly rewritten papers. When it comes to politically and culturally sensitive topics, such as religion, history, literature, and art, Glocipedia has consistently moved toward referring to more right-leaning news sources compared to Wikipedia. This finding suggests that automated systems provide local, topic-specific adjustments rather than a complete overhaul of political bias.
“Rather than systematically ‘correcting’ Wikipedia’s alleged biases, as was claimed at the time of its initial launch, our findings suggest that AI-generated encyclopedias such as Glocipedia are selectively reshaping existing knowledge,” said Taha Yasseri, director of the Joint Center for the Sociology of Humans and Machines at Trinity College Dublin and the Dublin Institute of Technology and lead researcher on the study. “This creates a patchwork system where some content is copied while other content is reinterpreted in ways that are less transparent and difficult to scrutinize.”
Researchers are concerned about the long-term effects of relying on automated knowledge generation. “Online encyclopedias are at the heart of public knowledge,” said Saeed Mohammadi, lead author of the study and a doctoral candidate at the Center for the Sociology of Humans and Machines and the Irish Center for Research and Data Science Basic Research and Training. “They are also being used to train future generations of large-scale language models.”
“Our findings raise important questions about how public knowledge is generated, reproduced, verified and managed,” Mohammadi added. “Unlike Wikipedia, where biases are visibly contested through human editing, AI-generated systems operate largely opaquely, meaning changes in viewpoints and sources can occur without clear accountability or editorial oversight.”
“Simply put, AI generation does not remove bias; it changes how and where bias enters the system, often making it less noticeable.”
Although the study provides useful evidence of new differences between AI-generated and human-edited encyclopedic knowledge systems, the researchers acknowledge that focusing on Wikipedia’s most edited English-language pages is likely to over-represent high-profile and controversial topics.
Automated similarity metrics also assess text format and lexical consistency, but cannot verify factual accuracy or detect illusory claims. The opacity of automated platform training data limits our ability to determine exactly why these particular ideological shifts occurred. Yasseri noted that these findings point to broader societal concerns.
“Transparency, oversight and regulation are desperately needed in this area,” Yasseri said. “Our information environment is rapidly changing. We have already seen how a lack of editorial accountability on social media platforms enables the generation and circulation of misinformation and disinformation, often with devastating consequences for elections, public health, and social stability.”
“Currently, we are witnessing a large-scale black box reproduction of information by large-scale language models that is outside the scope of public scrutiny,” Yasseri continued. In future research, scientists may investigate the underlying mechanisms of how these models choose which information to cite and display. You can also explore including academic and government sources to build a more comprehensive picture of automated knowledge generation.
The study, “Selective Differences Between Glocipedia and Wikipedia Articles,” was authored by Saeedeh Mohammadi and Taha Yasseri.
