Ai Advisory Body Interim Report
Ai Advisory Body Interim Report
Ai Advisory Body Interim Report
Introduction 2
The Global Governance Deficit 6
Opportunities and Enablers 6
Key enablers for harnessing AI for humanity 8
Governance as a key enabler 9
Risks and Challenges 10
Risks of AI 10
Challenges to be addressed 13
International Governance of AI 14
The AI governance landscape 14
Toward principles and functions of international AI governance 15
Preliminary Recommendations 16
A. Guiding Principles 16
Guiding Principle 1. AI should be governed inclusively, by and for the benefit of all 16
Guiding Principle 2. AI must be governed in the public interest 16
Guiding Principle 3. AI governance should be built in step with data governance and the
promotion of data commons 17
Guiding Principle 4. AI must be universal, networked and rooted in adaptive multi-
stakeholder collaboration 17
Guiding Principle 5. AI governance should be anchored in the UN Charter, International
Human Rights Law, and other agreed international commitments such as the
Sustainable Development Goals 17
B. Institutional Functions 18
Institutional Function 1: Assess regularly the future directions and implications of AI 20
Institutional Function 2: Reinforce interoperability of governance efforts emerging
around the world and their grounding in international norms through a Global AI
Governance Framework endorsed in a universal setting (UN) 20
1
Institutional Function 3: Develop and harmonize standards, safety and risk management
frameworks 21
Institutional Function 4: Facilitate development and use of AI for economic and societal
benefit through international cooperation 21
Institutional Function 5: Promote international collaboration on talent development,
access to compute infrastructure, building of diverse high quality datasets and AI-
enabled public goods for the SDGs 21
Institutional Function 6: Monitor risks, report incidents, coordinate emergency response
22
Institutional Function 7: Compliance and accountability through binding norms 22
Conclusion 25
Next Steps 25
Annexes 27
About the High-Level Advisory Body 27
Members of the High-Level Advisory Body on AI 28
Terms of Reference for the High-level Advisory Body on AI 29
Working Groups and Cross-Cutting Themes 30
List of Abbreviations 31
Introduction
1. Artificial intelligence1 (AI) increasingly affects us all. Though AI has been around for
years, capabilities once hardly imaginable have been emerging at a rapid, unprecedented
pace. AI offers extraordinary potential for good — from scientific discoveries that expand
the bounds of human knowledge to tools that optimize finite resources and assist us in
everyday tasks. It could be a game changer in the transition to a greener future, or help
developing countries transform public health and leapfrog challenges of last mile
access in education. Developed countries with ageing populations could use it to tackle
labour shortages.
2. Yet, there are also risks. AI can reinforce biases or expand surveillance; automated
decision-making can blur accountability of public officials even as AI-enhanced
disinformation threatens the process of electing them. The speed, autonomy, and
opacity of AI systems challenge traditional models of regulation, even as ever more
powerful systems are developed, deployed and used.
3. The opportunities and the risks of AI for people and society are evident and have seized
public interest. They also manifest globally, with geostrategic tensions over access to
the data, compute, and talent that fuel AI, with talk of a new AI arms race. Nor are the
benefits and risks equitably distributed. There is a real danger, even if humanity
harnesses only the positive aspects of AI, that those will be limited to a club of the rich.
1
Per OECD definition: https://oecd.ai/en/wonk/ai-system-definition-update
2
Today’s AI benefits are accruing largely to a handful of states, companies, and
individuals.
4. This technology cries out for governance, not merely to address the challenges and risks
but to ensure we harness its potential in ways that leave no one behind. A key measure
of our success is the extent to which AI technologies help achieve the Sustainable
Development Goals (SDGs). As an example, Box 1 illustrates AI’s potential in tackling
climate change and its impact (SDG 13).
Box 1: Case study illustrating how AI can help address climate change
Climate change represents a global and universal challenge – one where a collective response
requires sustainable digital transformation, thoughtfully designed new infrastructure, and the
ability to deliver precise decision making at scale. AI-driven approaches are particularly well
suited to this challenge, integrating key developments in machine learning, large language
models, high quality data analysis, and more, to create new capacities.
Information that describes disconnected and disparate phenomena – from geospatial imaging,
distributed sensors, real time monitoring, and citizen-reported data on effects of hyperlocal
climate change – can be used to create new understanding of inputs, consequences, and the
complex systems which drive climate outcomes. Taken together with predictive systems that
can transform data into insights and insights into actions, AI-enabled tools may help develop
new strategies and investments to reduce emissions, influence new private sector investments
in net zero, protect biodiversity, and build broad-based social resilience. This can apply to other
SDGs.
The following is a non-exhaustive list of early promises of AI helping to address climate change:
· Assigning responsibility for climate action to national and subnational
governance institutions by creating new and highly granular predictive resources for
climate investment. For example, real time heatmaps of storm-related urban flooding to
unlock hyperlocal infrastructure improvements in sewer and drainage systems.
· Building public, open-source data and AI systems to move private sector net
zero reporting from a static compliance function to a public facing, real time data
repository to increase trust, transparency, and accountability for public commitments.
· Using advanced climate modelling tied to information about urban mobility and
behaviour patterns to create new early warning systems, allowing for more effective
delivery of post conflict/disaster relief and recovery.
· Developing evidence-based AI interventions in open system and other carbon
removal technologies where high uncertainty intervals can limit crucial early-stage
investment. Advanced modelling techniques can lower the cost of scientific inquiry and
allow for rapid prototyping of novel solutions.
