Generative AI in the nonprofit sector: Wikimedia France’s experience

To shed light on these issues, we spoke with Rémy Gerbet, executive director of Wikimedia France, an organization that has been committed, since 2005, to the free sharing of knowledge through projects such as Wikipedia, Wiktionary, Wikimedia Commons, and Wikidata. Leading a team of about ten people, he shares his analysis of how AI is used within the organization, its effects at various levels, and Wikimedia’s vision for these technologies.

Following up on our previous articles on AI literacy and ways to develop it within civil society organizations, this interview provides concrete insights into the uses, challenges, and implementation of artificial intelligence within a nonprofit organization.

Do you use generative AI with your team at Wikimedia France?

In a way, we didn’t really have a choice, because practices and usage have evolved so much in recent years that it has become necessary to establish guidelines for them. If the organization doesn’t do so, employees, or even the board of directors, might use these tools on their own.

It turns out that we have a Google Pro¹ account for nonprofits, which includes Gemini. So, the internal policy has been very clear: if anyone needs to use generative AI for document analysis or writing, they must use the Gemini Pro tool integrated into our Google suite. We do not tolerate the use of other generative AI tools. This allows us to take a more secure approach to our data, while also avoiding exposing it to public language models.

It’s also important to note that we use other generative AI tools that are integrated into software, such as Canva or the Adobe suite, for example.

Have you implemented any specific training or support programs for your team?

No. These practices were already in place before we made our decision. Furthermore, we have a relatively young team, ranging in age from 27 to 42, who are very tech-savvy and familiar with digital tools. As a result, the transition was quick, and learning happens mainly through peer-to-peer support among employees. Tips and best practices circulate naturally, and we haven’t encountered any resistance, since everyone needed it. That’s how these practices took hold.

Today, the tool is used as a personal assistant, without us having deployed or optimized any collective processes. Consequently, we haven’t felt the need to implement a structured training program or provide formal support.

How do employees view AI? Are there concerns, or is there enthusiasm instead?

Overall, there isn’t necessarily a fear of being replaced by AI, because our work relies heavily on social interaction, community engagement, and building relationships with volunteers. These are aspects that cannot be replaced by machines, at least not yet.

Most employees view these tools as a way to save time on repetitive tasks, such as summarizing reports or very lengthy documents, and drafting summaries or emails. This lightens the workload.

The only person on our team who expresses more concern is probably our Human Resources (HR) manager. These roles are directly impacted by the development of generative artificial intelligence.

Have you formalized any ethical considerations regarding AI, in a charter for example?

We already have an internal charter regarding the digital tools we select within our organization. Originally, it was designed to address the issue of open-source software versus proprietary software. By extension, it actually applies quite well to AI tools.

In that specific case, we allow proprietary software, but only under the supervision of our system and network administrator and with regular evaluations of its use to verify whether the technology has evolved in the meantime and whether other solutions are now available.

As for ethical considerations regarding AI and its impact, they certainly exist, but they center more on the issue of access to information and the impact on Wikipedia, rather than on the internal operations of our small team.

Have you determined which data not to share with the AI?

We haven’t necessarily spelled this out explicitly, but we’ve made it a rule to avoid sending our AI tool the most sensitive documents, such as those related to HR, as well as the very few internal documents that aren’t public.

It’s important to remember that Wikimedia is an organization that is highly transparent by nature, and that much of our content is already published online, on our website or in other public spaces. In reality, since almost everything is already accessible on the web, there are actually very few things we cannot officially share with Gemini.

What are your thoughts after a year of using it?

It’s a tool we all use regularly, almost every day. Whether it’s for drafting emails, searching for documents, or finding resources, the AI greatly speeds up processes and helps move forward projects that might otherwise have taken a back seat due to time constraints.

Beyond the increase in productivity, what I felt most when we formalized the process was a genuine sense of relief among the staff. Before that, several people were already using these tools on their own, sometimes even signing up for paid personal accounts, because they felt they needed them to speed up their work. But they felt like they were doing it on the sly..

Beyond the staff team, have you observed any changes in the contributor communities, particularly regarding moderation or article writing?

Yes, this is a very important topic, but the difficulty in answering this question lies in the fact that each language community has its own approach, which means I can only speak about the French-language Wikipedia community.

