Wikipedia Italy has a problem with content produced by generative AI

For some years now, a question has plagued the global community of Wikipedia curators. What will become of the free encyclopedia? Of the portal capable of keeping in one place, all human knowledge, in every language, produced by humans for humans? What will become of all this with the power achieved by generative AI? “We have been discussing it for years, at least since ChatGpt was opened to the general public,” explains Ferdinando Traversa, 20 years old, from Bari, the new president of Wikimedia Italia, to Italian Tech.

“But lately many of us have noticed an increase in entries written with AI tools. That's why we decided to give ourselves some rules. We're discussing it, but the goal is only one: to always keep the human at the center and the human control over the final product,” he adds. The community is divided, the discussion is heated. But the foundation has decided on an approach that is “not totally closed” to AI, which is estimated to have written or contributed to writing 5% of the entries in the entire encyclopedia. On the contrary. “But we want the entries to remain written by humans, who can help each other in creating tables or in repetitive activities in the drafting,” he reasons.

The risk of collapse, or that AIs will train their own products

The caution of the Wikipedia Italia community is due to several risk factors: the hallucinations that generative AIs continue to suffer from, the risk of serious inaccuracies, the lack of sources and citations, but above all a possible short circuit in the results if the AIs began to train with articles written by themselves.

“It is no mystery that AIs have fed on Wikipedia in all languages ​​to learn how to generate texts. Our servers have suffered greatly from the continuous request for access to our pages of these generative models. But few see a related risk. Model collapse, which would cause everyone to lose, users and Wikipedia itself”. Model collapse is a phenomenon that is quite discussed in the context of AIs. It occurs when models are trained with products of the AIs themselves. They feed on their own mash. Effect? ​​The model becomes increasingly impoverished, the language becomes repetitive, less creative, imprecise, and tends to reinforce errors and incorrect information.

From Wikipedia Search to Chatbot Request: What's Missed

Traversa also sees a challenge posed by AI. “More and more people, especially younger ones, no longer do online searches, but ask Chatbots directly.” If Wikipedia wants to maintain its task of protecting, keeping together humanity’s knowledge and giving it to the reader, it will therefore have to “study new ways to help the reader himself in consultations.”

For Traversa, asking a chatbot directly means “giving up the beauty of getting lost in reading and maybe discovering a link, a reference, and reading again, understanding more, learning new things.” And the risk is that that chatbot has been trained on a dataset that contains errors, giving bad feedback. All issues that will need to be addressed over time.

Today Wikipedia is in good health. There are 3,000 members of the community active every month in Italy. In April it was read 57 million times, 396 thousand are the changes made to the entries. But if its popularity at the moment is not affected by the pressure of AI, those of politics represent a greater obstacle.

On Musk's attacks: "We are not for sale"

“We have been attacked several times by Musk. He believes that we are biased, that we contain false information about him and Tesla. He would like us to write things that are not true, at least not verified, and this has led him to attack us several times, asking Americans to stop donating to us.”

Message failed. Donations are “in line with what the foundation expected,” Traversa says. Next year Wikipedia will be 25 years old. “Our mission will not change, our way of working will not change, we are not for sale and cannot be bought like Twitter, we are the last piece of free internet in the era of the dominance of the tech giants,” Traversa concludes.