This page analyses the potential impacts of using generative artificial intelligence (AI) and large language models (LLMs) to create Wikipedia content.  This page also provides several recommendations for the Wikimedia Foundation (WMF) to consider as it decides what role  AI and LLMs will play in Wikipedia’s future.

• Reduced Motivation due to Perceived Diminished Value of Contributions: The introduction of generative AI tools to create Wikipedia content could reduce editor motivation. Editors may feel that their contributions are less valuable if AI can generate content more efficiently.^[1] Similar dynamics have been observed when paid labor is introduced into Wikipedia: volunteers may reduce their contributions if they perceive that paid editors are doing the work. In both cases, editors may feel that their efforts are less necessary or impactful, leading to decreased motivation.^[2]
  • Motivation is crucial for online communities because it drives participation, which is the lifeblood of these communities.^[3] Without sufficient motivation, individuals are less likely to contribute content, engage in discussions, perform tasks that benefit the community as a whole, or stay engaged with Wikipedia in the long term.^[4] Individuals are more motivated when they believe that their contributions are valuable, impactful, and meaningful to the community.^[1] If the introduction of AI content diminishes the perceived utility of editor's contributions, it will de-motivate them to participate in Wikipedia. Given that the WMF's mission includes the goal of "engaging people," this is a serious drawback to consider.
• Potential Bias in AI Content Could Erode Community Trust and Article Neutrality: A big part of Wikipedia's appeal and value is tied to the fact that it is noncommercial and strives for neutrality.^[2] This makes the site more trustworthy than other information sources that might be biased by political or financial motivations. However, generative AI models are trained on existing data, which can reflect societal biases and perspectives. If not carefully developed and moderated, AI-generated content could introduce biases into Wikipedia content, undermining the neutrality principle^[5] and eroding trust in Wikipedia content. If Wikipedia visitors feel like they can't trust the content, they will likely seek other sources for their information.
• Erosion of Community Identity: Wikipedia's culture is deeply rooted in the values of collaboration, volunteerism, and the pursuit of knowledge.^[1] Generative AI could erode the sense of shared ownership and community identity that has been essential to Wikipedia's success.^[2] If AI becomes the primary source of content, it could lead to a decline in the social aspects of Wikipedia, potentially discouraging newcomers and reducing the overall vibrancy of the community.^[6][7][8]

• Enhanced Motivation through Collaboration: Generative AI could be used to enhance collaboration between human editors and AI tools. For example, SuggestBot is a tool that directs editors toward work that matches their interests and competence.^[9] Generative AI could be used to develop similar tools that make it easier for editors to find tasks that are both meaningful and aligned with their skills. This could lead to increased editor satisfaction and motivation.
• Improved Experience for Newcomers: Generative AI could be used to develop tools and resources that make it easier for newcomers to learn the ropes and contribute to Wikipedia. Newcomers lack the established commitment and investment that long-time members possess. They are more likely to leave in response to even minor negative experiences. Providing a welcoming and supportive environment from the outset is essential to foster their sense of belonging and encourage continued participation.^[10][1] This is important because Wikipedia currently struggles to attract and retain new editors.^[11]
  • Personalized Tutorials: AI could be used to generate personalized tutorials or to identify tasks that are well-suited for newcomers. Research has shown that it is important to provide newcomers with clear and accessible guidance on Wikipedia's policies, norms, and editing practices.^{[1][6][7][11]}
    • AI could analyze a newcomer's initial edits, interests, and interactions to identify areas where they might need additional support. This personalized approach could make learning more efficient and engaging, leading to improved retention and contribution.
    • For instance, if a newcomer's edits consistently struggle with proper citation formatting, the AI could generate a tutorial specifically focused on that topic, providing targeted instruction and examples relevant to their current editing challenges. This personalized approach could address knowledge gaps more effectively than generic tutorials, leading to faster skill development and increased confidence.
  • Sophisticated and Tireless Moderation Tools: AI could also be used to develop more sophisticated moderation tools that can distinguish between good-faith edits by newcomers and vandalism or spam. This could help to create a safer and more supportive environment for newcomers.
    • Wikipedia newcomers often have negative interactions with experienced editors, including harsh criticism, reverts, and mistrust.^[6]
    • AI-powered moderation tools could help to mitigate these negative experiences by:
      • Identifying and flagging potentially harmful or unconstructive feedback directed at newcomers, giving human moderators an opportunity to intervene and promote a more supportive tone.
      • Distinguishing between good-faith errors made by newcomers and deliberate vandalism or spam, preventing the unnecessary rejection of well-intentioned contributions and reducing the likelihood of newcomers feeling discouraged.

• The WikiMedia Foundation already uses a machine learning platform called ORES (Objective Revision Evaluation Service) to predict the quality and good faith of edits.^[6] Building on these existing efforts, AI could further enhance moderation accuracy and responsiveness, creating a safer and more welcoming space for newcomers to learn and contribute.

• The Wikipedia Teahouse is a successful example of an intervention that was designed to support newcomers. The Teahouse provides a safe space for newcomers to ask questions, get help from experienced editors, and learn about Wikipedia's norms and policies.^[7] Research has shown that newcomers who are invited to the Teahouse are more likely to continue editing Wikipedia than those who are not.^[6] However, one of the teahouse's challenges is sustainability and scalability. As a volunteer-run initiative, the long-term sustainability and scalability of the Teahouse pose ongoing challenges. While the Teahouse's design aimed to minimize the technical burden on volunteers, ensuring a consistent and active pool of hosts to provide support for newcomers is crucial for its continued success.