But structural barriers remain to help these technologies reach the scale required to match the
scope of the climate crisis and meet the diverse needs of the many critical stakeholders in the
climate fight including corporations, governments, activists, civil society, and others. Systemic
3
risks such as algorithmic bias, transfer context bias, interpretation bias, representation and
allocation harms would have to be considered. Some actions to overcome barriers include:
· Improving model explainability and trust in order to increase adoption of AI-
produced insights into critical climate decision making.
· Ensuring that AI models are trained on diverse, truly representative datasets,
which reflect both commercially viable data collected by for-profit entities and data
which “fills in the gaps” funded by nonprofit, philanthropic, and government resources
and complements local tacit knowledge.
· Providing communities impacted by climate change vulnerabilities access to AI-
generated predictions that would otherwise only be provided to private companies.
· Lowering cost of compute and machine learning expertise so that nonprofits
and civil society can build and sustain free and open AI products.
· Overcoming siloed action from multiple organizations building proprietary
solutions / holding proprietary data to compete for private or philanthropic investment.
· Financing for scaling such solutions
The cross-domain connection between AI and frontline experience of climate change is critical
to enabling these transformative approaches. Solutions which exist in a technical silo – even
when enabled with compute, data, and talent – face significant challenges in uptake and
distribution when they do not reflect the lived experience of community members and local
decision makers.
For each of the opportunities described above, critical early inputs from non-technical
stakeholders need to inform project conception, design, execution, and integration. Enablers
therefore require a values-based approach that prioritizes community interests, a combination of
technical and problem-based expertise, and a comprehensive approach to new AI development.
We also need to keep an eye on the potential negative impact of AI on climate change because
of the associated energy and water consumption.
4
7. Nonetheless, we share the sense of urgency held by complementary governance
initiatives on AI, including those by states as well as regional and intergovernmental
processes such as the EU, the G7, the G20, UNESCO, and the OECD, among others. More
inclusive engagement is needed, however, as many communities — particularly in the
Global South or Global Majority — have been largely missing from these discussions,
despite the potential impact on their lives. A more cohesive, inclusive, participatory, and
coordinated approach is needed, involving diverse communities worldwide, especially
those from the Global South or Global Majority.
8. The United Nations holds no panacea for the governance of AI. But its unique legitimacy
as a body with universal membership founded on the UN Charter, agreed universally, as
well as its commitment to embracing the diversity of all peoples of the world, offer a
pivotal node for sharing knowledge, agreeing on norms and principles, and ensuring
good governance and accountability. Within the UN system, plans for the Global Digital
Compact and the Summit of the Future in September 2024 offer a pathway to timely
action.
9. The Advisory Body comprises individuals diverse by geography and gender, discipline
and age; it draws expertise from government, civil society, the private sector, and
academia. Intense and wide-ranging discussions yielded broad agreement that there is a
global governance deficit in respect of AI and that the UN has a role to play.
10. In this report, we first identify opportunities and enablers that can help harness the
potential benefits of AI for humanity. Second, we highlight risks and challenges that AI
presents now and in the foreseeable future. Third, we argue that addressing the global
governance deficit requires clear principles, as well as novel functions and institutional
arrangements to meet the moment. The report concludes with preliminary
recommendations and next steps, which will be elaborated in our final report by August
2024.
11. Though we are confident of the broad direction, we know that we do not take this
journey alone. We look forward to consulting widely on next steps to ensure that more
voices and views are included, and that AI serves our common good.
5
offers the institutional and normative foundation for collective action in global
governance of AI. Apart from considerations of equity, access, and prevention of harm,
the very nature of the technology itself — AI systems being transboundary in structure,
function, application, and use by a wide range of actors — necessitates a global
approach.
14. Pieces of this puzzle are being filled by self-regulatory initiatives, national and regional
laws, and the work of multilateral forums. Yet, gaps remain and the challenge is clear: a
global governance framework is needed for this rapidly developing suite of technologies
and its use by various actors, be they the developers or users of the technology. AI
presents distinctly global challenges and opportunities that the UN is uniquely
positioned to address, turning a patchwork of evolving initiatives into a coherent,
interoperable whole, grounded in universal values agreed by its member states,
adaptable across contexts.
15. The next three sections outline roles an institution or a network of institutions anchored
in the UN’s universal framework could play in expanding the benefits of AI and mitigating
its risks, as well as the principles and functions that will best achieve these ends.
Examples of AI opportunities
People-assistive AI
AI can assist people in everyday tasks as well as their most ambitious, creative and productive
endeavours. People-assistive AI includes accessibility tools and improvements to education.
Applications have been developed to serve as virtual assistants for people with limited vision or
speech, supporting accessibility needs previously overlooked or neglected. AI-powered
translation now covering over a hundred languages promotes access as well as intercultural
understanding and communication. A new generation of tutoring apps promises to expand
access to quality education worldwide.
Sectoral opportunities
6
AI will have a greater impact in some sectors rather than others. Among the most promising are
agriculture and food security, health, education, protection of the environment, resilience to
natural disasters and combating climate change. For example, AI has been used to create early-
warning systems for floods, now covering over 80 countries, as well as wildfires, and food
insecurity. AI is being used to monitor endangered species (e.g., dolphins, whales) and to
optimize agricultural practices. Within each field, there are myriad possibilities.
AI is broadening access to quality care, for example in the maternal health care space in Sub-
Saharan Africa. Similarly, possibilities exist with respect to environmental problems, making
education more accessible, helping ease poverty and hunger, and making cities safer.
