Education/News/June 2025/Does Wikipedia has future in the times of Chat-GPT
Does Wikipedia has future in the times of Chat-GPT
Points where Wikipedia wins:
1. Are language models reliable sources of information?
This question is essential. Today, many people use language models (LLMs) as quick reference tools, among many other functions. When I began exploring this topic, I wasn't entirely clear on whether there were defined criteria for evaluating the information these models offer, nor on what is Wikipedia's criteria. The question, beyond a simple comparison, invites us to question how we are currently consuming information and with what level of awareness we do so.
What is a reliable source of information? The CRAAP test
Sarah Blakeslee, a former librarian at California State University (Chico State), developed the CRAAP test in the early 2000s, a model for evaluating the reliability of information sources. CRAAP is an acronym that stands for Currency, Relevance, Authority, Accuracy, and Purpose. In other words, a reliable source must be up-to-date, relevant, come from a recognized authority, present verifiable data, and have a clear purpose that doesn't distort the information.
.png)
And LLMs? They're not exactly sources.
This is where LLMs stand out. These models aren't sources of information in and of themselves, but rather generative systems trained on large volumes of text scraped from the internet (including Wikipedia). From that training, they produce new answers, which don't necessarily replicate any specific source. They're predictive tools: they don't give us "the truth," but rather a probability of which word or phrase follows the previous one based on statistical patterns. As the European Data Protection Supervisor explains, these models are trained on billions of words from a variety of public and private sources, and their performance is measured by the number of parameters they use. But as impressive as they are from a technical perspective, their operation makes information traceability difficult.
2. Wikipedia: A Cooperative Human Network
In contrast, Wikipedia is the result of a collaborative network of people who discuss, review, and edit articles. This community—the Wikipedians—has sought since its inception to include diverse voices in terms of nationality, gender, education, social class, and ideological perspectives, in order to balance biases and offer more complete representations of knowledge. We know that this ideal is not always achieved. Actively participating in Wikipedia editing requires time, technical skills, and a certain level of subject matter expertise, which leaves many people excluded from the process. Some describe its organizational model as an "adhocracy," where those who participate the most acquire the most responsibility, but without fixed hierarchical structures. However, even with these limitations, Wikipedia remains a deliberative, transparent, and ethical endeavor in its knowledge construction. The community also operates under clearly defined standards, aligned with criteria such as those of the CRAAP test: • Verifiability: All content must be supported by published and accessible sources. • Reliable sources: Accredited authors, recognized media outlets, and peer-reviewed works are prioritized. • Relevance: A topic is considered relevant if it has been covered significantly by independent sources. "Librarians"—experienced users elected by the community—enforce these standards.
3. LLM: Power without Traceability
In generative language models, such as those based on Transformer architectures, text generation occurs through probabilistic calculations. That is, the model predicts the next word based on statistical patterns derived from its training. This logic can produce convincing, coherent, and even useful texts, but it also introduces the phenomenon of "hallucinations," that is, false or inaccurate information presented with complete certainty. Furthermore, the lack of transparency in training sources makes any attempt at verification difficult. We often don't know if an answer is based on a scientific article, a discussion forum, or an unsubstantiated post. In many cases, not even the developers can explain exactly why the model generated a specific response. This is known as the "black box" problem, and it is one of the main ethical and technical challenges facing current generative AI. In fact, as models become more powerful, they have also become less transparent. Studies like Stanford University’s Foundation Model Transparency Index (FMTI) show that the most recent LLMs tend to offer less information about their training sources, which raises issues in terms of trust and traceability.
4. What's Behind It? Economic Interests and Different Missions
It's worth remembering that LLMs are developed by for-profit companies. Their goal is to generate revenue. Whether they do so through tools that appear to offer knowledge is another story. The promise of mass access to information can coexist with business models focused on data monetization, as other technological initiatives have already done.
Wikipedia, on the other hand, operates under a different paradigm. Its purpose is the collective construction of freely accessible encyclopedic knowledge. It does not pursue economic profit, and its sustainability depends on donations and the commitment of a voluntary community. In this sense, it represents an alternative model for information production: decentralized, open, and cooperative.
