Hi everyone,
We recently shared a Staff Blog post announcing a pilot program to test LLM-assisted translations on a small number of wikis, which has understandably caused a lot of confusion and concerns. The communication wasn’t as clear as we hoped, so we want to correct that.
First off, we believe in "enablement not replacement” with AI and in no way do we intend to replace editors with AI tools. We hope the tools and features we create for you are useful in helping you build out your pages and communities, and our goal is to expand the online communities you work so hard on. The idea is to make it easier for international editors to contribute to wikis in their own language, especially in cases where starting from zero can feel overwhelming.
We have taken some time to release this blog as we wanted to get it right for you, and include as much information as possible. That can take a little time, but thank you for being patient through this.
Why We’re Doing This[]
Our main goal with this experiment is to figure out how we can reach fans and potential editors in non-English-speaking regions. We have not given as much attention to our non-English wikis as we have to our English communities, and we’re trying to change that. We asked ourselves a bunch of tough but necessary questions:
- Can we grow non-English traffic and justify renewed focus in those communities?
- Can we attract new editors if there’s already a foundation of content in place?
- What kind of support do these new communities need?
We chose AI-assisted translation as a way to test this idea quickly. It’s not perfect, but it gives us something to build on. We very deliberately chose to translate wikis where no German editors had created a wiki - and in some cases, where even the English original was created by Fandom staff.
If LLM-assisted wikis can act as a starting point for communities to grow around, then we’ll know we’re on the right track. We’re looking at whether these translated wikis can bring in new readers and editors. We want to provide readers with content - and editors with starting help - where otherwise no content or community might exist. We want to make it easier for fans in different languages to get involved, not remove the need for human contribution.
How the Pilot Works[]
Appreciate that a lot of you know this already, but to clarify:
Pilots are small tests run to see if something works, collect some info, and spot any problems before doing it on a bigger scale. We have included no more than five wikis in this pilot. Some are Fandom-created and others are community-created. The first one we worked on is the Shogun wiki, created by Fandom staff.
We chose three Fandom-seeded wikis because, as we were the ones that started the English wikis, we were able to translate them without being too intrusive to the Editor community. For others, we went through a list of high-impact Wikis, and dropped any that already had an existing translation. We conducted an SEO review for wikis that did not have German translations, to see if it would give us statistically meaningful results for the pilot. We then reached out to the Admins, and gained their consent to move forwards.
Some of you have asked for clarification on the model we are using: we’re using Gemini 2.0 Flash from Google Vertex, which we’ve found performs better than traditional translation models like Google Translate. We first looked at traditional machine learning models, but they weren't quite cutting it, so we decided to go with a large language model instead for better fidelity.
Every wiki in this pilot goes through a human quality check by our internal team. On the first one alone, over 300 edits were made after the AI translation to make sure the final result was accurate and readable. We use glossaries, and clarify the in-universe terms and manually correct inaccuracies. Human review is absolutely essential to this process. The AI cannot get these things right, and we do not intend to leave them incorrect, which is exactly why we are also focused on collecting feedback for the accuracy of these pages even after our manual intervention. We’re being careful about where we apply this. If a wiki already has a translation, we’re not touching it. We’re also making sure that all generated content is clearly labeled, and is stated in the licensing note as well as in each page’s history.
What About SEO?[]
One concern we’ve seen is whether translated content could affect the performance of English wikis in search engines, which is a super valid concern, but one we’ve prepared for and are actively monitoring. We’ve set up proper hreflang tags so that search engines know which language version to show to which users. We’re keeping a close eye on SEO performance, as it is one of our pilot metrics. We are tracking performance for both the EN and DE versions of each wiki that is launched with this pilot, alongside some human-translated DE wikis of similar search volume to serve as control groups for comparison.
What Happens Next[]
The remaining pilot wikis are being worked on now and will roll out in the coming weeks. Because of the manual checking stage, we can’t give you a concrete timeframe yet, but we will tell you as each one is rolled out. Again, this is a mix of Fandom-created, and user-generated wikis with expressed consent from their admin teams. Once all pilots are live, we’ll take a step back and look at the results. Here’s how we plan to keep everyone informed:
- Frequent updates for the first few months as the pilot is expanded and results are clearer
- Ad-hoc updates after that, with deeper analysis on things like audience growth, editor contributions, and improvements made based on feedback
All updates will be published in both English and German to make sure they’re accessible to the communities we’re trying to reach.
Moving Forward[]
We know this kind of initiative can raise concerns, especially around AI's role in editing. We’ve heard from many of you, including those who signed the petition about AI and editor value. We understand that the sentiment of AI is turbulent, but we believe that using LLMs for what we have envisioned here makes a lot of sense. LLM-assisted translation is NOT fully-automated publishing. It is a tool, not a substitute.
If we roll this out further at any point, it will always remain opt-in.
For this pilot, we have a feedback form in place using our site notice on any page that is being utilised for this. It is a way for us to understand any translation errors and the quality of our output. This is really important and valuable to us. We also want to make sure that editors participating in this pilot feel supported. They’re helping us test something new, and we ask that everyone in the community treats their work with respect and kindness. We’re taking what we learn from this pilot to inform what comes next. If it doesn’t hit the mark, we’ll say so. If it does, we’ll keep building on it with your help and feedback.
We’re committed to making Fandom a place where fans everywhere feel welcome, seen, and empowered to contribute. Thank you to everyone who’s shared their thoughts so far.
We hope that staying transparent through this pilot will help build some trust, and we will continue reading and processing your feedback through this.
Click here to follow the Fandom staff blog.
Click here to sign up for the From the Desk of Community email newsletter.
Join our Official Discord server for registered editors!