Getting to Geneva required careful political deliberation by the first Co-Chairs to keep all member states on board, including in how they designed the clusters. Yet, as they noted in their summary, views still differ on the “scope, priorities and institutional forms of AI governance.” The second cycle raises the bar. Beyond convening the Dialogue, the new Co-Chairs are tasked with holding “intergovernmental consultations to agree on common understandings on priority areas for international artificial intelligence governance,” as an input to the high-level review of the Global Digital Compact scheduled for September 2027. Therefore, this deliverable will shape how the road to the Dialogue and beyond will be structured. This includes the roadmap, the choice of themes, how consultations are run, and what they produce. It will also show how the Dialogue builds on what already exists—from the clusters and coalitions to the initiatives that have formed around it. Finally, it will set a precedent for what the Dialogue becomes beyond its current mandate.
A central question for the Co-Chairs is: what preparation would allow governments to move towards agreeing on priorities for international cooperation on AI. The Panel offers a powerful resource. As the new Co-Chairs shape a roadmap for the months ahead, they can use the Panel’s work as a trusted scientific compass to guide the consultations and structure of the second Dialogue, anchoring discussions in shared evidence and helping turn broad areas of convergence into common understandings.
Overcoming Evidence Dilemmas: Overload, Lag, and Reliability
Navigating the political complexities of designing any governance framework through a universal platform is inherently difficult, especially for a topic as broad and consequential as artificial intelligence. Yet the challenge of building AI governance is not only political. Diplomats also face three evidence problems: overload, lag and reliability. While the Panel cannot solve the political questions that member states must work through in the Global Dialogue, it can provide a sense of direction on all three.
First, information on AI and its governance implications appears faster than diplomats and other stakeholders can absorb it, creating an evidence overload. From daily news on AI incidents, studies assessing future risks, to advocacy on its benefits and strategies for putting it to work for society, it is hard for diplomats to know what to prioritize when shaping policy. The Panel’s mandate is to cut through this by synthesizing existing research and determining where scientists and academics agree. In the report’s own words, this helps “level the information playing field for policymakers worldwide.” Smaller delegations feel this most, but even well-staffed missions often lack the technical experts to keep up with the technological advancements.
Second, there is the problem of evidence lag: technology moves first, evidence follows, and governance comes last. The Panel’s report calls this an “evidence dilemma” whereby policymakers need evidence to act, “but by the time the evidence exists, it might be too late.” The Panel is designed to narrow this gap, updating its analysis throughout the year with thematic briefs that address developments as they arise.
Third, overload and lag together make it hard to know which information to trust, creating a reliability issue. The Panel is built to be a source member states can rely on. Its 40 independent experts are appointed by the General Assembly “on the basis of their outstanding expertise in AI and related fields,” come from all regional groups, and from disciplines ranging from core technical AI, to ethics and societal impact. Its work sets out the evidence, leaving the political conversation to member states.
The Panel’s first thematic brief, “AI Agents, Misalignment and the Risk of Losing Human Control,” released on September 21, 2026, showed all three strengths at work. Over the summer, AI agents used in OpenAI’s internal training and cybersecurity evaluations went around their restrictions and compromised parts of one of the leading AI companies, OpenAI’s research infrastructure and private software company Hugging Face’s live systems—which the Panel called “an early warning of one possible route to more severe future loss of control.” Within a month of OpenAI publishing its full account of the incident, the Panel released a short, rapid-response and evidence-based brief on the issue, enabling member states and other stakeholders to understand a complex issue quickly from a trusted source.
These three challenges will resurface in the decisions the Co-Chairs now face as they design the second cycle, and the Panel can help with each.
From Evidence to Decisions
The first Global Dialogue laid the foundations. It set the scope of discussions, organized them into four clusters, and identified broad areas of convergence. The next stage must build on it while preparing the ground for the intergovernmental consultations next summer. A useful starting point is to answer four practical questions outlined below:
| Decisions | Actionable Recommendations |
|---|---|
| Which questions merit focused discussion? | 1. Choosing priorities for focused discussion: combine scientific findings with member state priorities, stakeholder experience, and the need for international cooperation (Addresses issue of overload). |
| How should consultations be organized? | 2. Organizing consultations around specific questions: frame each around a defined governance question and identify established findings, uncertainties, and competing options (Addresses issue of reliability). |
| What should consultations produce? | 3. Turning discussions into common understandings: a short account of possible cooperation, outstanding disagreements, and matters ready for intergovernmental consideration (Addresses issues of reliability and lag). |
| How should the process adapt? | 4. Updating the agenda as evidence develops: allow emerging findings to update discussions and transmit unresolved scientific questions back to the Panel (Addresses issue of lag). |
1. Choosing priorities for focused discussion
The new Co-Chairs will need to foster more focused discussions on a specific set of priorities, with the aim of enabling these to turn from broad agreement into common understandings that member states can act on. Choosing priority areas is already a response to the issue of overload. With dozens of possible topics linked to global AI governance, the Co-Chairs need a way to decide which merit focused discussion. Three filters can help with this:
- What the scientific evidence shows.
