Can we bring sovereign states together in a multilateral forum to agree on how to govern powerful technologies? We’ve got good news this week. Following two years of extensive preparation, the United Nations Educational, Scientific, and Cultural Organization (UNESCO) adopted in Paris a Recommendation on the Ethics of Artificial Intelligence, which has been hailed as the first global standard-setting international instrument on the matter.
The final text was approved by acclamation on 23 November by UNESCO’s General Conference, chaired by Brazilian Ambassador Santiago Mourão. The Recommendation has two pillars: one focused on values and principles, and the other on areas of policy action, where specific proposals are listed for Member States to consider, as appropriate, in line with the non-binding nature of the document. If willing to do so, governments would apply on a voluntary basis the provisions put forward by taking concrete measures for their implementation. The journey to arrive at this point, however, started a long time ago.
UNESCO has consistently been promoting a humanistic approach to ethics and policy in response to emerging challenges related to AI, including reflections on what it means to be human in the face of disruptive technologies. The World Commission on the Ethics of Scientific Knowledge and Technology (COMEST), an advisory body created in 1998, pioneered workshops and roundtables, leading up to the publication in 2017 of a report on robotics ethics. COMEST was also responsible for producing a preliminary study on the ethics of AI in 2019.
These studies on technical and legal aspects stressed that AI is not confined to a tangible location, which makes regulation of AI technologies more puzzling. Due to their transnational character, durable solutions need to be found at the global level. In UNESCO’s vision for a human-centered AI future, a normative instrument on the ethics of AI should serve as a means of mainstreaming universal values into AI systems, which must be compatible with internationally agreed human rights and standards.
Last year, despite the pandemic, a specially appointed ad hoc expert group of 24 members, invited in their individual capacity, worked through online meetings and consultations to prepare the first draft (outcome document) of the Recommendation, which was made available for public consultation. In addition to principles drawn from or inspired by the international consensus achieved in the 100+ published frameworks related to AI ethics, policy action was given special attention to foster implementation. The experts pushed for a holistic approach with due regard to human rights. They mostly agreed that the United Nations had a significant role to play, akin to a “beacon”, empowering people and increasing participation by civil society.
Intergovernmental negotiations at expert level were held between April and June 2021, when Member States reached a consensus in painstaking and time-consuming virtual meetings. Detailed discussions on the lengthy text, paragraph by paragraph, required almost five weeks of negotiations. Member States expressed concerns related to broader political issues, national sovereignty, human rights, gender, commercial interests, and other economic issues that proved controversial. They eventually agreed on the ultimate goal of this process, which is to contribute to the ethical development of AI for good, for the benefit of humanity, sustainable development, and peace.
Preventing harm is high among the objectives of the Recommendation (paragraph 5). It also aims at “stimulating the peaceful use of AI systems”, an amendment suggested by the Brazilian delegation during the negotiations. The principle of peaceful uses of AI should be read in conjunction with paragraph 36 on human oversight and determination. As humans are ethically and legally responsible for all stages in the life cycle of AI systems, it follows that an AI system “can never replace ultimate human responsibility and accountability”. Therefore, the text notes, “as a rule, life and death decisions should not be ceded to AI systems”.
This is an important and far-reaching principle, echoing recommendation 3C on AI and autonomous intelligent systems in the 2019 report of the High-Level Panel on Digital Cooperation convened by the UN Secretary-General, in which the experts reaffirmed the foundational moral principle that “life and death decisions should not be delegated to machines”. Clearly, this emerging principle of international law has been unfolding in a steady fashion.
Other new elements should also be highlighted, such as the notion of proportionality in paragraph 25 (AI technologies must not exceed what is necessary to achieve legitimate aims and should be appropriate to the context), and protection of the environment in paragraphs 22-24 (AI systems must contribute to the peaceful interconnectedness of all living creatures with each other and respect the natural environment). The Recommendation is strong as far as gender is concerned, stressing that AI must not reproduce the gender inequalities found in the real world. A number of suggestions were offered for the empowerment of women and girls in technological research, education, and employability.
