Wikipedia has been the first stop for those looking for quick explainers for nearly a quarter-century.
Since 2022, however, that instinct has been changing. People are increasingly skipping the online encyclopedia entirely and going straight to ChatGPT for fast, conversational answers.
A new study from Columbia Business School, MIT, and Dartmouth offers the first large-scale evidence of that change. While overall Wikipedia traffic remains steady, the researchers find that specific types of articles—those whose content closely resembles what ChatGPT would generate—have seen a noticeable drop in readership since ChatGPT’s launch. Editing activity on these articles may also be declining, though the evidence there is less conclusive.
“People are now choosing which articles to use Wikipedia for over LLMs and vice versa,” says Hannah Li, an Assistant Professor of Business in CBS’s Decision, Risk, and Operations Division and one of the paper’s authors.
How the Research Was Done
To measure ChatGPT’s impact on Wikipedia, the researchers first compiled a large set of Wikipedia articles and prompted ChatGPT to generate responses on those same topics. Then, using machine learning, they classified each article based on how closely its content aligned with ChatGPT’s output, labeling them as either “similar” or “dissimilar.”
The goal was to see whether readership and editing behaviors changed differently across these two groups before and after ChatGPT's public debut. By comparing trends in similar versus dissimilar articles over time, the researchers were able to isolate ChatGPT's impact on Wikipedia traffic and contributions.
What the Researchers Found
The most striking result of the researchers’ findings is that ChatGPT appears to be cannibalizing attention from Wikipedia for certain topics. Readership declined noticeably for articles that closely mirrored what ChatGPT would produce in response to the same topic.
The effect was less clear, though still suggestive, regarding editing activity. There was a mild decline in contributions to similar articles, but the data was noisier and not statistically significant. Still, it raises concerns about Wikipedia’s long-term vitality—not just as a repository of knowledge, but as a living, crowdsourced platform.
“Wikipedia is known for its long tail of niche topics that are hard to find anywhere else,” Li says. “If people stop contributing, the knowledge base weakens and so do the AIs trained on it.”
The overall pattern points to what the authors call a “heterogeneous effect”: not a general decline across the board, but a selective shift in attention based on the strengths and weaknesses of LLMs.
Why the Research Matters
Wikipedia is the backbone of much of the internet’s factual infrastructure. It’s also a primary training source for large language models like ChatGPT. If usage and contributions decline, the quality and breadth of Wikipedia may suffer. And that could, paradoxically, degrade future generations of the LLMs that are drawing people away from it.
The study also raises more profound questions about human-AI behavior. Most AI research focuses on algorithms, but this paper flips the lens: what happens when people change their behavior in response to AI?
“In education, health care, and business,” Li says, “we’re seeing these hybrid decision-making processes emerge. AI flags a risk or suggests a solution, and a human makes the final call. Understanding how people interact with these tools is going to be essential.”
