Bridging the AI Capacity Gap through Open Science and Open Source: Inria Chile’s Participation in the UN Open Science Conference

From October 16 to 18, the 4th United Nations Open Science & Open Scholarship Conference was held in Tokyo at the United Nations University. During the session on “Artificial Intelligence, Open Science, and the Global Digital Divide,” Nayat Sánchez-Pi, Director of Inria Chile and the Franco-Chilean Binational Center on Artificial Intelligence, participated in the panel alongside international experts such as Roheena Anand, Executive Director of Global Publishing Development at Public Library of Science; Sadao Kurohashi, Director of the National Institute of Informatics and professor at Kyoto University; and Bishesh Khanal, Director of the Nepal Applied Mathematics and Informatics Institute.

The Conference built upon discussions from the three previous editions and was organized around three themes. The first focused on “AI, Open Science, and the Global Digital Divide.” It highlighted how artificial intelligence (AI) aims to revolutionize science by using open science outputs (data, code, and papers) as an essential foundation for its training. However, the session addressed the paradox that many AI tools trained with these open outputs are not themselves open.

The debate explored what it means to apply an open framework for AI centered on human rights. It was emphasized that international cooperation is crucial, especially given that global resource inequality exacerbates power imbalances. The alignment with the Global Digital Compact (which seeks to advance equitable approaches) and the Pact for the Future (which integrates a human rights perspective into technology regulation) was discussed. Finally, it was proposed that the values of open science can provide a framework for managing AI-generated content, ensuring privacy and prioritizing verifiability.

Nayat Sánchez-Pi, Director of Inria Chile and the Franco-Chilean Binational Center on Artificial Intelligence, shares the main points she presented at the conference. These focused on advances in artificial intelligence, open source, and open science, as well as the importance of discussing the ethical and responsible use of these tools for the benefit of society.

One of the main drivers behind the rapid progress of artificial intelligence in recent years has been the AI community’s strong commitment to the principles of open science. Of course, this progress has also been fueled by advances in theory and the continual evolution of computing hardware. Yet, what stands out is the collective understanding—almost a tacit agreement—within the AI community that openness accelerates discovery.

However, open science is incomplete without open-source software. A research paper describing a machine learning method is merely a recipe without the accompanying code. When that code remains proprietary, the work cannot be reproduced, and the knowledge cannot be fully shared. In this sense, no true open science exists without open source.

Therefore, open science in AI takes different forms:

• The open dissemination of scientific results through open-access publications.

• The creation and sharing of public datasets and benchmarks.

• The development and adoption of open-source software, enabling others to reproduce results and easily build upon prior work.

• And, increasingly, the release of “open weights”, where trained model parameters are made publicly available when training from scratch would be prohibitively expensive.

However, open science alone is not enough. It is a necessary condition for sustaining AI’s momentum—but not a sufficient one. There is a widening gap between those who can effectively leverage open scientific outputs and those who cannot, because of either limited computational resources or a shortage of skilled practitioners.

In Inria and Inria Chile, we have a special focus on open science, open source and furthermore, open collaboration. It's in the DNA of the institute. At Inria, we have created more than 1500 actively used open source software, and many more actions.

In the context of AI, we can highlight scikit-learn, which is one of the main libraries for doing machine learning. scikit-learn is a good example of how we address that issue of the growing gap, as it has focused on enabling users to do effective machine learning in personal computers and other easily accessible hardware.

There is another remarkable Inria project that I think should be highlighted here that is called Software Heritage, supported by Unesco. Its mission is to preserve and index all publicly available source code, ensuring its long-term traceability. Modern “Library of Alexandria”, international, non profit, long term initiative addressing the needs of industry, research, culture and society as a whole.

Interestingly, in the age of large language models, this project has found new, urgent relevance, as these open code repositories are now required for training and evaluating LLMs. What we’re doing is preserving the world’s digital knowledge for the next generation of AI developers.

But beyond sharing artifacts, code, papers,etc., we have been committed for a long time with open collaboration. Part of our mission in Inria Chile is actually closing that widening gap, that digital divide, by both connecting the European and Latin American computer science ecosystems.