Developing countries actively participated in the negotiations. Latin American delegations, for instance, supported by other countries of the Global South, called for the inclusion of an amendment that resulted in paragraph 120: actors developing AI systems in countries that have established or adopted ethical standards on AI should respect these standards when exporting these products, developing or applying their AI systems in countries where such standards may not exist, while respecting applicable international law and domestic legislation, standards, and practices of these countries. The idea here was to avoid that some AI systems could be deployed without supervision or safety precautions in developing countries, which could be used as testbeds for experimentation and data predation by more advanced countries.
Proper follow-up will be necessary to ensure that the Recommendation can effectively provide universal guidelines for policy and legislation. Paragraph 53 states that governments should adopt a regulatory framework that sets out a procedure to carry out ethical impact assessments on AI systems to predict consequences, mitigate risks, avoid harmful consequences, facilitate citizen participation, and address societal challenges. Oversight mechanisms should also be established, including auditability, traceability, and explainability of algorithms, data, and design processes.
While these measures are to be implemented by governments, the Recommendation sets an ambitious agenda for UNESCO. The Organization is requested to support Member States in those tasks by developing a methodology for an “Ethical Impact Assessment” (EIA), based on rigorous scientific research and grounded in international human rights law (paragraph 131). Additional efforts will include capacity-building to train government officials, policymakers, and other relevant AI actors. In the end, it is entirely up to Member States to establish mechanisms for monitoring and evaluation, such as ethics commissions, AI ethics observatories, experience-sharing bodies, AI regulatory sandboxes, and an assessment guide to evaluate adherence to policy recommendations (paragraph 134).
Will governments adopt all these proposals? This is undoubtedly an ongoing endeavor and much remains to be done, but UNESCO has taken the lead and it will continue to pursue initiatives in this direction. Last March, a new International Research Centre on Artificial Intelligence (IRCAI) was inaugurated in Ljubljana, Slovenia, under the auspices of UNESCO. IRCAI is poised to become a global network of institutions and experts across the world, promoting cutting-edge AI research to help achieve the SDGs. More recently, the Globalpolicy.ai portal was launched as a global observatory to promote intergovernmental cooperation on AI governance, providing policymakers with data, research, use cases, and best practices.
One of the philosophical foundations of UNESCO’s approach is acknowledging that AI is both a technological turning point and a major anthropological disruption. These long-term implications will become apparent over the years, as AI outperforms humans in cognitive tasks once taken for granted as a preserve of our brains. We need to be prepared and learn how to cope with non-human knowledge produced by non-biological entities, a social and intellectual phenomenon with profound consequences. We are presently just scratching the surface of this emerging transformation.
UNESCO has a unique perspective to add to this debate given its global membership, multidisciplinary expertise, and multicultural perspectives aiming at promoting human dignity, the rule of law, and a culture of peace. It brings together both developed and developing countries in a culturally diverse international institution.
Within its mandate, UNESCO supplements more traditional approaches for AI governance, many of them chiefly concerned with economic priorities and technological solutions. Interestingly enough, both the Organization for Economic Cooperation and Development (OECD) and UNESCO have their headquarters in Paris. The 2019 OECD AI principles are highly influential. They have been setting international standards in a wide range of areas, from research and development to national legislation. UNESCO’s contribution has incorporated aspects that are generally neglected concerning education, science, culture, information, and other social dimensions that affect people in their daily lives and aspirations.
Looking ahead, thanks to its unmatched range and universality, the United Nations can offer a neutral, bona fide platform for deeper normative discussions on AI in several domains. The UN Secretary-General’s Roadmap for Digital Cooperation, released in 2020, identified three crucial tasks: increase representation from the Global South in AI deliberations; improve overall coordination of existing initiatives; and capacity-building, particularly in the public sector. One of the proposals envisaged an AI Advisory Body, a multistakeholder group of international experts, yet to be appointed, to provide advice and set the stage for consensus-building.
The UN has convening power, legitimacy, and worldwide representation to be the locus for comprehensive negotiations on AI policymaking among states. It can offer a credible and much-needed multilateral track towards ethical and peaceful uses of AI systems. We hope states will take up this challenge and rise to the occasion.