Experts in Artificial Intelligence are to collaborate on a transatlantic research project looking at the relationship between Wikipedia and generative AI.
Curating the Commons - AI, Wikipedia, and the Reconstruction of Notability will explore how the globally renowned repository of information both shapes, and is shaped by, the rise of AI and large language models (LLMs).
Bringing together scholars from the University of Exeter and the University of North Carolina, the project will consider how this dynamic represents a challenge to the integrity and trust of Wikipedia and wider knowledge on the Web, potentially reproducing and reinforcing biases and inaccuracies.
The work has been funded by a £171,000 grant from the Arts and Humanities Research Council as part of its Bridging Responsible AI Divides (BRAID) programme, and will run for the next 2 years. All of the BRAID projects establish and strengthen collaborations between US and UK researchers and address the ethical, legal and societal implications of AI.
"Wikipedia occupies a unique position at the heart of global knowledge production, frequently serving as the first port of call for online information seekers," says project lead Dr Patrick Gildersleve, Lecturer in Communications and Artificial Intelligence, and Co-Director of Exeter's Critical AI Centre. "Its articles often form the basis for the automated information summaries users encounter every day on major search engines and AI-generated tools, shaping the understanding and beliefs of millions globally."
"Concurrently, AI-generated content is increasingly being integrated into Wikipedia itself, either directly through automation or indirectly through human editors using AI-assisted tools," adds Dr Francesca Tripodi, co-investigator and founder of the Search Prompt Integrity and Learning Lab at the University of North Carolina. "The challenge this project addresses lies in understanding how this repeating, looping relationship between Wikipedia and AI models affects the quality and reliability of public knowledge."
The researchers will study the training data and outputs of AI tools such as Perplexity AI, Google's AI Overviews, and ChatGPT and look for examples where AI outputs directly or indirectly refer to Wikipedia. They will also analyse how these tools judge what and who is considered "important" or "notable", revealing how they may rely on, reproduce, or reinforce Wikipedia's own biases around gender, race, and geography.
Throughout the project, the team will work with Wikipedia editors, exploring their perceptions of AI-driven content by attending and hosting online and in-person 'edit-a-thons' to observe how editorial decisions are reached on content inclusion. To avoid placing undue pressure on the live site, a smaller-scale platform dubbed WikipedAI will be built to simulate the editing process.
