In August 2025, the United Nations adopted a resolution forming the International Scientific Panel on AI (ISP-AI). This was the culmination of several years of work, beginning with the creation of the secretary-general’s AI Advisory Body in 2023. Two of the AI Advisory Body’s final recommendations—the ISP-AI and the creation of the Global Dialogues on AI—were included in the Global Digital Compact, which the UN General Assembly adopted as an annex of the Pact of the Future in 2024. Costa Rica and Spain then led an arduous process during which member states negotiated the terms of reference of the ISP-AI. This was therefore a two-year endeavour, made even more difficult by the high stakes of AI governance and member states’ diverging perspectives on the issue.

However, this work is just getting started. Once operational, the panel will provide an evidence-based assessment of the risks, opportunities, and impacts of AI on a yearly basis. It will analyze existing research rather than conducting new research—opening the door to contributions from scientists globally. This initiative is a milestone in global AI governance, because it will allow member states with different perspectives on risk management, opportunities, and use cases for AI to approach cooperation with a foundation of mutual understanding. Global technical consensus, which is what the ISP-AI hopes to achieve, is a necessary prerequisite for any global commitment or declaration on protecting human beings from AI harms, and harnessing emerging technologies for good.

The mandate of evaluating and synthesizing AI risks and opportunities on a yearly basis is monumental, especially as the technology advances so rapidly. Nevertheless, building scientific and technical consensus on risks of global concern is not new to the United Nations. This has been the basis for several international science and technology cooperation initiatives, most famously the Intergovernmental Panel on Climate Change (IPCC).

Nevertheless, evaluating the many risks of AI adoption is difficult. The MIT AI Risk Repository, for example, has seen its database swell to 1,600 AI risks, organized into seven domains (discrimination and toxicity, misinformation, socioeconomic and environmental harms, etc.). Attempts at developing a standardized categorization of risks for policymakers have included the EU AI Act’s risk severity levels (no or low risk, high risk, unacceptable risk) and the pillars of harm in the US National Institute of Standards and Technology’s AI Risk Management Framework (harm to people, harm to an organization, harm to an ecosystem).

The ISP-AI will clearly need a systematic approach to evaluating risk, ideally one that can be replicated from year to year to allow for comparison. After all, how will we be able to tell if AI policy is having the desired effect unless we can compare risk levels from one year to the next?

The United Nations, in its long history of risk monitoring, has typically taken one of four approaches across several areas of multilateral concern, such as climate change, food insecurity, and nuclear safety.

Approach 1: Global Risk Assessment Conducted by the United Nations

• Who collects the data: United Nations
• Who conducts the analysis: United Nations
• What is the scope of the analysis: Global

The first method is used by the IPCC and others, such as the World Bank and the UN Office for Disaster Risk Reduction’s (UNDRR) Global Risk Assessment Framework. It aims to provide a cohesive and comprehensive view of risks that transcend national boundaries, ensuring that global trends, challenges, and emerging threats are identified and addressed, possibly through global governance mechanisms.

This approach can establish uniform standards and methodologies for assessing risks, facilitating comparability and consistency across different regions. It can also help optimize the allocation of resources and international support to the areas of highest risk or greatest need. Finally, it can encourage international collaboration and policy harmonization to manage risks with cross-border implications.

At the same time, a global assessment alone may overlook or inadequately address specific local or national contexts and unique challenges faced by individual countries. This is why many risk-assessment initiatives have adopted a more granular approach.

Approach 2: National Risk Assessment Conducted by the United Nations

• Who collects the data: United Nations
• Who conducts the analysis: United Nations
• What is the scope of the analysis: National

The second option is inspired by the food insecurity monitoring process adopted by the Integrated Food Security Phase Classification (IPC). In this scenario, the UN assesses the risk of famine in highly vulnerable countries to inform prevention and intervention. The method uses a threshold: once a certain level of food insecurity is crossed (level 2 out of 5), intervention can be triggered. Other initiatives using this methodology include the Interagency Standing Committee and the Joint Research Centre of the European Commission’s INFORM Risk tool, which produces an annual Global Risk Index.

This approach allows the UN to assist countries in developing their risk assessment capabilities, particularly in countries with less experience or fewer resources. It facilitates comparisons between countries, helping to identify best practices and areas where additional support or intervention may be required. Finally, it ensures that global standards are adapted to local contexts, making them more effective and relevant.

However, this approach is resource-intensive and requires consistent and effective coordination between the UN and numerous national agencies and organizations, which can be complex and bureaucratic.

