G-Science Academies call for establishing an AI research collaborative: Recognizing harmful effects of advanced technologies while promoting global benefits

Organizations similar to CERN (European Organization for Nuclear Research) but focused on AI should be established to provide researchers worldwide with broad and equitable access to computing resources, enable dataset construction, and support multi-way learning among researchers from different countries. The G-Science Academies, composed of academies from G7 countries, including the Science Council of Japan, released a joint statement on May 13 with scientific policy recommendations. This year, under the chairmanship of the Royal Society of Canada, the statement addresses three themes: "Advanced Technology and Data Security," "Sustainable Migration," and "Climate Action and Health Resilience."

The proposal to establish a "CERN for AI" appears within the "Advanced Technology and Data Security" section.

The recommendations first emphasize that governance and regulation are essential to protect people adversely affected by the various opportunities and impacts of new technologies, and to prevent these technologies from further concentrating economic and political power or exacerbating existing inequalities.

However, it also acknowledges challenges in implementation. For example, regarding prevention of unintended data leaks and ensuring data quality, the EU's AI Act and other laws recommend data pseudonymization, but privacy experts note, "This is often insufficient, requiring stronger measures like differential privacy." Additionally, while the demographic distribution of training data used for useful inference should match the target population, methods to achieve this without compromising data confidentiality remain undeveloped.

Policymakers are required to bridge these gaps through two-way communication with experts and the general public to translate legal frameworks into technical requirements. For example, the statement recommends that to prevent worsening inequality, research subjects should be sampled to properly reflect demographic attributes such as language, age, race, and gender, and data should be securely managed. It also suggests that researchers should present the directions of priorities for developing technologies so as to make legal and regulatory compliance easier for practitioners and enable regulatory authorities to appropriately enforce laws.

The statement further points out issues such as discriminatory AI applications, challenges posed by specific vulnerabilities, the need for regulatory authorities to develop their own expertise and capabilities, and the importance of clarifying regulatory responsibilities. While recognizing that advanced technologies raise national security concerns, it emphasizes that academic institutions and governments have a responsibility to consider benefits for humanity and the environment globally, necessitating cooperation for peace and global security. Based on this, it proposes establishing a "CERN for AI," which would provide researchers worldwide with broad and equitable access to computing resources, enable dataset construction, and support multi-way learning among researchers from Global North and Global South countries.

Due to difficulties in training experts suitable for regulatory enforcement, policymakers should incentivize open-source model initiatives, the statement suggests. Such incentives could include support in the form of dedicated funding or resource distribution to help open-source communities maintain software and its integrity. However, decisions about permitting or restricting powerful AI systems as open source should be made under democratic oversight, and safety regulations applied to commercial systems should equally apply to open-source systems.

Generative AI models can produce extremely high-quality media and are being misused for fraud and deception. Such models flood the internet with misinformation, which can be repeatedly utilized by AI models, potentially leading to model deterioration and further misinformation. While watermarking is one solution, it remains vulnerable. Policymakers should encourage the development of various technologies that make data provenance verifiable.

Cloud-based AI technologies like large language models (LLMs) directly impact global climate issues through power consumption. For example, energy usage associated with ChatGPT queries far exceeds that of simple web searches. This issue cannot be addressed by unsubstantiated claims such as "increased energy usage creates incentives to accelerate the switch to sustainable power sources." Governance and regulation of data and its processing should be harmonized with policies for environmental and energy sustainability.