How AI Threatens the Social Contract of Free Knowledge

Short summary: Following our new annual plan, the Foundation is reviewing our licenses and legal policies to help address the ways that new AI tools have shifted how people obtain and consume knowledge and protect the social contract of free knowledge.

In 2022, when ChatGPT first went out to consumers, it looked like an interesting copyright problem for the law nerds, one that we jumped on quickly for legal analysis. In the ensuing three years, two major things have happened together: the number of people using various forms of generative AI has skyrocketed, and the demand for continuous data to support the various types of training and fine tuning of these tools has soared equally high. These changes threaten the basic social contract that has allowed the free knowledge movement to thrive over the past decades.

The Foundation has already written from a technical angle on the way that automated web crawling and scraping impacts the operation of the Wikimedia projects (see here and here). As our technical response has evolved, we continue to keep in mind the importance of reaching people where they are finding their knowledge as new tools develop. And we are happy to see some partners who are willing to work with us and support the Wikimedia movement’s free knowledge mission. But as time has passed, the nature of the issue has also moved beyond just technical traffic management. 

In our annual plan, we highlight the way that new AI tools have created a long-term shift in the way Wikipedia is used and the need to consider responsible reuse. This year, we’ve also committed to reviewing our licenses and legal policies to protect the knowledge commons. Here, we want to expand on the way structural shifts on the internet have created a fundamental question for the free knowledge movement: how can free knowledge be sustainable in an environment where people get their information without ever seeing its source or even knowing that the source exists? 

The social and legal structure of free knowledge

Historically, knowledge production and dissemination has been expensive and controlled: printing presses were expensive, and the printers and publishers of the world tended to organize into guilds and societies for exclusivity and control. Governments supported that control with legal rights that granted exclusivity to most creators of knowledge and made it easier for government power to control what people knew. The idea of free knowledge thus arose as a social movement: to empower people by reducing and eliminating barriers to accessing knowledge. Historically, this meant encouraging and promoting physical copies of a work, and as the internet came into being, it meant creating legal terms for spreading digital copies that didn’t involve locking them behind money, confidentiality, or other restrictions. At the same time, free knowledge has often kept some restrictions to support the ultimate goal of empowering people: for example, requirements for credit ensure that people who create knowledge are themselves recognized for their work, which encourages more creation. Requirements to keep knowledge free ensure that new people who build upon existing knowledge, won’t lock it behind new walls. Some creators and communities have added limitations on use, such as making knowledge free but forbidding derivatives that adapt the original, or by setting usage restrictions limiting reuse to noncommercial purposes in order to prevent harmful or exploitative reuses.   

The free knowledge movement has created legal structures to reflect and support this social structure. In the 21st Century, the most important of these legal structures has been the Creative Commons family of licenses. With a few exceptions, the Wikimedia projects all use Creative Commons (CC) licenses: mostly the Attribution-ShareAlike license (CC BY-SA 4.0). The Wikimedia community may already be intimately familiar with these licenses, but it’s worth taking a moment to look at how closely this tracks to the intended social contract. To meet the desire of people who want credit for their work, CC BY, the most basic license (the “by” meaning literally “who is this work by?”), sets legal terms that the information about a work sticks with it through the internet by requiring reusers keeping its title, author, source, and license with it. The SA, which stands for ShareAlike, is an addition to the basic license that requires that modifications to the work be shared freely, which is important for creators (such as Wikimedians) who want to ensure that everything built on their contributions to free knowledge also remains free. ShareAlike is expressed in legal language that requires derivative works (also called adaptions) to be published under the same free license. 

The threats to the social contract of free knowledge

For the last 25 years, this structure basically worked: people contributed free knowledge to be used and built on by others and as it spread around the world, it would link back to the original source. For example, people who learned from the Wikimedia projects would find out where their knowledge came from through citations, and people who contributed could see their work meaningfully used on the projects and beyond. Today, the combination of mass scraping and mass usage by some companies threatens to break this social contract of the free knowledge commons in three ways: 

1. Many new tools do not explain how their knowledge was created, removing the credit and gratitude that is supposed to accompany free knowledge creation.

2. Many tools do not cite their sources, breaking the link for new learners to find out where their knowledge comes from and contribute.

3. Many tools do not effectively contribute knowledge back to the free knowledge commons, making it much more difficult to build on and improve free knowledge.

Let’s look at these points in some more detail.

First, the data about what has gone into training generative AI tools is often sorely lacking. AI models are built on massive, frequently updated data sets; the training process is complex; the companies building the models treat their methods as trade secrets; and in some cases crawlers that collect data from the Internet simply lack adequate bookkeeping and logging functionality. This could be done better! There could be information about what goes into a data set used for training, what materials were actually acquired by the entity doing the training, and what actually showed up to be rated in a statistical process by either machines or humans. You can see what this might look like in the Foundation’s own model cards.  In the case of others reusing a wiki, this does not need to be a heavy lift: the url and time when a given wiki was crawled combined with the history function of Mediawiki software means that any given person could find exactly what the wiki looked like and who worked on it with just that minimal data. 

Second, the methods generative AI tools use to produce outputs often separate their users from the sources of the information they provide. Many outputs cite no sources, even to the point of potentially violating existing copyright law. In many cases today, tools also use retrieval augmented generation (RAG) in which the tool checks a website to answer a question instead of just relying on its original training data. RAG makes the possibility of citing a specific source much easier, since there is a technical check to at least one specific source directly associated with the specific answer from the AI tool. The Wikimedia Foundation is already experimenting and offering tools to partners that help ensure Wikimedia’s human-created content is fairly credited through a dedicated Attribution API and framework. But today there are still many tools where the answer provided may not have the right sources, may fail to match its sources and mislead the user, or may provide source information in a way that the user never notices. 

Finally, the structure of generative AI tools makes it very difficult for people to contribute back to the free knowledge ecosystem. The environment of working with an AI tool for knowledge often means never leaving that tool’s interface. Combined with poor or no attribution, the people using the tools never learn that there’s a free knowledge community that they could join and contribute to in a range of ways. And even if people do consider contributing, starting from an AI interface can create a host of new difficulties: the quality of work most tools generate is still highly variable.  These uncertainties can make a great deal of work for editors of the Wikimedia projects dealing with AI-generated contributions, and we’ve seen both English and German Wikipedia create restrictions to contributing AI-generated content to their language Wikipedia as a result (here and here). 

In this environment, where so many people are obtaining knowledge in new ways, we think that there needs to also be new and better ways for the companies hosting AI tools and the users receiving knowledge through them to support and sustain the knowledge commons. 

The future of the knowledge commons

Right now, we’re looking at ways that we can try to match the social promise of the free knowledge commons to the new environment of today. We want good, high-quality free knowledge to be able to reach people wherever they’re looking for it, including through the use of a range of AI tools. We want to find new ways to strengthen the social contract that gives appropriate recognition to everyone who contributes to free knowledge. And we want to find the right way to help that knowledge stay free as new people build and grow it. In the annual plan for this fiscal year, we’ve committed to reviewing our licenses and legal policies to protect the knowledge commons, and we hope to share more ideas in the coming months as we identify specific actions we think will help.