As artificial intelligence writing tools proliferate, Wikipedia editors are encountering a surge of AI-generated articles characterized by inaccuracies and fabricated references. In response, the volunteer community of Wikipedia has organized efforts to manage the influx of low-quality content, a movement described by Wikimedia Foundation product director Marshall Miller as an adaptive “immune system” response.
Miller emphasized the vigilance of Wikipedia contributors, stating, “They are committed to ensuring the content remains neutral and trustworthy. As the digital landscape evolves with new elements like AI, this reflects how the community adapts to emerging challenges and finds ways to address them.”
One tactic employed by Wikipedia editors is the swift removal of poorly constructed articles through a “speedy deletion” process. A Wikipedia reviewer supporting this rule noted being inundated with “atrocious drafts” and highlighted that expedited deletion would significantly facilitate the cleanup process, saving countless hours spent correcting the errors AI-generated content leaves behind. Another contributor pointed out that the “falsehoods and bogus references” prevalent in AI outputs consume extensive time for experienced editors to rectify.
Articles flagged for deletion typically undergo a seven-day discussion period, during which users assess the necessity of removal. The newly approved directive permits administrators to bypass these discussions when an article is evidently AI-generated and lacks prior human review. To determine this, editors will primarily look for three telltale signs:
- Text directed at the user, such as “Here is your Wikipedia article on…” or “I hope that helps!”
- Inaccurate citations, including those that inaccurately reference authors or publications.
- Non-existent citations, such as broken links, invalid ISBNs, or unverifiable DOIs.
Wikipedians are also observing various other indicators of AI-generated content. The ongoing WikiProject AI Cleanup focuses on addressing the increasing number of unsourced and poorly written AI-produced entries, compiling a catalog of common phrases and formatting traits found in content created by chatbots.
This guide involves identifying excessive usage of em dashes, which have been noted for their frequent appearance in AI writing, as well as the overreliance on specific conjunctions like “moreover,” along with promotional phrases such as “breathtaking.” The documentation suggests that contributors should also remain alert for formatting anomalies like curly quotation marks in place of straight ones.
Nevertheless, Wikipedia’s speedy deletion guidelines clarify that these characteristics alone should not be the exclusive criteria for determining AI authorship, nor should they trigger removal. Additionally, the speedy deletion policy covers more than just content produced by AI; it also allows for the removal of pages that harass subjects, contain hoaxes, or include incoherent text.
The Wikimedia Foundation, which oversees the platform but does not dictate its editorial policies, has experienced conflicts with its volunteer community regarding the integration of AI. In June, the organization suspended an initiative that featured AI-generated summaries atop articles following criticism from contributors.
Despite the diverse opinions on AI among Wikipedia editors, the Foundation supports leveraging AI tools that enhance the accuracy and quality of content.
Miller articulates this viewpoint, stating, “AI presents a double-edged sword. While it facilitates the creation of lower-quality content at higher rates, it can also serve as a valuable resource for volunteers if applied thoughtfully. For instance, the Wikimedia Foundation is already employing AI to identify edits that may involve vandalism, and its recent AI strategy includes aiding editors with tools that automate repetitive tasks, such as content translation.”
The Wikimedia Foundation is also in the process of developing a tool called Edit Check, which does not utilize AI and aims to assist new editors in adhering to the site’s guidelines. In the future, it may also mitigate the challenges posed by unverified AI-generated submissions. Currently, Edit Check can prompt users to add citations if significant portions of text lack references and can evaluate the tone to ensure neutrality.
The Foundation is also exploring the incorporation of a “Paste Check” feature, which would require users who paste large blocks of text into an article to confirm their authorship. Contributors have shared various suggestions to enhance this tool, including the proposal that those suspected of using AI should clarify how much content was generated by a chatbot.
Miller concludes, “We are attentive to our communities and their productive findings. For the time being, our focus on machine learning within editing is primarily about assisting users in making constructive edits and enabling those monitoring edits to prioritize their efforts effectively.”
