Work

Solutions for frontier AI safety: the case for science-driven international cooperation and standardization at a time of radical transformations

Year: 2026

Type: article

Abstract: As AI systems and their underlying technologies continue to advance at extraordinary speed, eminent scientists are warning of risks, including extreme risks. While high-profile governmental and intergovernmental initiatives such as AI summits, OECD and UN initiatives have raised hopes, substantial gaps remain between the severity of potential harms and the capacity of these initiatives to address them. Owing to limited research, private sector dominance, minimal investment in AI safety, the voluntary nature of most measures recommended, and the current geopolitical context, it is unlikely that multilateral actions will deliver the effective and timely measures required. Other approaches are needed and are possible. They should not be regarded as alternatives to multilateral efforts, but more as complementary actions to provide concrete and timely solutions that, at some point, can be integrated into the multilateral framework. We take inspiration from historical examples of scientific, technical and organizational collaboration in which a plurality of committed entities has achieved outstanding results. These historical examples show that, even with limited resources, collaboration can enormously increase the collective ability to attain far reaching goals. We propose the creation of an international network led by top scientists and composed of AI labs and research groups, think tanks and other organizations keen to share ideas, materials and data, while pursuing three common goals: 1) advance AI safety science; 2) develop practical techniques, tools, best practices, and technologies for building safe AI systems; 3) create standards for operationalizing the use of these tools and technologies, including conformity assessment mechanisms, to support their widespread adoption.

Source: Frontiers in Artificial Intelligence

DOI: 10.3389/frai.2026.1950290

Authors: Daniele Gerundino, Leonardo Chiariglione, Yishu Mao, Zheng Liang