After further rounds of consultations with member states, the co-facilitators from Spain and Costa Rica have published Revision 3 for the modalities of the Independent International Scientific Panel on AI and the Global Dialogue on AI Governance. We support Rev. 3 as a hard-wrought and reasonable compromise that provides a functioning structure for both panel and dialogue.
Panel structure - from Rev. 1 to Rev. 3
As a reminder, when we last discussed the panel structure of Rev. 1, there was a 40-member scientific steering body called the “Panel”, which had a 5-member executive body called the “Bureau”. Member States and the Secretary General selected a 20-member “Independent Appointment Committee”, which in turn selected the Panel.
Fig. 1: Structure of the Independent International Scientific Panel on AI in Rev. 1
In the panel structure of Rev.3, the “Independent Appointment Committee” has been replaced with a direct, criteria-based selection of the “Panel” that involves the Secretary General with confirmation through the General Assembly. Similarly, the governmental experts in the Plenary of AI have been replaced by the General Assembly Plenary.
Fig. 2: Structure of the Independent International Scientific Panel on AI in Rev. 3
Working groups remain essential for scientific independence
The intermediate draft of Rev. 2 had temporarily omitted working groups. We are strongly supportive of their return in Rev. 3. Working groups have never been a particularly contentious issue amongst member states. However, based on consultations with a wide range of experts and lessons from successful science-policy bodies, we have long identified them as a key requirement for the design of a functional scientific panel:
- The scientific panel has the mandate to synthesize large amounts of research and to write assessment reports. This requires a panel with more working capacity than a panel whose mandate it is to provide an advisory opinion.
- Senior experts typically have limited time to dedicate to the panel’s “leg work”. Experts at this career stage tend to provide guidance and manage other researchers. They will typically not have the time to conduct vast literature reviews by themselves - a gap that working groups can fill.
- Expertise is specialized, even within a field. There are more than 200,000 scientific publications on AI per year. The ability to tap specialized experts in specific subfields is important to ensure that the report produced by the panel is state-of-the-art.
- Best practice: Successful scientific panels aimed at the synthesis of large fields of research have working groups with external experts. For example, the International AI Safety Report has a narrower focus than the Independent International Scientific Panel on AI, yet, it still required a writing group.
Areas of contention
During the negotiations of Rev. 2 & 3, the following have been the core areas where there is a need to find compromises:
- Selection of the experts for the panel: Some stakeholders would want a more prominent role of member states in selecting the experts, others fear that member state election leads to higher politicization of expert selection.
- Interaction of the panel with experts from member states: Some stakeholders would prefer more extensive and more binding interaction of the panel with a plenary of experts from member states; others would want to keep this to a minimum.
- Participation support: Some stakeholders would want a fellowship that supports researchers in an early career stage and/or from developing countries in participating as experts in the panel & dialogue, others oppose any potential cost expansion in the current UN budget environment.
- Outcome formats of the Global Dialogue: Some stakeholders want negotiated outcome statements, arguing that they are excluded from other international AI governance processes where such statements are standard. In contrast, other stakeholders oppose mandating strong outcome formats, wanting to retain control on whether, when, and how the role of the UN in AI governance evolves.
Rev. 3 is a good compromise
Any arrangement between 193 member states requires compromises. Rev. 3 is a compromise within and across all the four areas above.
The Secretary General makes a key selection of experts for the Panel, but this is preceded by nominations and followed by a formal approval by member states. Rev. 3 explicitly mentions travel support for representatives and stakeholders of developing countries, but only on a voluntary basis. Rev. 3 foresees that the Global Dialogue will lead to an UNGA resolution, but only in 2027, and more as a review of discussions than as an aspirational vision.
The question of outcome formats for the Global Dialogue remains the most contentious issue. At the same time, it’s good for negotiators to not lose sight of the forest for the trees. First, there have already been US-led and China-led UNGA resolutions on AI with broad and mutual support, showing that there is room for UNGA resolutions on AI with mutual agreement. Second, AI will not wait for a specific UN process, and by 2027 the AI governance discussion might look significantly different from today. For example, we might see the emergence and diffusion of AI models that pose significant catastrophic risks.
Rev. 3 is not perfect, and it is likely that stakeholders will find a few details with which they are only somewhat happy (e.g. we would prefer selecting panel experts exclusively by their scientific credentials). However, from our perspective, the question is not whether we need to find pragmatic compromises or not. It’s whether these compromises still yield a functional structure for the panel & dialogue. The answer to this is yes. As such, Rev. 3 is a good compromise, and we support its adoption to begin a timely implementation.