Wikimedia principled LLMs

The hype around LLMs does not escape the Wikimedia movement. Not only because we are obviously feeling the effects of its presence in multiple ways (server loads, changed behavior, and lost readership). Many Wikimedians also either feel an urge or a pressure to use such tools themselves. There are plenty of conversations about how to use them going on in our movement. From policy discussions onwiki, sessions at our hackathon, and chat groups running with it. I have a lot of feelings about this, but I also haven’t really formed my thoughts fully yet. Apologies in advance for some threads being left hanging unanswered. However, I feel that I need to publish something before Wikimania as there are so many sessions around it and side events happening, and here is why.

Recently, I went to the AI-Bridges symposium, and in a Wikimedia-focused side event, I was relieved to hear that there were more people than me feeling that the issue is sided with many different problems. There was one quote during the day that resonated strongly with me (I cannot share who said it due to the meeting being under Chatham House rules). The essence of it: “What would principles for LLMs have looked like if the Wikimedia movement had come up with them before the current hype?” And that is where I want to focus this blog post. If nothing else, I hope it will serve as one perspective to bring into some of the 22 Wikimania sessions in the AI track so that we are not getting swept away by the current hype flood.

Free Libre and Open Source

The values of the free and open source software movement are deeply ingrained in the Wikimedia movement. And to me, it has been clearer and clearer about why it should be. Perhaps the best writing I read about this, and that is also related to AI, is the essay Software Freedom as Civic Care by Audrey Tang. I highly recommend reading it. Free Libre and Open Source may seem like an obvious requirement for LLMs as everything we want to deploy on our servers must be open source, LLMs or not, but there are some nuances to consider. First, the open source community already recognized that just slapping a license on a model isn’t the same as publishing source code with it. Partly, because it isn’t really the source. So the Open Source AI definition was developed, but this still only requires a description of what kind of training data has been used, not explicitly providing it. If, on the other hand, the training data was explicitly referenced and made accessible, it would be possible to tweak it for improvements and analyze for biases. Having all training data under a free license or in the public domain would not only enable providing this open access, it would also align with our values and how we publish our content for reuse.

Now I am not an LLM developer myself and only have limited knowledge of how they are built. So if there are other things that are essential (how evals are done?, the weights themselves?, the framework they run on?) to the LLM, these should also be published so that they can be used, studied, improved, and distributed at will.

Having all this would be a good start, but I don’t think it is enough.

Ethics

Even though it, at least for now, seems to be legal in the USA to train on fully copyrighted material, I think it is passing some clear lines and that we should stick to more explicit allowances. Training on only free content is at least more ethical than also harvesting everything published under full copyright, but it is likely that many people using free licenses didn’t foresee their content being used in this way. Predicting the future is challenging, but at least this content was published for general reuse. For more recent content, we should make sure that for whatever opt-out technologies and signals there are, these should have been respected when collecting the training data.

In a recent session by the Open Knowledge Foundation, Prof. Adriana Baravalle (Universidad Austral, Buenos Aires) described five principles for responsible design. I liked them quite a lot. The first one was around traceability of the training data; the four others were about ethics and perhaps honesty. Not all can be easily “solved”, but I think they are all worth thinking about for our movement as well. Perhaps one we already are familiar with is the variance of resources in different languages. It’s not easy to level this playing field, but we are aware of it, and several in our movement are also working on this. Similarly, the issue about access to the tools in places with poor infrastructure is known, and we should keep bearing that in mind. The part she calls ecological honesty could perhaps have been under transparency below, but I think it fits as an ethical issue as it is a process issue and not about the content or results. My view is that we should only use LLMs whose developers make a proper effort to report on what ecological impact the creation of the LLM had. And we should also be role models in reporting on what resources we use to run them or require that data to be public too. The principle of hers that I find novel and might require some innovation is that a tool should have a “pedagogy of uncertainty”. By that she means that they should communicate what they don’t know and not offer false confidence. Perhaps this is impossible given current technology, but as a Wikimedian, I would certainly like tools to offer that.

Transparency

Most of the transparency would be taken care of by being more open about the training data and the code. However, I think it may be worth repeating that this would not only be from an ideological viewpoint; it would also give a clue of what levels of trust can be had with a given model. At the same time, this verifiable provenance chain would also be very Wikipedian.

On Meta, we already have a good practice of describing machine learning models with model cards. The template for this could be expanded with everything mentioned above, and every model we use should then, of course, be described using a model card.

Joy

While there are almost countless discussions on projects on how we can use LLMs they are often centered on the usefulness of the output, and sometimes the legal issues are discussed. But I also want to highlight a point Christophe Henner made in an email thread on Wikimedia-l earlier this year:

“Human running the loop is our number one value. Wikimedia projects are great, but we have also successfully created something bigger: communities that across topics, languages, and geographies share a vision on what quality is. Even with the differences across projects, this is our biggest achievement. No AI system can replicate community-validated knowledge with editorial accountability. That’s our moat. Not the content itself, but the human process behind it.”

Combining that with what Alex Stinson wrote in the Signpost in June:

“Volunteering is a privilege of free time; volunteering on information dense content in a world of information overload is a rare personal choice; and if the assumed pathway to contribution is the same labor being outsourced to AI models in most other parts of our lives – we are narrowing the funnel not growing it.”

Both of these rhyme really well with this quote from 2024:

“I want AI to do my laundry and dishes so that I can do art and writing, not for AI to do my art and writing so that I can do my laundry and dishes.” — Joanna Maciejewska

So the point I want to make is that when we use LLMs we shouldn’t turn into reverse centaurs as Cory Doctorow would put it. We should instead make sure that anything we innovate brings us humans that constitute the Wikimedia movement more joy and a stronger sense of community rather than more chores and not forget the process that we have created.

Where do we go from here?

This was mostly a post to describe how our tools could look if we designed them from (our) first principles. I am sure I have missed some important aspects. Yet, obviously we aren’t there. I don’t know of any existing LLM that would meet all these criteria. I am acutely aware that existing “frontier models” are lightyears away from meeting these and that there additionally are plenty of other issues with them and perhaps in particular the companies behind them, but others have written about those topics better than I ever could (for example, Amnesty and Greenpeace). There is also a difference between the tools we deploy on our servers and the tools that people run locally (although I have views on that too). I still hope I have given some food for thought for our upcoming discussions. This is not the last of it, certainly not even my own last thoughts or words on the matter.

The grammar and spelling were checked with the open source Language tool. No other tools have been used for generating text or “bouncing ideas”.