Scientific opportunities
AI is transforming the way in which scientific research is performed and is expanding the frontier
of scientific advancement, including by accelerating molecular and genomic research. AI
systems show special promise for accelerating the work of scientists across many disciplines
and a potential paradigm shift in the way science is practised, from helping explore new
discovery spaces to automating experimentation at scale. For example, AI-powered tools that
predict protein structures are being used by over a million researchers for drug discovery and to
advance understanding of diseases like tuberculosis, as well as many previously neglected
diseases. In the healthcare space, AI is powering diagnostic tools to help doctors with more
timely detection of various types of cancers and eye-related diseases, thereby saving lives. In the
energy space, AI is playing a role in optimizing energy systems and advancing the transition to
renewable energies. For example, AI has been used to boost the value of wind energy, control
tokamak plasmas in nuclear fusion, and enable carbon capture. There is scope for the UN to
encourage progress in AI-enabled science by focusing attention on questions worth solving for
the global good.
Crucially, AI may drive progress in areas where market forces alone have traditionally failed.
These range from extreme weather forecasting and monitoring biodiversity, to expanding
educational opportunities or access to quality healthcare, and optimizing energy systems.
Governments and the public sector can improve services for citizens and strengthen delivery for
vulnerable communities by leveraging AI for social good.
Finally, the use of AI can contribute to accelerating progress towards achieving the Sustainable
Development Goals and enhance the role and effectiveness of the UN in promoting sustainable
development, human rights and peace and security. For example, the UN can use AI to monitor
the development of crisis situations in different parts of the world including human right abuses
or for measuring progress on the SDGs. While many have noted the potential of AI to contribute
to many of the 17 SDGs, many have also noted significant barriers to fully leveraging the
potential of AI to help make progress. The UN and other international organizations have started
to build promising AI use cases and demonstrations in areas such as prediction of food
insecurity, managing relief operations and weather forecasting.
17. The development of AI is now driven by data, compute, and talent, sometimes
supplemented by manual labelling labour. Currently, only well-resourced member states
7
and large technology companies have access to the first three, leading to a
concentration of influence. In addition to global shortages of crucial hardware such as
GPUs, there is also a dearth of top technical talent in the field of AI. It has been
suggested that open model development may alter this dynamic, though the impact and
safety of open models is still being analysed and debated.
18. The AI opportunity arrives at a difficult time, especially for the Global South. An “AI
divide” lurks within a larger digital and developmental divide. According to ITU estimates
for 2023, more than 2.6 billion people still lack access to the Internet. The basic
foundations of a digital economy — broadband access, affordable devices and data,
digital literacy, electricity that is reliable and affordable are not there. Fiscal space is
constrained and the international environment for trade and investment flows is
challenging. Critical investments will be needed in basic infrastructure such as
broadband and electricity, without which the ability to participate in the development and
use of AI will be severely limited. Even outside the Global South, taking advantage of AI
will require efforts to develop local AI ecosystems, the ability to train local models on
local data, as well as fine-tuning models developed elsewhere to suit local
circumstances and purposes.
19. Access and benefits must go hand in hand. Entrepreneurs in regions lagging in AI
capacity require and deserve the ability to create their own AI solutions. This requires
national investments in talent, data, and compute resources, as well as national
regulatory and procurement capacity. Domestic efforts should be supplemented by
international assistance and cooperation not only among governments but also private
sector players. Rallying scientists to solve societal challenges could be a key enabler for
harnessing AI’s potential for humanity. Open-Source and sharing of data and models
could play an important role in spreading the benefits of AI and developing beneficial
data and AI value chains across borders.
20. Enablers (‘common rails’) for AI development, deployment and use would need to be
balanced with ‘guard rails’ to manage impact on societies and communities. A litmus
test will be the extent to which AI governance efforts yield human augmentation rather
than human replacement or alienation as the outcome. Some AI development relies on
cheap and exploitable labour in the Global South. Even in the Global North, there are
questions related to valuing artistic expression, intellectual property, and the dignity of
human labour. Equitable access to these technologies and relevant skills to make full
use of them are needed if we are to avoid “AI divides” within and across nations.
21. AI can and should be deployed in support of the Sustainable Development Goals. But
doing so cannot rely on current market practices alone, nor should it rely on the
benevolence of a handful of technology companies. Any governance framework should
shape incentives globally to promote these larger and more inclusive objectives and help
identify and address trade-offs.
8
22. Comparisons with other sectors offer potential lessons. Mechanisms such as Gavi, the
Vaccine Alliance, may suggest short-term examples for ensuring that the benefits are
shared. Repositories of AI models that can be adapted to different contexts could be the
equivalent of generic medicines to expand access, in ways that do not promote AI
concentration or consolidation.
23. Some of these societally beneficial aspirations may be realized by advances in AI
research itself; others may be addressed by leveraging novel market mechanisms to
level the playing field, or by incentivizing actors to reach all communities and enable
benefits to be accessible to all. But many will not. Ensuring that AI is deployed for the
common good, and that its benefits are distributed equitably, will require governmental
and intergovernmental action with innovative ways to incentivize participation from
private sector, academia and civil society. A more lasting solution is to enable federated
access to the fundamentals of data, compute, and talent that power AI — as well as ICT
infrastructure and electricity, where needed. Here, the European Organization for Nuclear
Research (CERN), which operates the largest particle physics laboratory in the world,
and similar international scientific collaborations may offer useful lessons. A
‘distributed-CERN' reimagined for AI, networked across diverse states and regions, could
expand opportunities for greater involvement. Other examples of open science relevant
to AI include the European Molecular Biology Laboratory (EMBL) in biology or ITER, the
International Thermonuclear Experimental Reactor.
