Wikipedia and Grokipedia: A Comparison of Human and Generative Encyclopedias

Problem Statement
While Large Language Model (LLM) based systems are increasingly used for search, summarization, and automated text production, their specific role in mediating encyclopedic content remains under-explored. Existing research has extensively examined Wikipedia as a collaborative knowledge system, focusing on editorial conflict, polarization, and governance. Conversely, research on generative systems has largely focused on hallucinations, bias, and factual accuracy. A significant gap exists in understanding how generative systems interact with established reference knowledge infrastructures. Specifically, it is unclear how generative mediation alters the structural organization of encyclopedic content, including how content is selected, rewritten, stabilized, and framed when drawn from human-curated sources like Wikipedia. This study addresses this gap by systematically comparing Wikipedia and Grokipedia, a generative encyclopedia built on top of Wikipedia, to determine how generative mediation reshapes reference content beyond individual outputs.

Methodology
The authors conducted a systematic, page-level comparison between Wikipedia and Grokipedia across three topical domains: U.S. politics, geopolitics, and conspiracy-related content. The analysis focused on four complementary dimensions:

1. Content Selection and Rewriting: The study modeled page inclusion in Grokipedia as a function of Wikipedia page popularity (views), reference density, and editorial activity (edits and reverts). Logistic regression was used to determine the probability of a Wikipedia page appearing in Grokipedia and the probability of that included page being generatively rewritten. Features were discretized into ordinal levels (Low, Mid, High, Very High) to handle heavy-tailed distributions.
2. Editorial Participation and Complexity: To assess changes in editorial dynamics, the authors analyzed editing activity on both platforms during a specific window (late October to November 2025). They applied a fitness–complexity framework (adapted from economic product analysis) to the editor–page bipartite matrix. In this model, editors are treated as "countries" and pages as "products," allowing for the estimation of page complexity based on the "fitness" (breadth and experience) of the contributing editors.
3. Narrative Structure: Narrative organization was reconstructed using Abstract Meaning Representation (AMR) to extract actor–relation networks from article texts. Entities served as nodes, and directed edges encoded relations (agent–target) with assigned polarities (supportive, conflictive, neutral). This allowed for the comparison of global narrative structures and local sentiment balances (outgoing and incoming) between the two sources.
4. Evaluative Framing: The authors analyzed the "lead section" of articles (the introductory text before the first heading) to measure framing along two dimensions: laudatory framing (praise/admiration) and conflict/controversy framing (disputes). An LLM (gemma3:4b) was used to classify sentences within lead sections, generating scores for the fraction of text exhibiting these framings.

Key Results

• Non-Uniform Selection: Grokipedia does not sample Wikipedia uniformly. Inclusion is positively correlated with page popularity, reference density, and editorial activity. Pages with higher visibility and recent editorial conflict are more likely to be included.
• Rewriting Patterns: Among included pages, the probability of generative rewriting is positively associated with reference density and recent editorial activity (edits and reverts). Conversely, highly popular pages are more likely to be reproduced verbatim. This suggests that Grokipedia intervenes more frequently on pages characterized by instability and disagreement.
• Editorial Dynamics: While both platforms show event-driven editing spikes, Grokipedia editors concentrated activity on broader, enduring topics (e.g., Elon Musk, historical figures) rather than the specific news events that drove Wikipedia editing. The fitness–complexity analysis revealed a weak positive correlation in page complexity between the two platforms, indicating that the composition of contributor communities differs significantly for the same topics.
• Narrative Stability: Despite generative rewriting, the global narrative structure remains largely consistent. Actor–relation networks show similar levels of reciprocity and cyclic structure, and the ranking of prominent entities is consistent across sources.
• Framing Shifts: While overall framing is correlated, localized shifts occur. Grokipedia exhibits more laudatory framing and a reduced emphasis on conflict for specific salient topics, particularly within U.S. politics (e.g., Donald Trump, Marjorie Taylor Greene) and conspiracy-related content.

Key Contributions
The paper provides a systematic empirical comparison of reference and generative encyclopedic systems, moving beyond isolated factual errors to analyze structural properties. It demonstrates that generative systems preserve the core narrative organization of existing reference content while significantly altering:

1. Selection: Prioritizing high-visibility and high-conflict pages.
2. Rewriting: Intervening more on pages with high reference density and editorial instability.
3. Framing: Introducing localized shifts toward laudatory and less conflict-oriented language for specific actors.

Significance
The study claims that generative encyclopedic systems should not be viewed solely as text generators but as mediating layers that reorganize existing knowledge infrastructures. The findings suggest that evaluations based solely on accuracy or ideological bias are insufficient. Instead, the impact of generative systems must be understood through the lens of content selection, the consolidation of contested material through rewriting, and alterations in evaluative framing at the point of presentation. The authors conclude that while the structural backbone of encyclopedic knowledge is preserved, the generative mediation layer fundamentally changes how content is selected, stabilized, and presented to the user.