Approach 3: National Risk Assessment Conducted at a National Level and Reported Back to the United Nations

• Who collects the data: National entity
• Who conducts the analysis: National entity first, and then United Nations to compile the analyses into a global view
• What is the scope of the analysis: Both national and global

A less resource-intensive approach for the UN would be for national governments to conduct the bulk of the data collection and analysis. In some cases where risk monitoring is a requirement of an international agreement, the UN provides a monitoring framework for member states to use and report back on. The UN then collates the reports it receives and may send follow-up questions or requests for more data to member states as appropriate. This option is used by the World Health Organization to monitor epidemic risks, as well as by both the International Civil Aviation Organization, with its risk-assessment tool for organizations and national regulating bodies, and the International Atomic Energy Agency, with national self-assessments of nuclear risk.

The main advantage of this approach is that it leverages national bodies’ deep understanding of their own contexts, legal frameworks, and societal needs. It encourages countries to take ownership of their risk assessments, fostering responsibility and accountability for managing those risks. It also provides the UN with an opportunity to aggregate reports from various countries to create a rich and diverse set of data points and perspectives, enhancing the global understanding of risks. Finally, it allows member states to adapt global guidelines to their specific circumstances, which can be critical to address localized risks and challenges effectively.

Nevertheless, the quality and thoroughness of risk assessments may vary significantly between countries, leading to inconsistencies and gaps in the overall understanding of risks. This approach requires strong centralized oversight, as differences in how the methodology is applied can create problems when combining data. National reports may also be delayed or left incomplete due to bureaucratic inefficiencies or lack of prioritization, reducing the timeliness and effectiveness of global risk monitoring.

Approach 4: National Risk Assessment Conducted at a National Level but Informed by the United Nations

• Who collects the data: United Nations
• Who conducts the analysis: National Entity
• What is the scope of the analysis: Both national and global

Some initiatives collect data to create a global data hub without necessarily providing a risk evaluation. This approach resembles the open-data function of the UN, the World Bank, and other institutions, which gather data from a variety of sources and ensure that it can be easily accessed. Historically, this data usually took the form of official statistics provided by national statistical offices and collated by multilateral institutions. Today, other kinds of data are also available, especially through data hubs with specific themes, such as the World Bank’s Climate Change Knowledge Portal. These hubs are often developed with partners such as universities, NGOs, and civil society organizations. In some cases, data has also been collected directly from individuals, such as through the UN Environment Programme’s citizen science data collection techniques

What This Means for the ISP-AI

While each approach has its uses, they are not mutually exclusive. It is likely that the ISP-AI will first lean toward Approach 1, as it may naturally decide to first focus on the global risks of AI—especially those that already coincide with the work of the UN, including on human rights, sustainable development, peace and security, humanitarian aid, and rule of law.

However, there are already immense differences between UN member states in terms of risk exposure (if there is not much AI adoption, there is not much AI risk), risk appetite (innovation-forward countries might have higher risk tolerance), and adoption of risk-management policies. Additionally, there is the ever-important question of whether these policies will work. Does a regulation protecting a country’s citizens against the risks of AI actually do what it is supposed to do? If the ISP-AI wishes to answer these questions, it will need to consider either Approach 2 (conducting its own analyses of national risk) or Approach 3 (asking countries to report back on risk according to specific guidance).

In this case, Approach 2 could be undertaken not by the panel itself but by an observatory that would inform its work. Several such observatories could expand to take on this mandate, such as the Organisation for Economic Co-operation and Development’s (OECD) AI Policy Observatory and UNESCO’s Global AI Ethics and Governance Observatory. These research entities would monitor AI risk across as many countries as possible according to a defined set of metrics and make the results available to use in the panel’s reports. At this level, Approach 4 would also apply, as an observatory would make the data publicly available for others to analyze as well.

Approach 3, where the assessment of risk would be conducted at the national level and then reported back to the UN, is complex but commonly used by multilateral bodies. In this case, emerging national entities for AI—such as standards institutes, AI risks institutes, or ministries or departments of AI—would be tasked with conducting a self-evaluation and making that data available for the ISP-AI to compile and provide a global perspective.

Should the panel aim to produce its first report for launch at a global dialogue in 2026, it may want to begin with Approach 1 and expand into Approaches 2 and 3 over time. However, even in this case, backward compatibility will be important. That is, even an evaluation of AI risks against multilateral objectives and commitments should be measurable so that progress can be tracked. While this will require a concerted effort by the members of the panel, it will ensure that their reports are relevant and impact the global adoption of AI while remaining neutral on what specific policies to adopt.

Eleonore Fournier-Tombs is Head of Anticipatory Action and Innovation at UN University’s Centre for Policy Research and former Research Lead for the UN’s High-Level Advisory Body on Artificial Intelligence.

This article was written with research support from Muznah Siddiqui, Nidhi Polekar, Andrew Ham, Charlie Plumb and Kendal Gee.