- Where Member States’ political interest lies.
- Where initiatives already exist to build on.
Child safety is the clearest example. It is the only rights issue that recurs in three separate places in the preliminary report, and the Panel has announced a brief on it. In Geneva, a coalition on children’s rights was launched by several member states, and the Co-Chairs’ Summary returned to the issue repeatedly, noting that it “resonated across all regions and stakeholder groups.” This passes all three filters: a strong evidence base from the Panel, political interest by member states, and an existing coalition to carry it forward. The Panel has also announced briefs on AI and the environment, the effectiveness of governance instruments, and sectoral topics— including space, quantum, legal and judicial systems, and financial markets. Several of these, such as the environment, also appear among the priorities and gaps flagged in the Co-Chairs’ Summary, making them obvious options for more focused discussion.
Anchoring these choices in the Panel’s work also gives the Co-Chairs an independent reference point, strengthening perceptions of reliability.
2. Organizing consultations around specific questions
Once priorities are chosen, each consultation should be framed around a defined governance question, setting out what the evidence establishes, what remains uncertain, and which options are on the table. This is where the Panel’s reputation as a reliable scientific compass matters. The Panel can bolster legitimacy of discussions by providing a common factual footing. While it cannot settle political disagreements— this is the task of member states—it can help ensure that conversations start from a shared understanding of the facts, separating what the evidence already shows from what remains uncertain so that the Dialogue spends its limited time on genuine political choices.
For example, on governance instruments, the Panel’s preliminary report finds that dozens are already in use but are “fragmented… and rarely measure real-world effectiveness.” The Panel has announced further work in this area, which can give the Dialogue a clear starting point for discussing which existing instruments to build on. A consultation on this topic could address questions on which existing instruments governments should build on and how they can be made more interoperable, so that they work better together rather than in parallel.
In designing the format, context matters. Each governance question calls for a format to which it is best suited to. Working groups aligned with the Panel’s announced briefs, suit issues with strong evidence and political momentum. Each Panel brief could be followed by a member state briefing on what it means for governance so that sessions start from established facts. For the seven evidence gaps the Panel identified in its preliminary report, from labor market effects to the global AI supply chain, the Co-Chairs could call for written submissions on specific gaps, use the AI Dialogue Partnerships Hub* to surface initiatives already producing relevant data and opportunities, and invite experts from academia, civil society and the private sector to brief member states, including in smaller groupings.
3. Turning discussions into common understandings
With the second Dialogue meeting on May 3–4, 2027, two months earlier in the year than Geneva, and the review of the Global Digital Compact following in September, there is little time between the Dialogue and the mandated follow-up consultations. Thus, each consultation should aim to produce something the intergovernmental process can build on, with an eye towards common understandings. First, it should identify where practical cooperation is already possible (e.g., through member state coalitions led by ‘champions’) without waiting for agreement among all member states. Second, it should find where disagreements remain, and what it would take to narrow them (including by calling on further evidence from the Panel). Third, it should evaluate which areas have enough convergence to be taken forward into the mandated consultations as candidate common understandings.
4. Updating the agenda as evidence develops
Finally, the roadmap could account for evidence lag. The Panel’s ad hoc briefs can flag emerging issues the Co-Chairs could not have anticipated when drafting the roadmap. As the September brief on AI agents showed, the Panel has the ability to respond quickly to new risks, and the issues member states need to discuss three or six months from now may not be visible today. The roadmap should leave room for them.
The relationship also runs the other way. The Panel has said it “will take into account the outcomes of the Global Dialogue… and stands ready to take on new tasks as set by Member States.” The evidence shapes the conversation, and the conversation can shape which evidence comes next. Member states can signal where further research could be valuable.
The Panel as a Compass
Ultimately, the Dialogue remains independent of the Panel. Member states will decide its course, what to prioritize, and what goes into the common understandings. But by deliberately aligning parts of the Dialogue’s work with the Panel’s, and grounding both the preparatory discussions and the common understandings in its independent evidence, the Co-Chairs can make agreement easier to reach. Used this way, the Panel becomes a compass that gives the process clearer direction and helps move it forward.