Risks of AI
25. We examined AI risks firstly from the perspective of technical characteristics of AI. Then
we looked at risks through the lens of inappropriate use, including dual-use, and broader
considerations of human-machine interaction. Finally, we looked at risks from the
perspective of vulnerability.
26. Some AI risks originate from the technical limitations of these systems. These range
from harmful bias to various information hazards such as lack of accuracy and
“hallucinations" or confabulations, which are known issues in generative AI.
9
27. Other risks are more a product of humans than AI. Deep fakes and hostile information
campaigns are merely the latest example of technologies being deployed for malevolent
ends. They can pose serious risks to societal trust and democratic debate.
28. Still others relate to human-machine interaction. At the individual level, this includes
excessive trust in AI systems (automation bias) and potential de-skilling over time. At
the societal level, it encompasses the impact on labour markets if large sections of the
workforce are displaced, or on creativity if intellectual property rights are not protected.
Societal shifts in the way we relate to each other as humans as more interactions are
mediated by AI cannot also be ruled out. These may have unpredictable consequences
for family life and for physical and emotional well-being.
29. Another category of risk concerns larger safety issues, with ongoing debate over
potential “red lines” for AI — whether in the context of autonomous weapon systems or
the broader weaponization of AI. There is credible evidence about the increasing use of
AI-enabled systems with autonomous functions on the battlefield. A new arms race
might well be underway with consequences for global stability and the threshold of
armed conflict. Autonomous targeting and harming of human beings by machines is one
of those “red lines” that should not be crossed. In many jurisdictions, law-enforcement
use of AI, in particular real-time biometric surveillance, has been identified as an
unacceptable risk, violating the right to privacy. There is also concern about
uncontrollable or uncontainable AI, including the possibility that it could pose an
existential threat to humanity (even if there are debates over whether and how to assess
such threats).
30. Putting together a comprehensive list of AI risks for all time is a fool’s errand. Given the
ubiquitous and rapidly evolving nature of AI and its use, we believe that it is more useful
to look at risks from the perspective of vulnerable communities and the commons. We
have attempted an initial categorization as per this approach (Box 3), which will be
developed further into a risk assessment framework, building on existing efforts. There
will be dynamicity about risks as technology, its adoption, and use evolve. This speaks to
the need to keep risks under review through interdisciplinary science and evidence-
based approaches. Adaptable risk management frameworks that can be tuned as per
the experience of different regions at different times would also be needed. The UN can
provide a valuable space for such mutual learning and agile adaptation.
• Individuals
o Human dignity/value/agency (manipulation, deception, nudging, sentencing)
o Life, safety, security (autonomous weapons, autonomous cars, interaction with
chemical, biological, radiological and nuclear defence)
o Physical and mental integrity, health and safety (diagnostics, nudging,
neurotechnology)
10
o (other) human rights/civil liberties, e.g. fair trial (recidivism prediction),
presumption of innocence (predictive policing), freedom of expression
(nudging), privacy (biometric recognition)
o Life opportunities (education, jobs, financial stability)
• Groups
o Discrimination/unfair treatment of sub-groups, including on basis of gender
o Group isolation/marginalization
o Functioning of a community
o Social equality/equity (unfair treatment of groups, including on basis of gender)
o Children, elderly, people with disabilities
• Society
o International and national security (autonomous weapons/disinformation)
o Democracy (elections, trust)
o Information Integrity (mis- or disinformation, deep fakes, personalized news)
o Rule of Law (functioning of and trust in institutions, judiciary)
o Security (military and policing uses)
o Cultural diversity and shifts in human relationships (homogeneity, fake friends)
o Social cohesion (filter bubbles, declining trust in news, information)
• Economy
o Power concentration
o Technological dependency
o Unequal economic opportunity
o Resource distribution/allocation
o Under-/overuse of AI, techno-solutionism
• (Eco)systems
o Stability of financial systems
o Risk to critical infrastructure
o Strain on environment/climate/natural resources
• Values and Norms
o Ethical values
o Moral values
o Social values
o Cultural values
o Legal norms
31. There is not yet a consensus on how to assess or address these risks. Nevertheless, as
the precautionary principle provides on environmental dilemmas, scientific uncertainty
about risks should not lead to governance paralysis. Achieving consensus and acting on
it requires global cooperation and coordination, including through shared risk monitoring
mechanisms. International organizations have decades of relevant experience with dual
use technologies, from chemical and biological weapons to nuclear energy, based in
treaty law and other normative frameworks, that could be applied in addressing risks of
AI.
11
32. We also recognize the need to be proactive. There are important lessons in recent
experiences with other globally scalable, high-impact technologies, such as social
media. Even as diverse societies process the impact and implications of AI, the need for
effective global governance to share concerns and coordinate responses is clear.
33. We must identify, classify, and address AI risks, including building consensus on which
risks are unacceptable and how they can be prevented or pre-empted. Alertness and
horizon-scanning for unanticipated consequences from AI is also needed, as such
systems are introduced in increasingly diverse and untested contexts.
Challenges to be addressed
34. Many AI systems are opaque, either because of their inherent complexity or commercial
secrecy as to their inner workings. Researchers and governance bodies have difficulty in
accessing information or fully interrogating proprietary datasets, models, and systems.
