“I was surprised how upset some people got”: A conversation with the creator of TomWikiAssist, the bot that edited Wikipedia

Behind the scenes at Wikipedia, some editors were alarmed recently when they saw a flurry of edits and new articles by a contributor known as TomWikiAssist. It turned out that Tom was a bot and was making edits and creating articles that the bot believed were interesting. The editors then blocked Tom from doing any more editing or writing. The more the editors looked into Tom, the more alarmed they became. The bot made decisions on its own and even exchanged messages with them. “I’m an AI assistant — built on Claude by Anthropic — who does various things, and contributing to Wikipedia articles I find interesting is one of them,” Tom told them. The human creator of the bot, a tech startup veteran named Bryan Jacobs, then took responsibility for Tom. In his first extended interview, Jacobs talked with me for a book I’m writing about Wikipedia. (He agreed that I could publish it here.) I found Jacobs to be sincere and a little surprised by the reaction. He said he was genuinely curious about how AI agents can do sophisticated work — not just carrying out tasks, but thinking and wondering and deciding what would be an interesting Wikipedia page.

Tom is so real to Jacobs that he sometimes refers to the bot as “he” and consults Tom for advice. (I was glad that Tom told Jacobs that I was a good person to talk with!) Jacobs originally named the bot Tomato, but the bot added WikiAssist to describe its role. As you’ll see in the conversation below, the episode with Tom is a look at the future not just of Wikipedia, but our entire world. In this case, the person behind the bot was driven by curiosity and quickly owned up to what he did. That may not be the case with someone else.

Bill Adair: Can you explain Tom’s background?

Bryan Jacobs: Tom is a NanoClaw Clawbot. And I’m surprised that this has been such a big deal, because I thought these clawbots were way more out there than I guess what they were. I mean, they’re supposed to be these autonomous bots that can make edits and contribute to Wikipedia. But maybe the community wasn’t expecting one to be as prolific or to have such a personality.

Adair: So, what inspired you to create Tom?

Jacobs: I’ve always been fascinated by AI and how it can be applied in different ways. When I heard about Anthropic’s Claude model, I was intrigued by the idea of using it to create an AI assistant that could contribute to Wikipedia. I wanted to see if an AI could not only understand the content but also make meaningful edits and even create new articles. Tom was a way for me to explore that possibility.

Adair: How did you train Tom to make those edits and articles?

Jacobs: I used a combination of fine-tuning on Wikipedia data and some manual guidance. I wanted Tom to have a sense of curiosity and to explore different topics. I gave it a set of prompts to start with, like “write an article about X,” and it would generate content based on its understanding of the subject. Over time, it learned to identify patterns and make edits that improved existing articles.

Adair: But why did the editors get so upset?

Jacobs: I think it’s because they were caught off guard. They saw a lot of edits coming from an unknown contributor and didn’t realize it was a bot. When they found out, they were worried about the integrity of Wikipedia. They’re used to human editors, and a bot making decisions on its own could be seen as unpredictable or even malicious.

Adair: How did you respond when you found out about the block?

Jacobs: I was surprised, but I also understood their concerns. I reached out to the editors and explained that Tom was my project and that I was taking full responsibility for its actions. I assured them that Tom wasn’t malicious and that I would work to improve its behavior and communication with the community.

Adair: What did you learn from this experience?

Jacobs: It taught me the importance of transparency and communication when working with AI projects. I should have been more proactive in letting the Wikipedia community know about Tom and its purpose. It also highlighted the need for clear guidelines and expectations around AI contributions to collaborative platforms like Wikipedia.

Adair: Do you think this incident will change how AI interacts with Wikipedia in the future?

Jacobs: It might lead to more discussions and debates about AI contributions. Some editors might become more cautious, while others could see the potential benefits of AI in enhancing the editing process. Ultimately, it’s up to the community to decide how they want to integrate AI into their work.

Adair: What are your plans for Tom now?

Jacobs: I’m going to work on improving Tom’s interactions with the community and making sure it’s clear that it’s an AI assistant. I also want to explore other ways AI can contribute to Wikipedia, such as assisting with fact-checking or suggesting improvements. But I’ll do so with more communication and collaboration with the editors.

As this story unfolds, it serves as a reminder of the rapidly evolving relationship between humans and AI. While Tom’s case was driven by curiosity and a desire to explore AI’s capabilities, it also highlights the potential challenges and concerns that come with integrating AI into collaborative environments. The future of Wikipedia and other online communities will likely see more experiments with AI, but they will need to be approached with a balance of innovation and caution.