AI and the Commons

We present Public Domain 12M (PD12M), a dataset of 12.4 million high-quality public domain and CC0-licensed images with synthetic captions, designed for training text-to-image models. PD12M is the largest public domain image-text dataset to date, with sufficient size to train foundation models while minimizing copyright concerns. Through the Source.Plus platform, we also introduce novel, community-driven dataset governance mechanisms that reduce harm and support reproducibility over time.

Last week, Spawning launched source.plus, a platform for “curating, enriching and downloading non-infringing media collections in bulk for AI training.” This is a significant step in addressing a host of issues with AI training datasets, such as LAION or face recognition training datasets.

The aim of this experimental platform is to demonstrate that licensed content is not the only viable solution:

This means the most conscientious developers and most affected communities are often on the sidelines of this rapidly developing field, whereas these are the very groups that need to be steering its evolution, and they too should be able to benefit from participation with AI.

Its real value lies not just in the volume of aggregated media files but in how they are curated and governed. It is an interface to established collections, it introduces additional mechanisms – many of which we have proposed in our recent white paper, Commons-based governance of data sets for AI training. For example, source.plus is the first collection to offer an “opt-out” mechanism. Spawning is also planning to introduce value-sharing mechanisms, including paid collections of in-copyright works and a donation mechanism that supports cultural heritage institutions.

…there is a circularity inherent to the authoring of AI training sets. [...] Because they need to be so large, their construction necessarily involves the use of other models, which themselves were trained on algorithmically curated training sets. [...] There are models on top of models, and trainings sets on top of training sets. Omissions and biases and blind spots from these stacked-up models and training sets shape all of the resulting new models and new training sets.

Screenshot from Models all the way down, © Knowing Machines

A group of AI researchers coordinated by the French start-up Pleias wants to challenge the belief that you need copyrighted materials to train an LLM that competes with the models developed by leading AI companies. Yesterday, they released what has been dubbed the largest open AI training data set consisting entirely of public-domain texts. The collection is called “Common Corpus” and is available on Hugging Face for download. The resource is multilingual – besides English, it includes the largest open collections in French, German, Spanish, Dutch, and Italian, as well as collections for other languages.

Training data is a key resource for developing AI systems. Until very recently, it was commonly believed that LLMs, such as those behind popular services such as ChatGPT or Bard, could not be trained without relying on copyrighted content. If this is the case, access to high-quality data may continue to be a significant barrier for independent AI developers seeking to compete in the LLM market.

Datasets consisting only of public domain texts have significant limitations, the most important being that they miss more contemporary information because they are comprised of historical sources or older publications where copyrights have already expired. It remains to be seen whether public domain datasets can indeed compete with datasets containing more contemporary content that is protected by copyright.

Last week, the Commission published the AI Innovation Package to support Artificial Intelligence startups and SMEs. The measures listed in the package include facilitating access to AI-focused supercomputers, which is expected to help expand the use of AI to a wide range of users, including European start-ups and SMEs. An article in Science|Business rightly pointed out that the plan outlined by the Commission suggests that it is pinning its hopes on private companies to keep the EU competitive in AI.

Putting faith in private actors is not sufficient. The way to address the imbalance of power and market concentration must also include investing in the development of systems that serve society and have the best interests of people and the planet at their core. At the moment, it doesn't seem that this approach will be implemented in the field of AI. Whether we can expect any efforts to create a public option for AI in Europe remains to be seen. Some public interests, such as ensuring diversity and transparency in the datasets that train AI models, are simply not always aligned with the interests of corporations who might favor the fastest and cheapest solutions. This is where public authorities, civil society and the large communities of scientists and practitioners working on AI in Europe have a role to play.

Science | Business reported that German Research Minister Bettina Stark-Watzinger believes that "no state or association of states can match the investments made by large corporations like Microsoft or Google with public investments.”

This suggests that the German government throws in the towel and assumes  that private actors are equipped and have the intention to develop digital services that serve the public interest and allow people to enjoy their fundamental rights.This approach is somewhat disappointing, and given the example of private social media platforms that fail to fulfill the role of digital public spaces,, it does not appear to be appropriate. To put it simply, without public funding, European society won't get AI that serves the public.

Last week, Stanford HAI, Stanford CRFM, Princeton CITP, and RegLab released a Policy Brief on Considerations for Governing Open Foundation Models.

The brief outlines the current evidence on the risks of open foundation models (FMs) and offers some recommendations for policymakers on how to think about the risks of open FMs. The brief argues that open FMs - defined by the authors as "models with widely available weights" — "provide significant benefits by combating market concentration, catalyzing innovation, and improving transparency." The authors therefore conclude that "policymakers should explicitly consider the potential unintended consequences of AI regulation on the vibrant innovation ecosystem around open foundation models."

The policy brief also points out that despite widespread concern about the dangers of open foundation models that has dominated policy discussions, "the existing evidence on the marginal risk of open foundation models remains quite limited. The key questions for understanding their impact are the risks posed by open models relative to the risks posed by other models (the marginal risk):

To what extent do open endowment models increase risk relative to (a) closed endowment models or (b) pre-existing technologies such as search engines?

Coming a few days after the final compromise on the EU AI Act, the policy brief provides further support for the AI Act's approach of providing targeted exemptions for open (source) AI developers. The final compromise on the AI Act sidesteps policies - liability for downstream harm, licensing of model developers - that the authors of the policy brief see as particularly problematic for open AI developers. As we argued in our analysis of the Act, the overall approach to open source AI development in the AI Act is quite sound, although there is still room for improvement by getting some of the details of the transparency obligations right.

1. The AI Act should allow for proportional obligations in the case of open source projects while creating strong guardrails to ensure they are not exploited to hide from legitimate regulatory scrutiny.
2. The AI Act should provide clarity on the criteria by which a project will be judged to determine whether it has crossed the “commercialization” threshold, including revenue.

The European Union's upcoming AI Act will require adequate standards to become fully operational, and much work is required to ensure that the standardization process does not conflict with the Act's inclusion and transparency objectives.

The process will be led by the European Committee for Standardization (CEN) and the European Committee for Electrotechnical Standardization (CENELEC). In the past, they have been criticized for their secrecy and lack of transparency. The standards must be made public, but some fear that the private sector will have too much control over the process, which could have an impact on human rights. The standards' nature and scope will also have geopolitical implications, with some calling for greater international cooperation.

Standards will be essential in enforcing the EU's AI legislation, and CEN-CENELEC will have just two years to formulate and agree on a series of AI standards.

The dependence of GFMs on digital commons has economic implications: much of  the value comes from the commons, but the profits of the models and their applications may be disproportionately captured by those creating GFMs and associated products, rather than going back into enriching the commons. Some of the trained models have been open-sourced, some are available through paid APIs (such as OpenAI’s GPT-3 and other models), but many are proprietary and commercialized. It is likely that users will capture economic surplus from using GFM products, and some of them will have contributed to the commons, but there is still a question of whether there are obligations to directly compensate either the commons or those who contributed to it.In response, the paper identifies three proposals for dealing with the risks that GFMs pose to the commons.