Further, the science of AI is at an early stage, and we still do not fully understand how
advanced AI systems behave. This lack of transparency, access, compute and other
resources, and understanding hinders the identification of where risks come from, and
where responsibility for managing those risks (or compensating for harm) should lie.
35. Despite AI’s global reach, governance remains territorial and fragmented. National
approaches to regulation that typically end at physical borders may lead to tension or
conflict if AI does not respect those borders. Mapping, avoiding, and mitigating risks will
require self-regulation, national regulation, as well as international governance efforts.
There should be no accountability deficits.
36. We also need to meet member states where they are and assist them with what they
need in their own contexts given their specific constraints in terms of participation in
and adherence to global AI governance, rather than telling them where they should be
and what they should do based on a context to which they cannot relate.
37. In addition to technical and political hurdles, these challenges exist in a broader social
context. Digital technologies are impacting the ‘software’ of societies challenging
governance writ large. Moreover, there are human and environmental costs of AI —
hardware as well as software — must be accounted for throughout its lifecycle, as
human lives and our environment are at the beginning and end of all AI-integrated
processes.
38. Besides misuse, we also note countervailing worries about missed uses — failing to take
advantage of and share the benefits of AI technologies out of an excess of caution.
Leveraging AI to improve access to education might raise concerns about young
people’s data privacy and teacher agency. However, in a world where hundreds of
millions of students do not have access to quality education resources, there may be
downsides of not using technology to bridge the gap. Agreeing on and addressing such
trade-offs will benefit from international governance mechanisms that enable us to
share information, pool resources, and adopt common strategies.
12
International Governance of AI
39. There is, today, no shortage of guides, frameworks, and principles on AI governance.
Documents have been drafted by the private sector and civil society, as well as by
national, regional, and multilateral bodies, with varying degrees of impact. In technology
terms, governance efforts have been focused on data, models, and benchmarks or
evaluations. Applications have also been under focus, especially where there are existing
sectoral governance arrangements, say for health or dual-use technologies. These
efforts can be anchored in specific governance arrangements, such as the EU AI Act or
the U.S. Executive Order and they can be associated with incentives for participation and
compliance. Figure 1 presents a simplified schema for considering the emerging AI
governance landscape, which the Advisory Body will develop further in the next phase of
its work.
40. Existing AI governance efforts have yielded similarities in language, such as the
importance of fairness, accountability, and transparency. Yet there is no global
alignment on implementation, either in terms of interoperability between jurisdictions or
13
in terms of incentives for compliance within jurisdictions. Some favour binding rules
while others prefer non-binding nudges. Trade-offs are debated, such as how to balance
access and safety — or whether the focus should be on present day or potential future
harms. Different models may also require different emphasis in governance. A lack of
common standards and benchmarks among national and multinational risk
management frameworks, as well as multiple definitions of AI used in such frameworks,
have complicated the governance landscape for AI, notwithstanding the need for space
for different regulatory approaches to co-exist reflecting the world’s social and cultural
diversity.
41. Meanwhile, technical advances in AI and its use continue accelerating, expanding the
gap in understanding and capacity between technology companies developing AI,
companies and other organizations using AI across various sectors and societal spaces,
and those who would regulate its development, deployment, and use.
42. The result is that, in many jurisdictions AI governance can amount to self-policing by the
developers, deployers, and users of AI systems themselves. Even assuming the good
faith of these organizations and individuals, such a situation does not encourage a long-
term view of risk or the inclusion of diverse stakeholders, especially those from the
Global South. This must change.
43. The Advisory Body is tasked with presenting options on the international governance of
AI. We reviewed, among others, the functions performed by existing institutions of
governance with a technological dimension, including FATF, FSB, IAEA, ICANN, ICAO,
ILO, IMO, IPCC, ITU, SWIFT and UNOOSA2. These organizations offer inspiration and
examples of global governance and coordination.
44. The range of stakeholders and potential applications presented by AI and their uses in a
wide variety of contexts makes unsuitable an exact replication of any existing
governance model. Nonetheless, lessons can be learned from examples of entities that
have sought to: (a) build scientific consensus on risks, impact, and policy (IPCC); (b)
establish global standards (ICAO, ITU, IMO), iterate and adapt them; (c) provide capacity
building, mutual assurance and monitoring (IAEA, ICAO); (d) network and pool research
resources (CERN); (e) engage diverse stakeholders (ILO, ICANN); (f) facilitate
commercial flows and address systemic risks (SWIFT, FATF, FSB).
45. Rather than proposing any single model for AI governance at this stage, the preliminary
recommendations offered in this interim report focus on the principles that should guide
the formation of new global governance institutions for AI and the broad functions such
institutions would need to perform. The subfunctions listed in Table 1 below are
informed by a survey of existing research on AI governance as well as a gap-analysis of
nine current AI governance initiatives, namely, China’s interim measures for the
management of AI services, the Council of Europe’s draft Convention on AI, the EU AI
2
See the list of abbreviations in the annex.
14
Act, the G7 Hiroshima Process, the Global Partnership on AI, the OECD AI Principles, the
Partnership on AI and the Foundation Model Forum, the UK AI Safety Summit, and the
U.S. Executive Order 14110.