Points Where LLMs Win
It's fair to acknowledge that LLMs also have their strengths as a reference tool: 1. Speed and Accessibility: This is one of the elements we discussed with other fellows in the brainstorming exercise. Performing a search using an LLM is fast and straightforward. The model summarizes the information, responds in natural language, and allows for further exploration of the topic with new questions. Wikipedia, on the other hand, requires more effort from the user: reading, navigating through different sections, comparing sources, and understanding the logic of the article. Readers may encounter conceptual barriers that they cannot easily explore on their own.
2. Efficient Navigation: LLMs can navigate and summarize large volumes of information almost instantly. In many cases, they are able to provide a "satisfactory" answer, although not necessarily a precise one. This makes them a useful tool for initial explorations or for understanding complex concepts in a simplified way.
3. Potential for constant improvement: One of Wikipedia's major challenges is the controversial articles and edit wars initiated by some users. One point in favor of LLMs is that they can potentially be less prone to bias, as they lack human attributes such as emotions, tastes, and personal morality. However, the human engineers who program the model are not immune to bias. There is a real effort to produce increasingly filtered tools that reduce bias and improve the quality of the responses generated. Some development teams have integrated ethical protocols and auditing mechanisms, which may pave the way for more responsible AI.
Integrating LLMs into Wikipedia: Between Pressures and Tensions
Now, after reviewing the potential and limitations of both Wikipedia and language models, I wonder if, instead of competing, they couldn't coexist and offer the best of both worlds?
This isn't just a fantasy. There are already projects that point in that direction. One of them was presented on EduWiki by Jackeline Bucio García, as part of a series of experiments exploring the use of Retrieval-Augmented Generation (RAG) approaches to support collaborative editing in open knowledge environments. This pilot phase, conducted with student teachers during the first semester of 2025, explored how the use of RAG could facilitate the preparation of edits on Wikipedia: from reviewing academic sources to overcoming language barriers, including content design with a multimedia approach.
Preliminary findings indicate that this integration can not only enrich the quality of the content, but also the learning process: it allows for better understanding of sources, prioritizing information, and generating questions that deepen knowledge. AI, when used well, can be a powerful ally in learning to search, read critically, and produce meaningful knowledge.
This opens the door to a bigger question: what would happen if these conversational tools were an integral part of the Wikipedia search and query experience? What if, instead of being external, they were native to the Wikimedia ecosystem? Imagine a specialized language model, trained on the Wikipedia corpus—respecting its standards, editorial logic, and community values—that facilitates navigation, that cites specific articles, suggests connections between them, summarizes content accurately, and, above all, allows for dialogues that help readers better understand what they are consulting.
They could facilitate a type of guided, adaptive navigation, especially useful for students or casual readers who need guidance to avoid getting lost in the vastness of articles.
Of course, there are tensions. There is concern among those who believe that Wikipedia should remain an organized index of sources, a clear, verifiable, and neutral encyclopedic repository. They fear that the introduction of generative models will be in detrimental of accuracy and dilute the project's core mission.
However, Wikipedia is no stranger to automation. For years, it has integrated bots that support specific tasks, such as moderating edits in cases of vandalism, where a rapid response is needed. These automated agents don't replace the community, but they do optimize repetitive processes that previously required human time and effort. Why couldn't it integrate generative AI tools?
Let's consider whether resisting these possibilities could be a missed opportunity. Wouldn't this be a way to bring Wikipedia, not only as a source of information, but as a living and expanding knowledge project, again, to old and new internet audiences?
The conversation isn't simple. Like any innovation process, it's influenced by ethical, technical, and political decisions. However, if Wikipedia has demonstrated anything in its more than two decades of existence, it's that collective knowledge is not a utopia, but rather a possibility built day by day by committed communities. Perhaps, with the careful integration of artificial intelligence, this experience can become even more accessible, without losing its essence. AI at the service of human knowledge—and not the other way around—could make a substantial difference compared to the indiscriminate use of generative models, by giving more room to collaboration not only between humans, but also between humans and machines.