• AI could improve newcomers' experience and help prevent teahouse host burnout by:
  • Proactive Outreach and Personalized Invitations: AI could be used to analyze newcomers’ initial edits and interests to identify those who would most benefit from the Teahouse and send them personalized invitations. For example, the AI could identify editors who:
    • Make edits that frequently get reverted.
    • Show interest in specific topic areas where the Teahouse has active hosts.
    • Exhibit editing patterns or behaviors associated with newcomers who have previously benefited from the Teahouse.
  • Identifying struggling newcomers in real-time: AI could build upon existing tools like ORES (Objective Revision Evaluation Service) to identify newcomers struggling with policy compliance or receiving negative feedback and proactively invite them to the Teahouse. This could prevent newcomers from becoming discouraged and leaving before they are aware of available support.
  • AI-powered Question Routing and Matching: AI could analyze the content of newcomer questions and match them with other newcomers who have previously asked or answered similar questions. This could:
    • Facilitate peer-to-peer learning and support.
    • Reduce the workload on Teahouse hosts.
    • Encourage a greater sense of community among newcomers
  • Development of AI-powered Chatbots or Virtual Assistants: AI could be used to create chatbots or virtual assistants capable of answering basic newcomer questions and providing guidance on common tasks. This could:
    • Provide 24/7 support for newcomers.
    • Reduce the workload on human volunteers.
    • Free up hosts to focus on more complex issues requiring human interaction and expertise
  • AI-assisted Host Training and Onboarding (because hosts need help too!): AI could analyze data on successful host interactions and best practices to develop training materials and resources for new Teahouse hosts. This could help ensure a consistent level of quality support and mentorship for newcomers, even as the pool of volunteers evolves over time.

• Increased Content Production and Scope: Generative AI could significantly increase the size and scope of Wikipedia by automating the process of content creation. While the use of bots for moderation is widespread across online communities like Reddit, some editions of Wikipedia are already using bots to generate articles. This could lead to the creation of more articles on a wider range of topics, potentially expanding Wikipedia's coverage to include topics that have been historically underrepresented. AI could be used to generate articles in languages where Wikipedia is currently underdeveloped, potentially fostering the growth of smaller language communities and promoting knowledge equity.
  • Cross-Language Collaboration: Generative AI could be used to facilitate collaboration between different language communities, for example, by providing high-quality machine translations of articles. This could help to reduce the knowledge gap between larger and smaller language communities^[2].

Given the potential drawbacks and benefits, as the WMF moves forward with AI and LLMs, I recommend they:

• Start with a pilot program. Introduce generative AI tools to a small group of experienced Wikipedia editors first. This would allow the WMF to get feedback on the tools and make any necessary changes before releasing them to the wider community. The pilot program would be similar to the Teahouse, a project designed to help new Wikipedia editors, and the pilot program for TWA, the interactive tutorial for new editors. This approach would also allow the WMF to monitor the impact of generative AI tools on the quality of Wikipedia content. The goal is to ensure that generative AI tools are used to improve the quality of Wikipedia articles, not to replace human editors.
• Are mindful of the potential for bias. Generative AI tools are trained on large datasets of text and code. If these datasets are biased, the AI tools will also be biased. The WMF should take steps to mitigate the potential for bias in generative AI tools, such as using diverse datasets and carefully evaluating the output of the tools. For example, if a generative AI tool is only trained on articles written by men, it is more likely to produce articles that reflect a male perspective. This could lead to a decrease in the quality of Wikipedia articles and an exacerbation of the existing gender bias on Wikipedia.

• Encourage collaboration between human editors and generative AI tools. Generative AI tools can be used to augment the work of human editors, not to replace them. For example, generative AI tools can be used to generate lists of potential sources for articles, or to summarize complex topics, or to help create a consistent "encyclopedic tone" in articles. Human editors would then need to review and verify the information generated by the AI tools. This collaboration can help to improve the quantity and quality of Wikipedia articles and help make the editing process more efficient. Encouraging collaboration is especially important because the WMF's mission includes engaging people in producing educational content.

• Develop clear guidelines for using generative AI tools. The guidelines should specify what types of content are appropriate for generative AI tools to create and how the tools should be used. For example, generative AI tools could be used to create first drafts of articles on non-controversial topics. Editors would then need to review and edit the drafts before they are published. Without clear guidelines, editors may use generative AI in ways that violate the norms of the Wikipedia community. This is similar to the need for policies and rules to prevent inexperienced Wikipedia editors from making well-intentioned but harmful edits.
• Provide training and support for editors using generative AI tools. If not used responsibly, generative AI could increase factual errors, biases, or plagiarism on Wikipedia. Training can help editors understand the limitations of AI tools and the importance of carefully reviewing and verifying all AI-generated content. It can also empower editors to confidently use AI tools while understanding their role in the content creation process. Editors can learn to use AI as a supportive tool rather than a replacement for human judgment and expertise.
• Be transparent about the use of generative AI tools. The WMF should be open about how generative AI tools are being used to create Wikipedia content. This transparency will help to build trust with the Wikipedia community. The WMF should make it clear which articles have been created or edited using generative AI tools.