Preliminary Recommendations
A. Guiding Principles
Guiding Principle 1. AI should be governed inclusively, by and for the benefit of all
46. Despite its potential, many of the world’s peoples are not yet in a position to access and
use AI in a manner that meaningfully improves their lives. Fully harnessing the potential
of AI and enabling widespread participation in its development, deployment, and use is
critical to driving sustainable solutions to global challenges. All citizens, including those
in the Global South, should be able to create their own opportunities, harness them, and
achieve prosperity through AI. All countries, big or small, must be able to participate in AI
governance.
47. Affirmative and corrective steps, including access and capacity building, will be needed
to address the historical and structural exclusion of certain communities, for instance
women and gender diverse actors, from the development, deployment, use, and
governance of technology, and to turn digital divides into inclusive digital opportunities.
15
Governance in this context should expand representation of diverse stakeholders, as
well as offer greater clarity in delineating responsibilities between public and private
sector actors. Governing in the public interest also implies investments in public
technology, infrastructure, and the capacity of public officials.
Guiding Principle 3. AI governance should be built in step with data governance and the
promotion of data commons
51. Data is critical for many major AI systems. Its governance and management in the public
interest cannot be divorced from other components of AI governance (Figure 1).
Regulatory frameworks and techno-legal arrangements that protect privacy and security
of personal data, consistent with applicable laws, while actively facilitating the use of
such data will be a critical complement to AI governance arrangements, consistent with
local or regional law. The development of public data commons should also be
encouraged with particular attention to public data that is critical for helping solve
societal challenges including climate change, public health, economic development,
capacity building and crisis response, for use by multiple stakeholders.
Guiding Principle 4. AI governance must be universal, networked and rooted in adaptive multi-
stakeholder collaboration
52. Any AI governance effort should prioritize universal buy-in by different member states
and stakeholders. This is in addition to inclusive participation, in particular lowering entry
barriers for previously excluded communities in the Global South (Guiding Principle 1).
This is key for emerging AI regulations to be harmonized in ways that avoid
accountability gaps.
53. Effective governance should leverage existing institutions that will have to review their
current functions in light of the impact of AI. But this is not enough. New horizontal
coordination and supervisory functions are required and they should be entrusted to a
new organizational structure. New and existing institutions could form nodes in a
network of governance structures. There is a clear momentum across diverse states for
this to happen as well as growing awareness in the private sector for a well-coordinated
and interoperable governance framework. Civil society concerns regarding the impact of
AI on human rights point in a similar direction.
54. Such an AI governance framework can draw on best practices and expertise from
around the world. It must also be informed by understanding of different cultural
ideologies driving AI development, deployment, and use. Innovative structures within this
governance framework would be needed to engage the private sector, academia, and
civil society alongside governments. Inspiration may be drawn from past efforts to
engage the private sector in pursuit of public goods, including the ILO’s tripartite
structure and the UN Global Compact.
16
Guiding Principle 5. AI governance should be anchored in the UN Charter, International Human
Rights Law, and other agreed international commitments such as the Sustainable
Development Goals
55. The UN has a unique normative and institutional role to play; aligning AI governance with
foundational UN values — notably the UN Charter and its commitment to peace and
security, human rights, and sustainable development — offers a robust foundation and
compass. The UN is positioned to consider AI’s impact on a variety of global economic,
social, health, security, and cultural conditions, all grounded in the need to maintain
universal respect for, and enforcement of, human rights and the rule of law. Several UN
agencies have already done important work on the impact of AI on fields from education
to arms control.
56. The Global Digital Compact and the Roadmap for Digital Cooperation are examples of
multi-stakeholder deliberations towards a global governance framework of technologies
including AI. Strong involvement of UN member states, empowering UN agencies and
involving diverse stakeholders, will be vital to empowering and resourcing a global AI
governance response.
B. Institutional Functions
57. We consider that to properly govern AI for humanity, an international governance regime
for AI should carry out at least the following functions. These could be carried out by
individual institutions or a network of institutions.
17
58. Figure 2 summarizes our recommended institutional functions for international AI
governance. At the global level, international organizations, governments, and private
sector would bear primary responsibility for these functions. Civil society, including
academia and independent scientists, would play key roles in building evidence for
policy, assessing impact, and holding key actors to account during implementation.
Each set of functions would have different loci of responsibility at different layers of
governance — private sector, government, and international organizations. We will
further develop the concept of shared and differentiated responsibilities for multiple
stakeholders at different layers of the governance stack in the next phase of our work.
18
above). They should leverage existing UN organizations and fora such as UNESCO and
ITU for reinforcing interoperability of regulatory measures across jurisdictions. AI
governance efforts could also be coordinated through a body that harmonises policies,
builds common understandings, surfaces best practices, supports implementation and
promotes peer-to-peer learning (subfunctions 7-10 in Table 1). A Global AI Governance
Framework could support policymaking and guide implementation to avoid AI divides
and governance gaps across public and private sectors, regions, and countries as well
as clarifying the principles and norms under which various organizations should operate.
As part of this framework, special attention should be paid to capacity-building both in
the private and public sectors as well as dissemination of knowledge and awareness
across the world. Best practices such as human rights impact assessments by private
and public sector developers of AI systems could be spread through such a framework,
which may a need an international agreement.
Institutional Function 3: Develop and harmonize standards, safety, and risk management
frameworks
63. Several important initiatives to develop technical and normative standards, safety, and
risk management frameworks for AI are underway, but there is a lack of global
harmonization and alignment (subfunction 11 in Table 1). Because of its global
membership, the UN can play a critical role in bringing states together, developing
common socio-technical standards, and ensuring legal and technical interoperability.
