As AI gains space as the first option for Europeans looking for information online, open knowledge platforms like Wikipedia are feeling the impact. Over the past year, the site has seen an estimated eight per cent drop in traffic, according to its own monitoring. This decline coincides with the rollout of AI models, such as search summaries and chatbots.
For a project that relies on a volunteer community and public donations, the shift raises existential questions. If fewer people visit Wikipedia, while AI systems extract and repurpose its content at scale, the open ecosystem that has sustained the world’s largest collaborative knowledge project for over two decades could be placed under real strain.
In Brussels, EU Perspectives spoke with Dimitar Zagorski, Policy Director at Wikimedia Europe, about how the organisation is responding to this moment of uncertainty, from the surge in AI scraping and the financial pressures it creates, to the risks of “weaponised AI,” and what it will take to keep Wikipedia a human-driven project.
Sustainability under threat?
Wikipedia’s traffic is reportedly down by around 8 per cent. What does this mean for long-term sustainability and for the volunteer community?
We’ve experienced fewer human users coming to Wikipedia over the past year. At the same time, we’ve seen an increase in traffic from crawlers and bots, which is an infrastructure challenge. Our servers are spread across the US and places like Amsterdam, Marseille, Singapore, and Brazil. With human readers, traffic is quite predictable, for example, during a football match in a certain country, articles about those clubs spike and get cached on the closest server, which is cheap and fast to deliver. Crawlers, especially those used by AI developers, behave totally differently. They open thousands of unrelated articles in seconds. Because this traffic is unpredictable, we can’t cache it locally, so we have to serve it globally across data centres.
Also, if fewer humans come to Wikipedia, two problems follow. First, fewer people will start editing, and we rely entirely on volunteer editors. Second, over 90 per cent of our income comes from fundraising banners. If fewer people visit the website, fewer see the banners. So we need to account for a potential drop in fundraising revenue. We haven’t seen the drop yet—our main fundraiser is in December—but the trend is not going in the direction we’d like.
You might be interested
AI crawlers
Have you considered blocking AI crawlers?
We really dislike the idea of blocking. Our philosophy from the beginning was to be free for everyone: commercial, non-commercial, humans and machines. Blocking is a last resort. Until recently, the only policy was basically a note in robots.txt saying: ‘Please don’t melt our servers’. That was literally it. When necessary, we can limit the number of requests from specific IPs or crawlers.
There is also the opt-out mechanism has not always been respected. Some companies simply ignore it. These companies are in a major race with billions at stake. My feeling is that their thinking is: in 10 years, only two or three will survive and become the new Google or Facebook. They take the legal risk now. The damage is also already done. Once the base model is trained, future legislation can’t retroactively change that.
Last time there was a copyright reform, I was all over it. I think copyright will be reopened and changed, but this will happen only in the next legislative term of the Commission. This might propose small fixes through an omnibus, but there won’t be a full-blown copyright reform now.
Our philosophy from the beginning was to be free for everyone: commercial, non-commercial, humans and machines. Blocking is a last resort. – Dimitar Zagorski, Policy Director at Wikimedia Europe
Handling large volumes
How are you adapting to handle these very large volume requests from crawlers?
We’re experimenting with Wikimedia Enterprise. This is uncomfortable for us because we’re proudly a nonprofit, but Wikimedia Enterprise is a 100 per cent Wikimedia-owned for-profit organisation. What it does is provide direct access to our servers and an API specifically built for very large users. It sells the speed and the access, not the content.
For them, the advantage is high speed, guaranteed uptime, and getting the data in a format they want. Otherwise, if they use our public APIs and melt our servers, we cut them off. With Enterprise, they get a perfect copy of the projects on a server they can access fast. We’re in talks with basically every big company you can think of. Many have already signed up, but most require nondisclosure agreements.
Often they asked why they should pay if the content is open. And we tell them that yes, the content is open, but our servers cost money. I don’t want the people who donate 5€ every year to Wikipedia to pay for delivering content to AI models.
Another interesting thing is that LLM developers ask us to please not include or limit AI content on our projects. We need the human, organic content — otherwise our development breaks.’ That’s something where suddenly there is leverage that projects like Wikipedia have.
AI summaries
How significant has the impact of AI summaries from Google been on Wikipedia traffic?
We believe the drop in human traffic is because people, when they search for something, get the answer directly — whether from ChatGPT or from Google. I consider it the same category. It’s an AI model that tries to give you a quick answer without clicking in any site.
What we really want is proper citations and inline citations. For us, it’s not enough. Some AI models are doing it better than others. Whenever they have a statement, they cite the source immediately and link it so you can check it. That’s what we want to see a lot more of.
How is Wikipedia dealing with AI-generated content?
Wikipedia relies on volunteer moderators, and its communities actively discuss and set rules for AI use. The German and English Wikipedias already ban undisclosed AI-generated content. When editors find such content, they delete it, and moderators may ban the user.
We’re also worried about weaponised AI. We ran an experiment asking OpenAI and Claude to subtly shift the narrative in the article about the Russian invasion of Ukraine. OpenAI did an incredibly good job. No explicit lies, just a restructuring that made it read like ‘it’s unclear who is to blame.’ Claude refused at first, but after a simple jailbreak prompt, it also produced the altered version. This kind of subtle manipulation at scale is something we have to watch carefully.
AI systems can’t exist on synthetic content alone. They need constant, organic human input. Our response is to lean into that: we are the human encyclopedia. – Dimitar Zagorski, Policy Director at Wikimedia Europe
Positive AI uses
Are there also positive uses of AI for Wikimedia?
Yes. Especially for small languages. Wikidata is already designed so humans and machines can use it. Wikimedia Deutschland has added grammatical features for under-resourced languages, and with some AI-assisted tools, machine translation has improved dramatically for them.
We also use machine-learning tools to support volunteer patrollers. These tools give a probability score on whether an edit is vandalism or productive.
But we have to be careful. In earlier experiments, we were using AI to guide editors on how to work on the pages, and AI suggestions showed bias. For example, for men, it proposed adding information about careers; for women, about personal life. AI amplifies existing biases, so we’re very conservative.
Do you think open projects like Wikipedia are in danger in the AI boom?
We hope we can survive. That’s why we’re experimenting with things like Wikimedia Enterprise, to make sure enough value returns to the commons. Our biggest worry is the ecosystem. Wikipedia is not original research. We rely on journalism, academia, and open research. If the media ecosystem collapses, Wikipedia collapses with it.
At the same time, AI systems can’t exist on synthetic content alone. They need constant, organic human input. Our response is to lean into that: we are the human encyclopedia.
But the broader question of money distribution won’t go away. Big organisations can negotiate deals. Smaller ones can’t. Maybe open-knowledge projects like Wikimedia, OpenStreetMap, Gutenberg need to team up.
