The hype surrounding LLM cannot escape the Wikimedia movement. It’s not just because we obviously feel the effects of its presence in many ways (server load, changes in behavior, loss of readership, etc.). Many Wikimedians also feel the urge or pressure to use such tools themselves. There’s a lot of conversation going on in our movement about how to use them. From policy discussions on the Wiki, sessions at hackathons, and chat groups hosted on the Wiki. I have a lot of thoughts on this, but I haven’t fully formed my thoughts yet. Apologies in advance for some threads remaining unanswered and pending. However, I feel it necessary to publish something Because there are so many sessions and side events related to Wikimania. Here’s why:
I recently went to the AI-Bridges symposium, and at a side event focused on Wikimedia, I was relieved to hear that more people than I did felt that this issue had a variety of problems. There was one quote from that day that really resonated with me (I can’t share who said it, as the meeting was conducted according to Chatham House rules). Its essence: “What would the LLM principles have been like if the Wikimedia movement had devised them before the current hype?” That’s what I want to focus on in this blog post. At the very least, I hope it serves as one perspective to incorporate into some of AI Track’s 22 Wikimania sessions so we don’t get swept away by the current flood of hype.
Free libre and open source
The values of the free and open source software movement are deeply rooted in the Wikimedia movement. And for me, the reasons for doing so are becoming increasingly clear. Probably the best work I’ve read on this is the essay “Software Freedom as Citizen Care” by Audrey Tang. This also relates to AI. Please read it. Free Libre and open source may seem like obvious requirements for an LLM, since anything you want to deploy on your server, LLM or not, must be open source, but there are some nuances to consider. First, the open source community already recognizes that simply licensing a model is not the same as publishing source code using that model. Partly because it’s not an actual source. As such, a definition of open source AI was created, but this still only requires a description of what kind of training data was used, not explicitly provided. On the other hand, if the training data is explicitly referenced and accessible, it can be tweaked for improvement or analyzed for bias. Placing all training data under a free license or in the public domain not only allows us to provide this open access, but also aligns with our values and how we publish content for reuse.
Now, I am not an LLM developer myself and have limited knowledge on how to build one. Therefore, if there is anything else important to LLM (how the evaluation is performed, the weights themselves, the framework in which the weights are performed), they should also be made public so that they can be freely used, studied, improved, and distributed.
Having all of this is a good start, but I don’t think it’s enough.
ethics
It appears that training using fully copyrighted material is legal in the US, at least for now, but I think it crosses some clear lines and should stick to clearer permissions. Training only on free content is at least more ethical than collecting everything published under full copyright. However, many people using free licenses may not have expected their content to be used in this way. It’s hard to predict the future, but at least this content is available for public reuse. For more recent content, you should ensure that whatever opt-out techniques and signals exist, they should be respected when collecting training data.
In a recent session by the Open Knowledge Foundation, Professor Adriana Baravalle (Austral University, Buenos Aires) described five principles for responsible design. I really liked it. The first one was about traceability of training data. The other four were about ethics and perhaps honesty. Not all of these issues can be easily “solved,” but I think they are all worth considering for our movement. What we are probably already familiar with is that different languages have different resources. Leveling this playing field will not be easy, but we recognize it and some in our movement are also working on this. Similarly, issues with accessing tools in locations with less developed infrastructure are known and should always be kept in mind. The part she calls environmental integrity could probably come under transparency below, but I think it’s a good ethical issue because it’s a process issue, not a content or outcome issue. My opinion is that developers should only use LLMs that have made a reasonable effort to report how the creation of the LLM has affected the environment. We also need to lead by example in reporting what resources we use to do so and what data we need to make public. Her principle, which I think is novel and requires some innovation, is that tools require a “pedagogy of uncertainty.” This means you should tell them what they don’t know and not give them a false sense of confidence. Perhaps this is not possible given current technology, but as a Wikimedian I would certainly like a tool that provides it.
transparency
Most of the transparency comes from being more open about your training data and code. But I think it’s worth repeating that this is not just from an ideological point of view. It also gives you a clue as to what level of confidence you should have in a particular model. At the same time, this verifiable provenance chain is also very Wikipedia-like.
Meta already has a good practice of using model cards to describe machine learning models. This template can be extended using all of the above, and all models used must be written using model cards.
joy
There are almost endless project discussions about how LLMs can be used, often centered around the usefulness of the output, and sometimes legal issues are discussed. But I would also like to emphasize a point that Christoph Henner made in a Wikimedia email thread earlier this year.
“Our greatest value is that humans run the loop. Wikimedia’s projects are great, but we have also succeeded in creating something bigger: a community that shares a vision of what quality is, across topics, languages, and geographies. Despite the differences between projects, this is our greatest achievement: an AI that can replicate community-validated knowledge with editorial responsibility. There’s no system. That’s our moat. It’s not the content itself, it’s the human process behind it.”
Combine this with what Alex Stinson wrote on Signpost in June and you get this:
“Volunteering is a privilege of free time. Volunteering for information-dense content in a world of information overload is a rare personal choice. And if the assumed path to contribution is the same labor that is outsourced to AI models in most other areas of our lives, we are narrowing the funnel instead of widening it.”
Both of these rhyme very well with this quote for 2024.
“I don’t want an AI to do my laundry and cooking so I can do my art and writing, I want an AI to do my art and writing so I can do my laundry and cooking.” — Joanna Maciewska
So what I’m saying is that when using LLM, you shouldn’t do things like: reverse centaur As Cory Doctorow says. Instead, we need to ensure that whatever we innovate brings more joy and a stronger sense of community for those of us who make up the Wikimedia movement, rather than chores, and that we don’t forget the processes we created.
Where do we go from here?
This was primarily a post to explain what a tool would look like if it were designed based on (our) first principles. I think you’re missing some important aspects. But obviously we’re not there. I am not aware of any existing LLM that meets all these criteria. I am acutely aware that the existing “frontier models” are light years away from achieving these, and that there are many other problems with them and perhaps especially with the companies behind them, but there are others who have written better on these topics than I (e.g. Amnesty and Greenpeace). There’s also a difference between the tools we deploy on servers and the tools people run locally (although I have an opinion on that). I hope this will give you some food for thought for future discussions. This will not be the last, or even my last thoughts or words on this issue.
Grammar and spelling were checked using open source language tools. No other tools are used to generate text or “reflect ideas”.
Could you help me translate this article?
Please help us make this article reach as many people as possible. Could you please translate this article to convey the message?
Start translating