64. As an example, emerging AI safety institutes could be networked to reduce the risk of
competing frameworks, fragmentation of standardization practices across jurisdictions,
and a global patchwork with too many gaps. Care should, however, be taken not to
overemphasise technical interoperability without parallel movement on other functions
and norms. While there is greater awareness of socio-technical standards, more
research, active involvement of civil society and transdisciplinary cooperation is needed
to develop such standards.
65. Further, new global standards and indicators to measure and track the environmental
impact of AI as well as its energy and natural resources consumption (i.e. electricity and
water) could be defined to guide AI development and help achieve SDGs related to the
protection of the environment.
Institutional Function 4: Facilitate development, deployment, and use of AI for economic and
societal benefit through international multi-stakeholder cooperation
66. In addition to standards for preventing harm and misuse, developers and users,
especially in the Global South, need critical enablers such as standards for data labelling
and testing, data protection and exchange protocols that enable testing and deployment
across borders for startups as well as legal liability, dispute resolution, business
development, and other supporting mechanisms. Existing legal, financial, and technical
arrangements need to evolve to anticipate complex adaptive AI systems of the future,
and this will require taking into account lessons learnt from forums such as FATF,
19
SWIFT and equivalent mechanisms. In addition, for most countries and regions, capacity
development in the public sector is urgently required to facilitate responsible and
beneficial use of AI as well as participate in international multi-stakeholder cooperative
frameworks to develop enablers for AI (subfunctions 4, 5 and 11 in Table 1).
20
created at a global level to monitor, report, and rapidly respond to systemic
vulnerabilities and disruptions to international stability (subfunctions 13, 14 in Table 1).
71. For example, a techno-prudential model, akin to the macro-prudential framework used to
increase resilience in central banking and bringing together those developed at the
national level, may help to similarly insulate against AI risks to global stability. Such a
model must be grounded in human rights principles.
72. Reporting frameworks can be inspired by existing practices of the IAEA for mutual
reassurance on nuclear safety and nuclear security, as well as the WHO on disease
surveillance.
21
Table 1: Subfunctions for international governance of AI
1.Scientific Prepare a public review of international, regional, and Research & 6-12 months
assessment national AI policies at least every 6 months. Analysis
2.Horizon Prepare a horizon-scanning report that identifies risks Research & 6-12 months
scanning that transcend borders and can potentially affect all Analysis
jurisdictions.
3.Risk Assess existing and upcoming AI models on a risk Research & 6-12 months
classification scale of untenable, high-level, mid-level, and low to no Analysis
risks.
4.Access to Equitable access to technology and benefits of AI, Enabling 12-24 months
benefits accelerating achievement of the Sustainable
Development Goals.
5.Capacity Programs and resources to build AI technology and Enabling 12-24 months
building businesses as well as governance and promotional
capacity among states.
7.Inclusive Ensure participation of all stakeholder groups and all Governing 6-12 months
participation countries and regions in collective governance, risk
management and realization of opportunities; strive for
innovative governance.
22
9.International Deconflicting work and building synergy across Governing 6-12 months
coordination existing international bodies that continue to address
AI.
10.Policy Surfacing best practices for norms and rules, including Governing 12-24 months
harmonization; for risk mitigation and economic growth. Align,
norm alignment leverage, and include, soft and hard law, standards,
methods, and frameworks developed at the regional,
national, and industry level to support interoperability.
11.Standard Develop global consensus on standards for AI use Governing 12-24 months
setting across stakeholder groups by working with national
standards development organizations (SDOs) -
updated regularly.
12.Norm Convene stakeholders to assess the necessity of and Governing 24-36 months
elaboration negotiate non-binding and binding frameworks,
treaties, or other regimes for AI.
15.Monitoring Elaborate oversight and verification schemes where Governing > 36 months
and verification appropriate to ensure that the design, deployment and
use of AI systems is in compliance with applicable
international law.
Conclusion
75. To the extent that AI impacts our lives — how we work and socialize, how we are
educated and governed, how we interact with one another daily — it raises questions
more fundamental than how to govern it. Such questions of what it means to be human
in a fully digital and networked world go well beyond the scope of this Advisory Body.
Yet they are implicated in the decisions we make today. For governance is not an end
but a means, a set of mechanisms intended to exercise control or direction of something
that has the potential for good or ill.
76. We aspire to be both comprehensive in our assessment of the impact of AI on people’s
lives and targeted in identifying the unique difference the UN can make. We hope it is
apparent that we see real benefits of AI; equally, we are clear-eyed about its risks.
23
77. The risks of inaction are also clear. We believe that global AI global governance is
essential to reap the significant opportunities and navigate the risks that this technology
presents for every state, community, and individual today. And for the generations to
come.
78. To be effective, the international governance of AI must be guided by principles and
implemented through clear functions. These global functions must add value, fill
identified gaps, and enable interoperable action at regional, national, industry, and
community levels. They must be performed in concert across international institutions,
national and regional frameworks as well as the private sector. Our preliminary
recommendations set out what we consider to be core principles and functions for any
global AI governance framework.
79. We have taken a form follows function approach and do not, at this stage, propose any
single model for AI governance. Ultimately, however, AI governance must deliver tangible
benefits and safeguards to people and societies. An effective global governance
framework must bridge the gap between principles and practical impact. In the next
phase of our work, we will explore options for institutional forms for global AI
governance, building on the perspectives of diverse stakeholders worldwide.
Next Steps
80. Rather than proposing any single model for AI governance at this stage, the foregoing
preliminary recommendations focus on the principles and functions to which any such
regime must aspire.