In practice, this means three things:

• Promoting robust collaboration between French and Latin American researchers, startups, and institutions.

• Supporting projects that make AI usable, accessible, and safe for local economies and communities.

• Ensuring openness is the foundation that makes all progress shared and sustainable

About a year ago, Inria and the Chilean Ministry of Science signed an agreement to create the Franco-Chilean Binational Center for AI. This new center is the operational tool for our shared priorities. This center implements a multi-stakeholder approach—engaging two governments, academia, industry, and civil society—to develop and implement projects that benefit all of society. AI must not be a luxury for major capitals. Latin America needs to solve regional problems with locally relevant, resource-efficient AI.

We focus on advancing an AI that is:

• Usable, so it can truly support society and the economy,

• Accessible, so that everyone, not just a few, can benefit from it, and

• Safe, ensuring responsible and trustworthy development.

This Center reflects the common values not only between France and Chile but among many countries committed to responsible, ethical and human-centered AI.

We can see that open science, including open source principles, are essential to keep the momentum that we have reached in AI. Needless to say, there is a lot of work to be done to manage the AI Capacity Divide.

First, by preserving what we have achieved. In the last few years we have seen how many actors, in particular private companies, have started to close their science, even when they claim to be “open” by name. I am convinced that a closed science approach leads to lack of innovation, and we have already seen the results.

Second, there is an important aspect that should also be addressed, as taking any idea to the extreme naturally brings risks, obvious and non-obvious risks: We must act with responsibility and manage the risks that openness brings.

We must all agree on clear “red lines” with respect to AI uses and AI behaviours for example: autonomous cyber attacks, weapons of mass destruction, and mass surveillance which should be non-negotiables.

However, we also need to address the nuanced risks of openness. This means recognizing that there are times when, even if we apply these models with a valid, legitimate goal and without ill intent, their application can still result in unintended harm.

The fact that these open models are more readably available to inexperienced practitioners implies that we should properly document not only their features but also their known shortcomings and possible misuses and develop strategies for responsible openness.

I’ll give an example, if we gather data about organisms, animals, and ecosystems for a well-meaning biodiversity study where the location is an important metadata. If we then share it in an open way, we risk that this data could then be used by hunters and poachers to go and hunt these animals. This sort of borderline cases should be analyzed, and we need to have global and local approaches to it. 

And third, for open science to be truly open, it must enable everyone to be an active participant and not just spectators or consumers of the science being done by big companies. 

It means there is a lot of work to be done to close the gap.

The good news is that we are already working on the solutions: the current, resource-intensive paradigm of Big AI, relying on massive datasets and colossal computing power, is antithetical to open science and sustainable development. We are doing more theory and science that enable frontier AI to be assimilated at more “common” levels of computation.

We are pursuing Frugal AI—an approach focused on achieving high scientific impact with minimal resources, less data, and less energy, which contributes to the democratization of AI. 

This is driven by three core necessities:

• Reproducibility and Falsifiability: for a result to be scientific, it must be reproducible. As a scientific paper is merely a "recipe" without the source code and training method, No Open Science is possible without Open Source.

• Sustainability: the energy footprint of massive, closed models is irresponsible. Frugal AI encourages the reuse of components and efficient, low-footprint design, making science more sustainable and accessible to institutions with limited computing infrastructure.

• Transparency and Regulation: effective regulation and evaluation—like that envisioned by the EU's AI Act—are impossible if models remain opaque. Transparency is not just an ethical ideal; it is a regulatory necessity that requires access to code for auditing bias and ensuring reliability in critical sectors.

Focusing on creating the means to ensure AI is safe not only by investing in evaluation tools but in mitigation strategies and creating collaborative programs as public good like the one Inria has created with the center in Chile, the Franco-Chilean Binational Center on AI that pools together brains and shared resources under the umbrella of our shared principles and values between France and Chile and also between the regions including those of open science.

Finally, openness and safety is no longer optional in the AI era. We need binding global rules, operational and international collaboration, to guide AI development to the promise of innovation and the wellbeing of our societies.