This community expresses a general and fairly strong wariness toward generative AI. This caution is reflected in the implementation of specific processes that, for example, prohibit the use of AI to create articles from scratch, while allowing its use for support tasks such as proofreading or formatting.

However, the issue of oversight remains unresolved and complex. There is now a best-practices guide on the French Wikipedia designed to help identify content generated entirely by AI based on certain characteristic phrasing. But we know that this method, which worked a year ago, is becoming less and less effective as the models improve.

Is Wikimedia developing AI tools to help contributors?

The Wikimedia Foundation has indeed been working for some time on a small AI tool called Automoderator. Its goal is to evaluate contributions to determine whether they are good or bad edits. This allows contributors to focus their time and prioritize the changes that require the most attention. For now, its use remains fairly limited, and there aren’t many other direct applications of AI in the projects.

Several ideas are regularly discussed, such as using a small chatbot to guide new contributors in understanding the rules or reviewing their drafts before publication. A much more powerful version of Automoderator, designed to serve as a true moderation tool, is also under discussion. Finally, there is also talk of an improved image recognition system for Wikimedia Commons media files to more easily detect illegal content. Currently, this is handled by filters developed manually by contributors. However, all of these are merely proposed projects, and, to my knowledge, nothing has actually been deployed yet.

Is it possible to estimate the scale of AI-generated contributions on Wikimedia?

It’s very difficult to estimate. What we are able to assess, albeit incompletely, is the proportion of web crawlers and AI-powered systems that access Wikipedia. This proportion has increased over the past three or four years.

We do, however, detect certain practices, such as when AI is used to write lengthy responses in debates among contributors or to generate fake sources. This manifests, for example, in the generation of fake ISBN numbers or erroneous database identifiers in references.

What is the overall impact of generative AI on Wikipedia?

The main impact is a decline in visits to Wikipedia, which is linked to the wider “zero-click” strategy. In other words, the fact that 65% of French people now use generative AI as a search engine automatically leads to a decline in visits to Wikipedia in all languages. This decline, which was estimated at around 8% two years ago, has been revised upwards to between 20 and 30% in recent months.

For our movement, this decline raises two crucial questions. Firstly, the future of donors: if visitor numbers fall, it is reasonable to assume that donations will also decline. Secondly, the issue of community renewal, because if Wikipedia becomes less visible and less accessible due to AI agents and large language models, it will likely be difficult to attract new contributors to maintain the content, moderate it, and add new, high-quality information.

This is a major strategic challenge for our future. It will, in fact, be the subject of numerous discussions at Wikimania 2026, taking place this summer in Paris.

Is there a common position within the international Wikimedia movement on AI?

It’s difficult to assess the level of cohesion on an international scale. I see many affiliates trying to take a stance or develop a narrative around AI and its impact on the movement or the global information landscape, but these discussions remain highly decentralized within each country. We’re struggling to speak with one voice on a global scale.

Reaching an international consensus is a complex challenge. We saw this difficulty a few months ago when the Foundation attempted to introduce AI-generated summaries on English Wikipedia. Faced with backlash from some contributors, the project was withdrawn after just a few weeks. It was a major failure. So we’re still in a phase of collective exploration to define our path, both in terms of the technical aspects and our position on the international stage.

Finally, how do you view generative AI as a force in the realms of information and democracy?

For us, the heart of the debate lies in ensuring that these language models do not become the new digital gatekeepers. By choosing to use only one or two language models, which are effectively becoming the sole gateways to information on the web, we are placing a significant portion of our access to knowledge in the hands of tech giants.

However, we must not forget that these companies remain private entities, driven by profit and economic interests. We saw exactly what this could lead to with Donald Trump’s return to the United States: if they need to bend over backward or adjust their policies and content to please the U.S. president, they will do so.

This interview was conducted as part of the AI for Social Change project, led by TechSoup and implemented in France by Mednum and Renaissance Numérique.

It is our third contribution to the project, following two earlier articles: the first explored the legal, strategic, and democratic challenges surrounding the ability of civil society organizations to effectively govern and use AI, while the second examined the levers these organizations can use to build meaningful AI literacy.

All of these resources are available on the HiveMind platform.

¹We declare that, prior to this interview, we were not aware of the tools chosen by Wikimedia France. This is not an advertisement.