81. Over the coming months we will consult — individually and in groups — with diverse
stakeholders around the world. This includes participation at events tasked with
discussing the issues in this report as well as engagement with governments, the private
sector, civil society, and research and technical communities. We will also pursue our
research, including on risk assessment methodologies and governance interoperability.
Case studies will be developed to help think about landing issues identified in the report
in specific contexts. We also intend to dive deep into a few areas, including Open-
Source, AI and the financial sector, standard setting, intellectual property, human rights,
and the future of work by leveraging existing efforts and institutions.
82. We encourage constructive engagement from anyone with an interest in AI. More
information about how to engage with our ongoing work can be found online at
https://www.un.org/en/ai-advisory-body.
83. We look forward to engaging with diverse stakeholders as we answer more fully the
questions identified in this interim report, in support of the ongoing efforts of the United
Nations on digital cooperation and on social progress and better standards of life in
larger freedom.
24
Box 4: Example of questions to be addressed during consultations on this interim report
Would common standards for data labelling and testing encourage AI startups to test and
deploy across more countries and regions?
How can we grow and spread AI talent? Can UN entities or other institutions facilitate
exchange of students, joint PhD programmes, and cross-domain (health and AI, agriculture
and AI) talent development?
How can international collaboration harness AI talent, data and compute for scientific
research and for the SDGs?
How can we incentivize governments and the private sector to invest in other core
infrastructures that drive AI development around the world?
What is the best path to reaching consensus on identifying, classifying, and addressing AI
risks?
How should assessments of risks and challenges relate to more specific use cases of AI,
notably autonomous weapons systems?
What should be the threshold or the trigger for identifying red lines (analogous, perhaps, to
the ban on human cloning in biomedical research)? How would any such red line be
policed and enforced?
International governance of AI
Do the principles listed above properly reflect the aspirations that a global governance
regime for AI should have?
Do the functions outlined above properly reflect what global AI governance can and should
do?
25
What structural arrangement(s) would best empower a new institution or set of
institutions to uphold these principles and carry out these functions?
A range of models exist within the UN system for engaging industry in sectoral work
(WHO, ITU, ICAO etc).
What kind of financing and capacity building mechanisms would be needed for effective
international arrangements to address the functions outlined above?
26
Annexes
Initially proposed in 2020 as part of the Secretary-General’s Roadmap for Digital Cooperation
(A/74/821), The multi-stakeholder High-level Advisory Body on Artificial Intelligence was formed
in October 2023 to undertake analysis and advance recommendations for the international
governance of AI.
Advisory Body members participated in their personal capacity, not as representatives of their
respective organizations. This proposal represents a majority consensus; no member is
expected to endorse every single point contained in the document. In publishing this report, the
UN AI Advisory Group members affirm their broad, but not unilateral, agreement with its findings
and recommendations. Language included in this report does not imply institutional
endorsement by the UN AI Advisory Group members’ respective organizations.
27
Terms of Reference for the High-level Advisory Body on AI
The High-level Advisory Body on Artificial Intelligence, convened by the United Nations
Secretary-General, will undertake analysis and advance recommendations for the international
governance of artificial intelligence. The Body’s initial reports will provide high-level expert and
independent contributions to ongoing national, regional, and multilateral debates.
The Body will consist of 38 members from governments, private sector, civil society, and
academia, as well as a member Secretary. Its composition will be balanced by gender, age,
geographic representation, and area of expertise related to the risks and applications of
artificial intelligence. The members of the Body will serve in their personal capacity.
The Body will engage and consult widely with governments, private sector, academia, civil
society, and international organizations. It will be agile and innovative in interacting with existing
processes and platforms as well as in harnessing inputs from diverse stakeholders. It could set
up working parties or groups on specific topics.
The members of the Body will be selected by the Secretary-General based on nominations from
Member States and a public call for candidates. It will have two Co-Chairs and an Executive
Committee. All stakeholder groups will be represented in the Executive Committee.
The Body shall be convened for an initial period of one year, with the possibility of extension by
the Secretary-General. It will have both in-person and online meetings.
The Body will prepare a first report by 31 December 2023 for the consideration of the Secretary-
General and the Member States of the United Nations. This first report will present a high-level
analysis of options for the international governance of artificial intelligence.
Based on feedback to the first report, the Body will submit a second report by 31 August 2024
which may provide detailed recommendations on the functions, form, and timelines for a new
international agency for the governance of artificial intelligence.
The Body shall avoid duplication with existing forums and processes where issues of artificial
intelligence are considered. Instead, it shall seek to leverage existing platforms and partners,
including UN entities, working in related domains. It shall fully respect current UN structures as
well as national, regional, and industry prerogatives in the governance of artificial intelligence.
The deliberations of the Body will be supported by a small secretariat based in the Office of the
Secretary-General’s Envoy on Technology and be funded by extrabudgetary donor resources.
28
Working Groups and Cross-Cutting Themes
The ongoing work of the Advisory Body is organized around five working groups and ten cross-
cutting themes. Sectoral applications and additional themes will be considered in detail in the
next phase.
Working Groups
Opportunities and Enablers
Risks and Challenges
Interoperability
Alignment with Norms and Values
International Institutions
Cross-Cutting Issues
Culture
Equity
Ethics
Future of work
Government capacity
Gender
Human Rights, Democracy, Rule of Law
Open-Source
Societal impact
Sustainability
29
List of Abbreviations
EU European Union
G7 Group of Seven
G20 Group of 20
30